Self-Driving Science: How Autonomous Platforms Like SAMPLE Are Revolutionizing Protein Engineering

Abstract

Autonomous protein engineering platforms are transforming the slow, labor-intensive process of protein design into a rapid, automated, and data-driven endeavor. This article explores the core architecture of these systems, focusing on the pioneering SAMPLE platform, which integrates artificial intelligence with robotic automation to navigate protein fitness landscapes without human intervention. We detail the methodological breakthroughs—from Bayesian optimization and protein language models to fully integrated biofoundries—that enable these platforms to achieve in weeks what traditionally took years. For researchers and drug development professionals, this review provides a comprehensive analysis of current applications, troubleshooting of experimental hurdles, validation of platform performance against traditional methods, and a forward-looking perspective on how autonomous experimentation will accelerate breakthroughs in biomedicine and therapeutic development.

The New Paradigm: Understanding Autonomous Protein Engineering

Defining Autonomous Protein Engineering Platforms

Autonomous protein engineering platforms represent a paradigm shift in biotechnology, integrating artificial intelligence (AI), robotics, and data science to create self-directed systems for designing and optimizing proteins. These platforms automate the classic Design-Build-Test-Learn (DBTL) cycle, dramatically accelerating the process of developing enzymes and therapeutic proteins with enhanced properties such as improved catalytic activity, stability, and specificity [1]. By minimizing human intervention, they address key challenges in traditional protein engineering: the vastness of protein sequence space, the time and cost of experimental workflows, and the dependency on specialist knowledge [1] [2].

The core value of these systems lies in their ability to close the loop between computational design and experimental validation. AI models propose promising protein variants, robotic biofoundries synthesize and test them, and the resulting data are fed back to refine the AI's predictions. This creates a rapid, iterative learning cycle that efficiently navigates sequence space which is intractable for manual methods [1]. As a result, autonomous platforms are poised to drive advancements across diverse fields, from drug development and diagnostic tools to the creation of novel biocatalysts for sustainable chemistry [3] [1].

Quantitative Comparison of Representative Platforms

The table below summarizes the performance and key features of several recently developed autonomous and semi-autonomous protein engineering systems.

Table 1: Performance Metrics of Selected Autonomous Protein Engineering Platforms

Platform / System Name	Core Technology	Reported Improvement	Timeframe	Key Innovation
T7-ORACLE [3]	Orthogonal DNA replication in E. coli	Evolution of antibiotic resistance enzymes surviving doses 5,000x higher than wild-type.	Less than 1 week	Continuous hypermutation (100,000x normal rate) inside living cells.
AI-Powered Platform (iBioFAB) [1]	AI (LLM & ML) + Robotic Biofoundry	90-fold improvement in substrate preference; 16-fold and 26-fold improvement in activity for two different enzymes.	4 rounds over 4 weeks	End-to-end automation integrated with protein language models for generalizable application.
METL Framework [4]	Biophysics-based Protein Language Model	Effective design of functional GFP variants from minimal data.	N/R	Pretraining on synthetic biophysical data for superior generalization from small datasets (~64 examples).
COMPSS Framework [2]	Composite Computational Metrics	Improved experimental success rate by 50-150%.	N/R	A benchmarked set of metrics for reliably selecting functional, computer-generated protein sequences.

Abbreviations: N/R: Not explicitly reported in the provided context; LLM: Large Language Model; ML: Machine Learning.

Detailed Experimental Protocols

This section outlines two foundational methodologies for autonomous protein engineering: a generalized platform for AI-driven engineering and a continuous evolution system.

Protocol: Generalized AI-Driven Enzyme Engineering on a Biofoundry

This protocol describes an end-to-end autonomous workflow for engineering enzymes, as demonstrated by the platform implemented on the Illinois Biological Foundry for Advanced Biomanufacturing (iBioFAB) [1].

1. Design Phase

• Input Requirements: Provide the wild-type protein sequence and a quantifiable assay to measure fitness (e.g., enzyme activity under specific conditions).
• Initial Library Design: Use a combination of computational models to generate a high-quality, diverse initial library of mutant sequences.
  • Protein Language Model: Utilize a model like ESM-2 to predict the likelihood of amino acids at specific positions based on evolutionary context [1].
  • Epistasis Model: Employ a model like EVmutation to account for synergistic effects between mutations [1].
  • Curation: Combine the outputs of these models to select a library of ~180 variants for the first round of experimentation.

2. Build Phase (Automated on iBioFAB) The following modules are executed automatically by the robotic foundry:

• Module 1: Mutagenesis PCR. A HiFi-assembly based mutagenesis method is used to construct variant genes, eliminating the need for intermediate sequencing verification and ensuring >95% accuracy [1].
• Module 2: DNA Assembly. Variant genes are cloned into an appropriate expression plasmid.
• Module 3: Transformation. Plasmids are transformed into a microbial host (e.g., E. coli).
• Module 4: Colony Picking. Robotic systems pick individual colonies and inoculate cultures in deep-well plates.
• Module 5: Protein Expression. Cultures are induced to express the target protein variants.
• Module 6: Crude Lysate Preparation. Cells are lysed to release the expressed proteins for functional assays.

3. Test Phase (Automated on iBioFAB)

• Module 7: Functional Enzyme Assay. An automated, high-throughput assay (e.g., a spectrophotometric readout) is performed on the crude lysates to measure the fitness of each variant [1].
• Data Recording: All experimental data is automatically recorded and structured for the learning phase.

4. Learn Phase

• Machine Learning Model Training: The experimental data from the "Test" phase is used to train a low-N machine learning model (e.g., a Bayesian optimizer) to predict variant fitness [1].
• New Variant Proposal: The trained model proposes a new set of variants for the next DBTL cycle, focusing the search on the most promising regions of the sequence space.

5. Iteration

• The cycle (Steps 1-4) is repeated automatically. Typically, 4 rounds of iteration over 4 weeks, involving the construction and testing of fewer than 500 total variants, can yield significant improvements [1].

Protocol: Continuous Protein Evolution with T7-ORACLE

This protocol leverages the T7-ORACLE system for the continuous, accelerated evolution of proteins in E. coli [3].

1. System Setup

• Strain Preparation: Use an engineered E. coli strain that hosts the orthogonal T7 replication system.
• Plasmid Construction: Clone the gene of interest (e.g., an antibiotic resistance gene, therapeutic enzyme, or antibody fragment) into a special plasmid (replicon) that is recognized by the engineered T7 replication machinery.
• Introduce Error-Prone Polymerase: The key to the system is an engineered T7 DNA polymerase that is error-prone, introducing random mutations specifically into the target plasmid at a rate ~100,000 times higher than normal cellular replication [3].

2. Continuous Evolution Cycle

• Culture Growth: Grow the transformed E. coli cells in liquid medium.
• Mutation Generation: As the cells divide (approximately every 20 minutes), the orthogonal T7 replication system continuously replicates the target plasmid, introducing random mutations into the gene of interest with each replication cycle [3].
• Selection Pressure: Apply a selective pressure to the growing culture. For example:
  • To evolve antibiotic resistance, add the antibiotic to the culture medium, gradually escalating the dose over time [3].
  • For other functions, use Fluorescence-Activated Cell Sorting (FACS) or other selection mechanisms.
• Harvesting Variants: After multiple generations of growth under selection (e.g., less than one week), harvest the culture. The surviving population will be enriched with plasmids encoding protein variants that have evolved to function under the applied selective pressure.

3. Analysis and Validation

• Isolate plasmids from the evolved population.
• Sequence the gene of interest to identify the beneficial mutations.
• Clone and express the evolved gene variants for functional validation in a clean genetic background.

Autonomous Protein Engineering DBTL Cycle

The Scientist's Toolkit: Research Reagent Solutions

Table 2: Essential Research Reagents and Tools for Autonomous Protein Engineering

Reagent / Tool	Function / Description	Example / Application
Protein Language Models (PLMs)	AI models trained on protein sequences to predict the likelihood of amino acids and infer variant fitness.	ESM-2 [1]; Used to design initial diverse variant libraries.
Biophysics-Based Models	AI models pretrained on molecular simulation data to capture sequence-structure-energy relationships.	METL framework [4]; Excels in prediction when experimental data is scarce.
Orthogonal Replication System	A separate DNA replication machinery within a host cell that mutates a target plasmid at a very high rate.	T7-ORACLE [3]; Enables continuous directed evolution in E. coli.
Robotic Biofoundry	An integrated suite of automated liquid handlers, incubators, and analytical instruments.	iBioFAB [1]; Executes the Build and Test phases of the DBTL cycle without human intervention.
Error-Prone DNA Polymerase	An engineered enzyme that introduces random mutations during DNA replication.	Engineered T7 DNA polymerase in T7-ORACLE [3]; Drives hypermutation of the target gene.
High-Throughput Assay	A quantifiable, automated method for measuring protein function (e.g., activity, binding).	Spectrophotometric enzyme activity assays [1]; Provides the fitness data for the Learn phase.
Composite Computational Metrics	A set of computed scores to filter and select generated protein sequences likely to be functional.	COMPSS framework [2]; Improves the experimental success rate by 50-150%.
(rac)-Exatecan Intermediate 1	(rac)-Exatecan Intermediate 1, CAS:10298-40-5, MF:C13H13NO5, MW:263.25 g/mol	Chemical Reagent
Heptadecanyl stearate	Heptadecanyl stearate, CAS:18299-82-6, MF:C35H70O2, MW:522.9 g/mol	Chemical Reagent

Autonomous Platform Core Architecture

The field of protein engineering is undergoing a transformative shift with the advent of autonomous laboratories that seamlessly integrate artificial intelligence (AI), robotics, and data analytics. These platforms implement a closed-loop Design-Build-Test-Learn (DBTL) cycle, where AI algorithms design protein variants, robotic systems build and test them, and the resulting data is used to learn and inform the next design cycle [5]. This integration enables the exploration of protein sequence spaces at an unprecedented scale and efficiency, moving protein engineering from a bespoke, human-dependent process to a scalable, automated science [6] [7]. The core value proposition of these platforms lies in their ability to operate with minimal human intervention, dramatically accelerating the rate of scientific discovery and application in areas such as therapeutic development, industrial biocatalysis, and renewable energy [5].

Core Architectural Framework

The architecture of an autonomous protein engineering platform is a sophisticated integration of computational and physical components. The process begins with an input protein sequence and a quantifiable fitness assay, and through iterative cycles, autonomously engineers improved proteins [6].

The Integrated Workflow Diagram

The following diagram illustrates the logical flow and component relationships within a generalized autonomous enzyme engineering platform.

Autonomous Protein Engineering DBTL Cycle

Detailed Workflow Description

• Design Phase: The process is initiated by AI models that design a library of protein variants. This often involves unsupervised models like protein Large Language Models (LLMs), such as ESM-2, and epistasis models, such as EVmutation, which predict beneficial mutations based on evolutionary patterns and sequence co-dependencies without requiring prior experimental data for the target enzyme [6] [5]. These models generate a diverse and high-quality initial library, increasing the likelihood of identifying promising mutants.

• Build Phase: The designed sequences are transferred to a biofoundry, such as the Illinois Biological Foundry for Advanced Biomanufacturing (iBioFAB). This phase is fully automated, handling gene synthesis, cloning, and protein expression. A key innovation is a high-fidelity mutagenesis method that achieves approximately 95% accuracy, eliminating the need for intermediate sequence verification and enabling a continuous workflow [6].

• Test Phase: Robotic systems conduct high-throughput functional assays to characterize the expressed protein variants. The platform uses a quantifiable fitness measure, such as enzymatic activity under specific conditions (e.g., neutral pH), to evaluate each variant [6]. This step is modular, allowing for robust operation and easy troubleshooting.

• Learn Phase: The assay data from the test phase is used to train supervised machine learning models (e.g., "low-N" regression models). These models learn the complex relationship between protein sequence and function from the experimental data and predict the next, potentially improved, set of variants to test, often by combining beneficial mutations [6] [5]. The resulting data closes the loop, informing the next Design phase.

Application Notes: Performance and Quantitative Results

The efficacy of autonomous platforms is demonstrated by their application to engineer specific enzymes with remarkable speed and efficiency. The table below summarizes key performance metrics from a landmark study that engineered two distinct enzymes in parallel.

Table 1: Performance Metrics of Autonomous Enzyme Engineering Campaigns

Engineered Enzyme	Engineering Goal	Key Improvement	Experimental Effort	Timeframe
Arabidopsis thaliana Halide Methyltransferase (AtHMT)	Improve ethyltransferase activity and substrate preference	~16-fold increase in ethyltransferase activity; ~90-fold shift in substrate preference [6] [5]	4 rounds; <500 variants screened [6]	4 weeks [6] [7]
Yersinia mollaretii Phytase (YmPhytase)	Enhance activity at neutral pH	~26-fold higher specific activity at neutral pH [6] [5]	4 rounds; <500 variants screened [6]	4 weeks [6] [7]

These results highlight the platform's generality—it requires only a protein sequence and a defined fitness assay—and its exceptional efficiency in navigating vast sequence spaces with minimal experimental effort [6]. In another demonstration, an industrial automated platform, iAutoEvoLab, was used for continuous evolution, successfully engineering a functional T7 RNA polymerase fusion protein with mRNA capping properties from an inactive precursor [8].

Experimental Protocols

Protocol 1: Automated Build Phase for Variant Construction

This protocol details the automated construction of plasmid DNA for protein variant expression on a biofoundry, as implemented for the engineering of AtHMT and YmPhytase [6].

• Objective: To reliably construct plasmid libraries for protein variant expression without the need for intermediate sequence verification, enabling a continuous workflow.
• Principle: A high-fidelity, HiFi-assembly based mutagenesis method is used to generate variants with high accuracy (~95%) [6].
• Materials:
  • Robotic Liquid Handler: (e.g., integrated via Thermo Momentum software).
  • PCR Thermocyclers.
  • DpnI Restriction Enzyme: For digesting template DNA.
  • E. coli Strain: For transformation.
  • Omnitray LB Plates: For plating transformations.
• Procedure:
  • Mutagenesis PCR Setup: The robotic liquid handler prepares PCR reactions in 96-well format using primers designed for the target mutations.
  • Template Digestion: DpnI enzyme is added to the PCR product to digest the methylated template plasmid DNA.
  • DNA Assembly: The digested product is used in a HiFi DNA assembly reaction.
  • Transformation: The assembly reaction is transformed into competent E. coli cells via a high-throughput microbial transformation protocol.
  • Plating: The transformation is plated on Omnitray LB plates with appropriate antibiotics using the robotic system.
  • Colony Picking & Plasmid Purification: Colonies are automatically picked and inoculated into deep-well plates for growth, followed by automated plasmid purification.

Protocol 2: Automated Test Phase for Phytase Activity Assay

This protocol describes a high-throughput assay for measuring phytase activity at neutral pH, used to screen YmPhytase variants [6].

• Objective: To identify YmPhytase variants with enhanced activity at neutral pH.
• Principle: Phytase hydrolyzes phytic acid to release inorganic phosphate, which can be quantified colorimetrically.
• Materials:
  • Assay Buffer (Neutral pH): e.g., Phosphate-buffered saline or Tris-HCl buffer, pH 7.0-7.5.
  • Substrate: Phytic acid solution.
  • Stop/Detection Reagent: Acidified molybdate reagent (e.g., malachite green) for phosphate detection.
  • Microplate Reader: For measuring absorbance.
• Procedure:
  • Lysate Preparation: In a 96-well plate, crude cell lysate containing expressed phytase variants is prepared via automated lysis.
  • Reaction Initiation: The assay buffer containing phytic acid substrate is dispensed into the lysate plate to start the enzymatic reaction.
  • Incubation: The reaction plate is incubated at a controlled temperature (e.g., 37°C) for a fixed time.
  • Reaction Termination: The stop/detection reagent is added to quench the reaction and develop color.
  • Absorbance Measurement: The plate is transferred to a microplate reader to measure absorbance at a suitable wavelength (e.g., 655 nm for malachite green).
  • Data Analysis: Phosphate concentration (and thus enzyme activity) is calculated from the standard curve. Data is automatically fed into the ML model for the Learn phase.

The Scientist's Toolkit: Research Reagent Solutions

The implementation of autonomous protein engineering relies on a suite of core reagents and computational tools. The following table catalogs essential components for establishing such a platform.

Table 2: Key Research Reagents and Tools for Autonomous Protein Engineering

Item Name	Type	Function / Application
ESM-2	Computational Model / Software	A protein large language model (LLM) used for zero-shot prediction of beneficial mutations based on evolutionary patterns in protein sequences [6] [5].
EVmutation	Computational Model / Software	An epistasis model that identifies co-evolving residues in protein sequences to guide the design of mutant libraries with higher functional potential [6].
iBioFAB	Hardware / Platform	A fully automated biofoundry that integrates robotic arms, liquid handlers, and incubators to execute the Build and Test phases of the DBTL cycle without human intervention [6] [7].
HiFi DNA Assembly Mix	Laboratory Reagent	A high-fidelity enzyme mix used for seamless and accurate assembly of multiple DNA fragments, crucial for the automated construction of variant libraries [6].
OrthoRep System	Molecular Biology Tool	A continuous in vivo evolution system used in some platforms (e.g., iAutoEvoLab) for growth-coupled selection and evolution of proteins over long trajectories [8].
Cyclo(L-Leu-trans-4-hydroxy-L-Pro)	Cyclo(L-Leu-trans-4-hydroxy-L-Pro), CAS:115006-86-5, MF:C11H18N2O3, MW:226.27 g/mol	Chemical Reagent
Azido-C6-OH	Azido-C6-OH, CAS:146292-90-2, MF:C6H13N3O, MW:143.19 g/mol	Chemical Reagent

The integration of AI, robotics, and data represents a paradigm shift in protein engineering. Autonomous platforms have moved from concept to proven technology, capable of outperforming traditional methods in speed, efficiency, and scalability. By closing the DBTL loop, these systems mitigate the primary bottleneck of human-dependent design and analysis. As the underlying AI models, such as protein LLMs, become more powerful and robotic systems more accessible, the democratization and broad application of this technology across biotechnology, medicine, and sustainable chemistry is imminent. The future of protein engineering lies in self-driving laboratories that continuously and autonomously explore the frontiers of protein function.

The Self-driving Autonomous Machines for Protein Landscape Exploration (SAMPLE) platform represents a transformative approach to protein engineering that fully automates the design-test-learn cycle. This platform integrates artificial intelligence with robotic laboratory systems to navigate protein fitness landscapes without human intervention, dramatically accelerating the process of engineering proteins with enhanced properties [9]. Traditional protein engineering remains slow, labor-intensive, and inefficient, limited by human cognitive constraints and manual laboratory processes. SAMPLE addresses these limitations by creating a closed-loop system where an intelligent agent learns sequence-function relationships, designs new proteins, and experimentally tests them through automated robotics [9]. This case study examines the SAMPLE platform's application in engineering thermally stable glycoside hydrolase enzymes, detailing the methodologies, outcomes, and practical protocols that demonstrate its capabilities.

Platform Architecture & Workflow

The SAMPLE platform architecture consists of two tightly integrated components: an intelligent software agent that makes computational decisions and a fully automated physical laboratory that executes experiments.

Computational Intelligence Components

The SAMPLE agent employs Bayesian optimization (BO) as its core decision-making framework to efficiently navigate the protein sequence space. This approach is specifically designed to balance exploration of unknown regions of the fitness landscape with exploitation of promising areas already identified [9]. The agent utilizes a multi-output Gaussian process (GP) model that simultaneously predicts whether a protein sequence will be functional and estimates its thermostability (T50), effectively modeling both the continuous fitness landscape and the "holes" representing non-functional sequences [9].

Two specialized BO methods were developed to enhance sampling efficiency:

• UCB Positive Method: Considers only sequences predicted to be functional (Pactive > 0.5) and selects those with the highest Upper Confidence Bound value
• Expected UCB Method: Calculates expected UCB by multiplying the UCB value by the GP classifier's Pactive, focusing sampling on sequences that are both likely to be functional and have high predicted fitness [9]

Benchmarking on cytochrome P450 data demonstrated that these methods could identify thermostable variants with only 26 measurements on average, representing a 3-4 fold improvement over standard UCB or random sampling approaches [9].

Robotic Laboratory Integration

The physical laboratory system provides seamless experimental execution of the agent's designs. The platform implements a streamlined, general pipeline for automated gene assembly, protein expression, and biochemical characterization that takes approximately 9 hours from protein design to functional data [9]. Multiple layers of exception handling and data quality control ensure reliability: the system verifies successful gene assembly via double-stranded DNA detection, validates enzyme reaction progress curves, and confirms activity above background levels before accepting experimental results [9].

Table: SAMPLE Platform Automated Workflow Timeline

Process Stage	Time Required	Key Operations
Gene Assembly	1 hour	Golden Gate cloning of pre-synthesized DNA fragments
PCR Amplification	1 hour	Expression cassette amplification with EvaGreen verification
Protein Expression	3 hours	T7-based cell-free protein production
Thermostability Assay	3 hours	Activity measurement across temperature gradient
Total Cycle Time	9 hours	Fully automated from design to data

SAMPLE Autonomous Control Logic: Diagram illustrating the decision flow and experimental cycle of the SAMPLE platform.

Application Case Study: Engineering Thermostable Glycoside Hydrolases

Experimental Design & Setup

The SAMPLE platform was deployed to engineer glycoside hydrolase family 1 (GH1) enzymes with enhanced thermal tolerance, a valuable property for industrial applications in biofuel production, food processing, and paper manufacturing [9]. Four independent SAMPLE agents were deployed, each starting with the same six natural GH1 sequences as initial seeds. The agents operated in a combinatorial sequence space composed of 1,352 unique GH1 sequences that incorporated natural sequence elements alongside computational designs from Rosetta and evolution-based fragments [9]. This diverse landscape ensured broad sampling of possible functional variants.

Each agent designed three new protein sequences per round based on the Expected UCB acquisition function and ran for 20 consecutive rounds without human intervention. The thermostability metric (T50) was defined as the temperature at which enzyme activity decreased by 50%, providing a robust quantitative measure for the optimization objective [9]. The experimental system demonstrated high reproducibility, with measurement errors of less than 1.6°C for T50 values across technical replicates [9].

Performance Results & Outcomes

All four SAMPLE agents successfully navigated the GH1 fitness landscape and converged on thermostable enzyme variants despite differences in their individual search trajectories. Within 20 rounds of experimentation, all agents identified enzymes with significantly enhanced thermal stability, at least 12°C higher than the starting natural sequences [9]. This improvement was achieved while searching less than 2% of the full combinatorial sequence space, demonstrating exceptional sampling efficiency [9].

The agents exhibited characteristic optimization behavior with early phases dominated by exploratory sampling to understand landscape structure, followed by progressive convergence toward fitness peaks in later rounds. Notably, the system maintained robust performance despite experimental noise and occasional instrumental challenges, with the cloud-based implementation enabling continuous operation despite physical laboratory constraints [10].

Table: SAMPLE Platform Performance Metrics

Performance Indicator	Metric Value	Comparative Benchmark
Experimental Cycle Time	9 hours	3-4 days (manual)
Thermostability Improvement	≥12°C T50 increase	Target-dependent
Search Space Coverage	<2%	Typically 10-20% (directed evolution)
Sampling Efficiency	26 variants (average to find optima)	3-4x better than random
Experimental Reproducibility	<1.6°C error	Industry standard

Detailed Experimental Protocols

Automated Gene Assembly & Verification

Principle: Generate full gene expression cassettes from pre-synthesized DNA fragments via Golden Gate cloning [9].

Procedure:

• Fragment Preparation: Dispense pre-synthesized DNA fragments (15-30bp overlapping ends) into 96-well PCR plates using liquid handling robotics
• Golden Gate Assembly:
  • Combine DNA fragments with T4 DNA ligase (5U/µL) and BsaI restriction enzyme (5U/µL)
  • Add 10X T4 DNA ligase buffer to 1X final concentration
  • Program thermal cycler: 25 cycles of (37°C for 2 minutes, 16°C for 5 minutes)
  • Final incubation: 60°C for 10 minutes, 80°C for 10 minutes
• Expression Cassette Amplification:
  • Transfer 2 µL assembly reaction to 50 µL PCR mix with Q5 High-Fidelity DNA Polymerase
  • Add EvaGreen fluorescent dye (1X final concentration) for real-time detection
  • Amplify with standard PCR protocol: 98°C 30s; 35 cycles of (98°C 10s, 72°C 30s)
• Verification: Confirm successful amplification by monitoring EvaGreen fluorescence signal; failed reactions are flagged and sequences return to design queue

Cell-Free Protein Expression & Characterization

Principle: Directly express proteins from amplified DNA cassettes and measure thermostability via enzymatic activity [9].

Procedure:

• Cell-Free Expression:
  • Combine 5 µL amplified DNA with 45 µL T7-based cell-free expression reagents
  • Incubate at 30°C for 3 hours with orbital shaking (300 rpm)
  • Centrifuge at 4,000 × g for 10 minutes to remove debris
• Thermostability Assay:
  • Prepare 96-well PCR plate with 10 µL expressed protein per well
  • Incubate plates at temperature gradient (30-80°C) for 10 minutes using thermal cycler
  • Cool to 4°C, then add 90 µL substrate solution (1 mM p-nitrophenyl-glycoside in appropriate buffer)
  • Monitor reaction progress at 410 nm for 30 minutes at 25°C
• Data Processing:
  • Fit progress curves to determine initial reaction velocities
  • Normalize activities to maximum observed velocity
  • Fit temperature-activity profile to sigmoid function to calculate T50
  • Flag and exclude experiments with irregular progress curves or low signal-to-noise ratios

Computational Protein Design Protocol

Principle: Employ Bayesian optimization with Gaussian process models to select protein sequences for experimental testing [9].

Procedure:

• Model Initialization:
  • Encode protein sequences using a valid structural representation
  • Initialize multi-output GP with radial basis function kernel
  • Set hyperparameters via maximum likelihood estimation
• Sequential Design:
  • Compute posterior predictive distribution for all candidate sequences
  • Calculate Expected UCB scores: EUCB(x) = μ(x) + βσ(x) × Pactive(x)
  • Select top 3 sequences with highest EUCB scores for experimental testing
  • Set exploration parameter β to 2.0 for balanced exploration-exploitation
• Model Update:
  • Incorporate new experimental data (sequence, T50, activity status)
  • Retrain GP hyperparameters via gradient-based optimization
  • Update internal landscape representation and uncertainty estimates
• Convergence Check:
  • Monitor performance improvement over last 5 rounds
  • Terminate after 20 rounds or when no improvement observed for 5 consecutive rounds

Research Reagent Solutions

Table: Essential Research Reagents for Autonomous Protein Engineering

Reagent / Material	Function	Specifications
Pre-synthesized DNA Fragments	Gene assembly building blocks	15-30bp overlapping ends, HPLC purified
T4 DNA Ligase	DNA fragment joining	5U/µL, high concentration for automation
BsaI Restriction Enzyme	Golden Gate assembly	5U/µL, thermostable
Q5 High-Fidelity DNA Polymerase	PCR amplification	Error rate: ~5 × 10⁻⁶ mutations/bp
EvaGreen Fluorescent Dye	DNA quantification	20X concentrate in DMSO
T7 Cell-Free Expression System	Protein synthesis	Pre-mixed reagents, -80°C storage
p-Nitrophenyl-glycoside	Enzyme activity substrate	1 mM stock in appropriate buffer
96-Well PCR Plates	Reaction vessels	Skirted, clear, automation compatible

Visualization of Protein Fitness Landscape Navigation

Fitness Landscape Navigation: Visualization of SAMPLE agent strategies for exploring and exploiting protein fitness landscapes.

The SAMPLE platform demonstrates that fully autonomous protein engineering is not only feasible but exceptionally effective at navigating complex biological design spaces. By completing the design-test-learn cycle in just 9 hours without human intervention, SAMPLE achieves a dramatic acceleration compared to traditional manual approaches [9]. The platform's ability to efficiently explore vast sequence spaces while requiring minimal sampling highlights the power of Bayesian optimization combined with robotic automation.

This case study establishes a framework for generalized autonomous experimentation in synthetic biology and protein engineering. The methodologies detailed here can be adapted to diverse protein engineering challenges beyond thermostability, including substrate specificity, catalytic efficiency, and pH stability optimization. As autonomous platforms like SAMPLE become more accessible and robust, they promise to transform protein engineering from a specialized, labor-intensive craft into a systematic, data-driven discipline capable of addressing pressing challenges in medicine, biotechnology, and sustainable energy.

The Protein Fitness Landscape and the Challenge of Navigation

The concept of the protein fitness landscape, first introduced by Sewall Wright, provides a powerful framework for understanding the relationship between a protein's sequence and its function [11]. In this analogy, the landscape is composed of peaks and valleys, where fitness peaks represent high-performing sequences and valleys correspond to suboptimal or non-functional variants [9]. Navigating this vast, high-dimensional landscape represents one of the most significant challenges in modern protein engineering, as the number of possible sequences for a typical protein far exceeds what can be experimentally tested.

Traditional approaches to protein engineering, such as directed evolution (DE), operate as empirical hill-climbing processes on this landscape [12]. While successful for incremental improvements, these methods are inherently limited when facing epistatic interactions (non-additive effects of mutations) and rugged terrain, which can trap exploration at local optima [12]. The emergence of autonomous laboratories represents a paradigm shift, combining artificial intelligence (AI), machine learning (ML), and robotic automation to create self-driving platforms capable of navigating the fitness landscape with unprecedented efficiency [9] [5] [6]. This Application Note details the operational protocols and core components of these systems, with a specific focus on the SAMPLE platform, framing them within the broader context of autonomous protein engineering.

The Autonomous Navigation Platform: Core Architecture

Autonomous platforms for protein engineering are built upon a closed-loop Design-Build-Test-Learn (DBTL) cycle, where intelligent agents make sequential decisions to explore and exploit the fitness landscape without human intervention. The SAMPLE (Self-driving Autonomous Machines for Protein Landscape Exploration) platform exemplifies this architecture [9].

The DBTL Cycle and SAMPLE Workflow

The following diagram illustrates the integrated, autonomous DBTL cycle as implemented in the SAMPLE platform.

• Design: The cycle begins with an intelligent agent that uses a Gaussian Process (GP) model to build a probabilistic understanding of the sequence-function relationship from limited data. To navigate the trade-off between exploring unknown regions and exploiting known promising areas, the agent employs a Bayesian Optimization (BO) strategy, such as the "Expected UCB" method, which focuses sampling on sequences predicted to be both functional and high-performing [9].
• Build: Designed protein sequences are sent to a fully automated robotic system. The SAMPLE platform, for instance, uses a streamlined pipeline of Golden Gate cloning for gene assembly, followed by PCR amplification and direct addition to cell-free protein expression systems [9].
• Test: The expressed proteins are characterized using automated, enzyme-specific biochemical assays. For thermostability engineering, the platform measures the T50 value (the temperature at which 50% of the activity is lost) with high reproducibility (error < 1.6°C) [9].
• Learn: Data from the experiments are used to update the GP model, refining the agent's internal representation of the fitness landscape. This step closes the loop, allowing the system to learn from empirical results and plan more informative experiments in the next cycle [9].

Key Experimental Protocol: Autonomous Navigation of a Glycoside Hydrolase Landscape

A specific deployment of SAMPLE involved engineering glycoside hydrolase (GH1) enzymes for enhanced thermal tolerance [9].

• Objective: To autonomously discover thermostable GH1 variants.
• Combinatorial Space: The platform searched a designed combinatorial space of 1,352 unique GH1 sequences, which introduced an average of 116 mutations compared to the starting sequences [9].
• Agent Configuration: Four independent SAMPLE agents were deployed, each seeded with the same six natural GH1 sequences.
• Execution:
  • Each agent used the Expected UCB acquisition function for sequence selection.
  • In each of 20 rounds, an agent selected three sequences for experimental testing.
  • The fully automated Strateos Cloud Lab platform executed the build-and-test process, with a total turnaround time of approximately 9 hours from design to data [9].
• Outcome: All four agents successfully converged on thermostable enzymes that were at least 12°C more stable than the starting sequences, despite searching less than 2% of the total landscape [9]. This demonstrates the remarkable sample efficiency of the autonomous approach.

Quantitative Performance of Autonomous Platforms

The performance of autonomous platforms can be quantified by their efficiency and effectiveness in optimizing protein function. The table below summarizes key results from the SAMPLE platform and another generalized AI-powered platform.

Table 1: Performance Metrics of Autonomous Protein Engineering Platforms

Platform / System	Target Protein	Engineering Goal	Key Experimental Result	Efficiency & Scale
SAMPLE Platform [9]	Glycoside Hydrolase (GH1)	Enhance thermal tolerance	Variants with ≥12°C improved thermostability	20 rounds, 4 agents, <2% of landscape searched
Generalized AI Platform [6]	Yersinia mollaretii Phytase (YmPhytase)	Improve activity at neutral pH	Variant with 26-fold higher specific activity	4 weeks, 4 rounds, <500 variants tested
Generalized AI Platform [6]	Arabidopsis thaliana Halide Methyltransferase (AtHMT)	Improve ethyltransferase activity & substrate preference	16-fold higher ethyltransferase activity; 90-fold shift in substrate preference	4 weeks, 4 rounds, <500 variants tested

The Scientist's Toolkit: Research Reagent Solutions

Implementing an autonomous protein engineering campaign requires a suite of specialized reagents, software, and hardware. The following table details essential components as used in platforms like SAMPLE and the generalized AI platform from iBioFAB.

Table 2: Essential Research Reagents and Tools for Autonomous Protein Engineering

Tool / Reagent Category	Specific Example	Function & Application
Machine Learning Models	Gaussian Process (GP) with Bayesian Optimization [9]	Models sequence-function relationships and decides which variants to test next in the SAMPLE platform.
Machine Learning Models	Protein Language Model (e.g., ESM-2) [6]	An unsupervised model used to design high-quality, diverse initial libraries based on evolutionary principles.
Machine Learning Models	Epistasis Model (e.g., EVmutation) [6]	Predicts the effect of mutations in the context of other mutations, helping to model non-additive effects.
DNA Construction & Cloning	Golden Gate Assembly [9]	A robust, automated method for assembling pre-synthesized DNA fragments into full genes.
DNA Construction & Cloning	High-Fidelity (HiFi) Mutagenesis [6]	A method achieving ~95% accuracy in site-directed mutagenesis, enabling continuous workflow without intermediate sequencing.
Protein Expression System	Cell-Free Protein Expression [9]	Allows for rapid protein synthesis directly from a DNA template, bypassing the need for cell culture.
Automation Hardware	Robotic Cloud Lab (e.g., Strateos) [9]	A fully integrated, remote-operated system that automates liquid handling, incubation, and assay measurements.
Automation Hardware	Illinois Biological Foundry (iBioFAB) [6]	An industrial-grade automated system for end-to-end execution of biological workflows, from DNA construction to assay.
Safety & Screening	Safety Protocol (MMSeqs2, FoldSeek) [13]	Scans designed protein sequences against databases of harmful proteins based on sequence and structural homology to mitigate risk.
CB-184	CB-184, MF:C22H21Cl2NO2, MW:402.3 g/mol	Chemical Reagent
Glimepiride sulfonamide	Glimepiride sulfonamide, CAS:119018-29-0, MF:C16H21N3O4S, MW:351.4 g/mol	Chemical Reagent

Advanced Applications and Protocol Extensions

The principles of autonomous navigation are being extended to tackle increasingly complex challenges in protein science.

Machine Learning-Assisted Directed Evolution (MLDE)

For laboratories not yet equipped for full autonomy, ML-assisted directed evolution (MLDE) offers a powerful intermediate step. This approach uses supervised ML models trained on experimental data to predict the fitness of unsampled variants, dramatically improving the efficiency of traditional directed evolution.

A comprehensive evaluation across 16 diverse combinatorial landscapes revealed that MLDE strategies consistently matched or exceeded the performance of standard DE, with the advantage becoming more pronounced on rugged landscapes with high epistasis and fewer active variants [12]. Focused training (ftMLDE), which uses zero-shot predictors to enrich initial training sets with higher-fitness variants, was shown to further accelerate the discovery of optimal sequences [12].

Fitness Landscape Design (FLD)

A frontier in the field is the move from navigating existing landscapes to actively designing them. Fitness Landscape Design (FLD) is an inverse approach that seeks to computationally define a target fitness landscape—for example, one that suppresses the fitness of viral escape variants—and then discover molecular interventions (like antibodies) that reshape the natural landscape to match the target [11].

The FLD-with-Antibodies (FLD-A) protocol involves:

• Biophysical Modeling: Deriving a model that maps viral protein sequence to fitness based on binding affinities to host receptors and antibodies [11].
• Designability Analysis: Determining the subset of possible fitness assignments to different genotypes that is physically realizable by an antibody repertoire [11].
• Inverse Design: Using stochastic optimization to discover antibody ensembles that force viral evolution onto the user-defined, low-fitness target landscape [11].

The challenge of navigating the protein fitness landscape is being met by a new generation of autonomous platforms. By integrating intelligent agents that learn and decide with robotic systems that build and test, platforms like SAMPLE close the DBTL loop without human intervention. The resulting systems are highly efficient, as demonstrated by their ability to discover significantly improved enzymes while sampling only a tiny fraction of the possible sequence space. The protocols and tools detailed in this Application Note provide a roadmap for researchers aiming to leverage these technologies, from fully autonomous laboratories to ML-enhanced directed evolution. As the field progresses towards the active design of fitness landscapes, the potential to not only navigate but also to sculpt the evolutionary terrain itself promises to revolutionize protein engineering and therapeutic design.

Inside the Machine: AI and Automated Workflows in Action

The Design-Build-Test-Learn (DBTL) Cycle as an Automated Loop

The Design-Build-Test-Learn (DBTL) cycle represents the cornerstone engineering framework in synthetic biology and protein engineering. In traditional implementations, each stage requires significant human intervention, judgment, and domain expertise, creating bottlenecks that limit the pace of discovery and optimization. The emergence of autonomous experimentation systems has transformed this iterative process into a continuous, self-driving loop that dramatically accelerates protein engineering campaigns. By integrating artificial intelligence (AI), robotic automation, and biofoundry infrastructures, these platforms can execute multiple DBTL cycles with minimal human supervision, achieving in weeks what previously required months or years of manual effort [6].

This paradigm shift is particularly impactful for protein engineering, where the sequence-function landscape is vast and complex. Autonomous DBTL platforms address this challenge by combining machine learning for intelligent design, automated workstations for high-throughput construction and testing, and data analysis pipelines for rapid learning. Recent demonstrations have achieved remarkable results, including engineering enzymes with 90-fold improvement in substrate preference and 26-fold enhancement in activity at targeted pH conditions within just four weeks and fewer than 500 variants tested [6]. This document details the protocols, components, and workflows that enable such accelerated engineering in autonomous protein engineering platforms.

Core Components of an Autonomous DBTL Framework

The Integrated Workflow

Autonomous DBTL implementation requires tight integration of computational and physical components. The process begins with an input protein sequence and a quantifiable fitness objective, concluding with characterized variants and updated AI models that inform subsequent cycles [6].

Diagram 1: Autonomous DBTL Workflow for Protein Engineering

Essential Research Reagent Solutions

Successful implementation of autonomous DBTL cycles depends on specialized reagents and materials that enable high-throughput, reproducible operations. The following table details key solutions required for automated protein engineering platforms.

Table 1: Essential Research Reagent Solutions for Autonomous Protein Engineering

Category	Specific Solution	Function in Workflow	Application Example
DNA Construction	HiFi Assembly Mix [6]	Enables highly accurate DNA assembly without intermediate sequencing verification	Modular plasmid construction for variant libraries
	Golden Gate Assembly with ccdB suicide gene [14]	Provides high cloning accuracy (~90%) without need for colony picking	Sequencing-free cloning in SAPP workflow
Cell-Based Systems	Auto-induction Media [14]	Eliminates manual induction step in protein expression	High-throughput protein production in 96-well formats
	OrthoRep Continuous Evolution System [8]	Enables continuous in vivo evolution with growth-coupled selection	Engineering complex protein functions
Cell-Free Systems	Cell-Free Expression Lysates [15]	Rapid protein synthesis without cloning; enables toxic protein production	Megascale data generation for machine learning
Screening Assays	Functional Enzyme Assays [6]	Quantitatively measures variant fitness in high-throughput format	Screening methyltransferase and phytase activity
Analysis Tools	SEC Chromatography Analysis Software [14]	Automates analysis of thousands of chromatograms for purity, yield, oligomeric state	Standardized data output in SAPP platform

Protocol: Implementation of an Autonomous DBTL Cycle

Stage 1: AI-Driven Protein Design

Objective

Generate a diverse, high-quality initial variant library using AI models to maximize the probability of identifying beneficial mutations while minimizing library size.

Materials

• Input protein sequence (FASTA format)
• Access to computational resources (GPU-enabled workstation or cluster)
• Protein language models (ESM-2 [6] [15] or equivalent)
• Epistasis models (EVmutation [6] or equivalent)
• Fitness prediction algorithms

Procedure

• Sequence Analysis: Input the wild-type protein sequence into the ESM-2 protein language model to calculate amino acid probabilities at each position based on evolutionary context [6].
• Epistasis Modeling: Process the same sequence through EVmutation to identify co-evolutionary patterns and epistatic interactions from homologous sequences [6].
• Variant Scoring: Combine predictions from both models to generate a ranked list of single-point mutations with predicted fitness scores.
• Library Design: Select 150-200 variants for the initial library, ensuring coverage of diverse mutation sites and types [6].
• DNA Sequence Output: Convert selected amino acid variants to DNA sequences with codon optimization for the expression host.

Technical Notes

• Balance exploration (diverse mutations) and exploitation (high-scoring mutations) in library design.
• For subsequent cycles, incorporate experimental data to train supervised machine learning models for fitness prediction [6].
• Consider structural constraints using tools like ProteinMPNN when structural data is available [15].

Stage 2: Automated Build Processes

Objective

Efficiently convert digital variant designs into physical DNA constructs and expressed proteins with minimal manual intervention.

Materials

• Oligo pools or synthesized DNA fragments
• HiFi DNA assembly master mix
• Automated liquid handling system (e.g., iBioFAB [6])
• 96-well deep-well plates
• Transformation-competent cells
• Auto-induction media [14]

Procedure

• DNA Assembly Setup:

  • Program liquid handler to dispense DNA fragments and assembly mix into 96-well PCR plates
  • Execute HiFi assembly reaction: 50°C for 60 minutes [6]
  • Perform DpnI digestion to eliminate template DNA

• High-Throughput Transformation:

  • Transfer assembly reactions to 96-well transformation plates containing competent cells
  • Execute heat-shock transformation protocol
  • Plate transformations on 8-well omnitray LB plates using robotic plating system

• Culture and Expression:

  • Pick colonies using automated colony picker into 96-deep well plates containing auto-induction media
  • Incubate with shaking for protein expression (24-48 hours, based on target protein)

• Quality Control:

  • Randomly select and sequence 5-10% of variants to verify assembly accuracy (expected >95% [6])

Technical Notes

• The HiFi assembly method achieves ~95% accuracy, eliminating need for sequence verification during iterative cycles [6].
• For cost-effective DNA construction, consider DMX workflow using oligo pools and nanopore sequencing, which reduces DNA synthesis costs by 5-8 fold [14].
• Modular workflow design enables error recovery without restarting entire process.

Stage 3: High-Throughput Testing & Characterization

Objective

Rapidly quantify fitness parameters for all library variants to generate high-quality datasets for machine learning.

Materials

• Cell lysates or purified protein variants
• Microplate readers with kinetic capability
• Assay-specific substrates and reagents
• Automated liquid handling systems
• Size-exclusion chromatography (SEC) plates [14]

Procedure

• Sample Preparation:

  • Use robotic systems to prepare crude cell lysates from expression cultures
  • Alternatively, implement automated purification (nickel-affinity and SEC) for selected variants [14]

• Assay Implementation:

  • Program liquid handler to dispense lysates and assay reagents into 96- or 384-well plates
  • For methyltransferase activity: Measure methyltransferase or ethyltransferase activity using appropriate substrates (e.g., methyl iodide vs. ethyl iodide) [6]
  • For phytase activity: Quantify phosphate release at target pH using colorimetric assay [6]
  • Record kinetic measurements using plate readers

• Data Collection:

  • Measure fluorescence/absorbance at appropriate intervals
  • Calculate enzymatic rates normalized to protein concentration or cell density
  • Export data in standardized format for machine learning analysis

Technical Notes

• Implement cell-free screening systems for ultra-high-throughput (>>10,000 variants) using droplet microfluidics [15].
• SEC analysis simultaneously provides data on purity, yield, oligomeric state, and dispersity [14].
• Assay conditions should reflect final application requirements (e.g., specific pH for phytase activity [6]).

Stage 4: Machine Learning-Driven Learning

Objective

Extract meaningful patterns from experimental data to improve variant predictions in subsequent cycles.

Materials

• Experimental dataset (variant sequences and fitness measurements)
• Machine learning workstation with adequate computational resources
• ML libraries (scikit-learn, PyTorch, TensorFlow)
• Data visualization tools

Procedure

• Data Preprocessing:

  • Normalize fitness metrics across plates and batches
  • Encode protein variants as numerical features (one-hot encoding, physicochemical properties)
  • Split data into training (80%) and validation (20%) sets

• Model Training:

  • Train supervised machine learning models (e.g., random forest, gradient boosting, neural networks) on sequence-fitness data [6]
  • For small datasets (<1000 variants), employ "low-N" machine learning methods optimized for limited data [6]
  • Validate model performance using cross-validation and hold-out test sets

• Variant Prediction:

  • Use trained model to predict fitness of unseen variants in the sequence space
  • Select top predictions for next DBTL cycle, balancing exploration and exploitation
  • Generate hypotheses about sequence-function relationships

Technical Notes

• For initial cycles with limited data, combine supervised models with unsupervised protein language model predictions [6].
• Prioritize interpretable models when seeking biological insights into sequence-function relationships.
• The entire DBTL cycle should require construction and testing of fewer than 500 variants to achieve significant improvements [6].

Performance Metrics and Case Studies

Quantitative Performance of Autonomous DBTL Platforms

Recent implementations of autonomous DBTL cycles have demonstrated remarkable efficiency and efficacy in protein engineering campaigns. The following table summarizes quantitative results from published studies.

Table 2: Performance Metrics of Autonomous DBTL Platforms in Protein Engineering

Engineering Target	Key Objective	DBTL Cycles	Timescale	Variants Tested	Performance Improvement
Arabidopsis thaliana halide methyltransferase (AtHMT) [6]	Improve ethyltransferase activity and substrate preference	4 rounds	4 weeks	<500	90-fold improvement in substrate preference; 16-fold improvement in ethyltransferase activity
Yersinia mollaretii phytase (YmPhytase) [6]	Enhance activity at neutral pH	4 rounds	4 weeks	<500	26-fold improvement in activity at neutral pH
Respiratory Syncytial Virus (RSV) neutralizer [14]	Develop multivalent antiviral proteins	1 round (SAPP platform)	1 week	58 multimeric constructs	IC50 of 40 pM (dimer) and 59 pM (trimer) vs. 5.4 nM (monomer)
Fluorescent Protein variants [14]	Improve yield, stability, and optical properties	1 round (SAPP platform)	1 week	96 variants	Identified designs with enhanced thermal stability and altered optical properties

Emerging Paradigm: LDBT Cycle

A paradigm shift is emerging in which the Learning phase precedes Design (LDBT), leveraging large datasets and pre-trained models to make accurate zero-shot predictions [15]. This approach utilizes:

• Protein language models (ESM, ProGen) trained on evolutionary sequences [15]
• Structure-based models (ProteinMPNN, RFdiffusion) for functional design [15]
• Megascale data generation through cell-free systems and ultra-high-throughput screening [15]

This reordering potentially enables a "Design-Build-Work" model where initial designs frequently function as intended, reducing or eliminating iterative cycling [15].

The transformation of the DBTL cycle into an automated, autonomous loop represents a fundamental advancement in protein engineering methodology. By integrating AI-guided design, automated biofoundry workflows, and machine learning-driven learning, these systems achieve unprecedented efficiency in navigating complex protein sequence spaces. The protocols and case studies presented demonstrate that autonomous DBTL platforms can produce order-of-magnitude improvements in enzyme function within weeks rather than years, dramatically accelerating the development of proteins for therapeutic, industrial, and research applications. As these platforms become more accessible and standardized through frameworks like the biofoundry abstraction hierarchy [16], autonomous protein engineering is poised to become the dominant paradigm for biotechnology innovation.

Performance Benchmarks in Autonomous Platforms

The integration of Bayesian Optimization (BO) and Protein Language Models (PLMs) like ESM-2 into autonomous platforms has demonstrated remarkable efficiency and effectiveness in protein engineering campaigns. The table below summarizes quantitative outcomes from recent studies.

Table 1: Performance Metrics of AI-Driven Protein Engineering Platforms

Platform / System	Key AI Components	Target Protein(s)	Engineering Goal	Key Quantitative Results	Experimental Scale & Duration
PLMeAE [17]	PLM (ESM-2), Multi-layer Perceptron	tRNA synthetase (pCNF-RS)	Improve enzyme activity	- Up to 2.4-fold improvement in enzyme activity [17].	- 4 rounds in 10 days.- 96 variants per round [17].
Generalized AI Platform [6]	Protein LLM (ESM-2), Epistasis Model, Low-N ML model	Halide Methyltransferase (AtHMT)	Improve substrate preference & ethyltransferase activity	- 90-fold improvement in substrate preference.- 16-fold improvement in ethyltransferase activity [6].	- 4 rounds over 4 weeks.- Fewer than 500 variants total [6].
Generalized AI Platform [6]	Protein LLM (ESM-2), Epistasis Model, Low-N ML model	Phytase (YmPhytase)	Improve activity at neutral pH	- 26-fold improvement in activity at neutral pH [6].	- 4 rounds over 4 weeks.- Fewer than 500 variants total [6].
SAMPLE [18]	Gaussian Process Model, Bayesian Optimization	Glycoside Hydrolase (GH1) enzymes	Enhance thermal tolerance (T50)	- Identified enzymes >12 °C more stable than starting sequences.- 83% accuracy in active/inactive classification.- ~26 measurements on average to find thermostable variants in simulation [18].	- 20 rounds of autonomous experimentation.- Searched <2% of the full sequence landscape [18].

Detailed Experimental Protocols

Protocol: Bayesian Optimization for Navigating Protein Fitness Landscapes

This protocol is adapted from the SAMPLE platform for the autonomous engineering of protein thermostability using Bayesian Optimization [18].

1. Problem Formulation and Agent Setup

• Define the Objective: Clearly specify the protein property to be optimized (e.g., thermostability, expressed as T50).
• Initialize the Agent: Seed the BO agent with a small set (e.g., 6) of known protein sequences and their experimentally measured fitness values.
• Configure the Model: Employ a multi-output Gaussian Process (GP) model. This model simultaneously performs two tasks:
  • A classifier that predicts the probability that a given sequence is stably folded and functional (P_active).
  • A regressor that predicts the continuous fitness value (e.g., T50) for sequences deemed active.

2. Iterative Design-Build-Test-Learn Cycle

• Design Phase:
  • The agent uses an acquisition function to select the next sequences for experimental testing. The "Expected UCB" method is recommended [18].
  • Calculation: Expected UCB = (Predictive Mean + β * Predictive Uncertainty) * P_active, where β is a parameter balancing exploration and exploitation.
  • The top sequences (e.g., 3 per round) with the highest Expected UCB scores are selected for testing.
• Build Phase (Automated):
  • Gene Assembly: Use a robust, automated method like Golden Gate cloning to assemble genes from pre-synthesized DNA fragments [18].
  • Protein Expression: Employ a cell-free protein expression system for rapid protein production without the need for cell culture [18].
• Test Phase (Automated):
  • Perform a colorimetric or fluorescent biochemical assay to measure the protein's activity.
  • For thermostability (T50), measure residual activity after a heat gradient treatment and fit the data to a sigmoid curve to determine the T50 value [18].
  • Implement automated data quality checks (e.g., successful PCR, valid reaction progress curves, activity above background).
• Learn Phase:
  • The new sequence-fitness data is added to the training dataset.
  • The multi-output GP model is retrained on this updated dataset to refine its understanding of the fitness landscape.

3. Termination

• The cycle repeats until a predefined fitness threshold is met, a set number of rounds is completed, or performance plateaus.

Protocol: Protein Language Model for Zero-Shot Variant Design

This protocol outlines the use of PLMs like ESM-2 for the initial design of protein variants, as implemented in the PLMeAE platform [17]. It consists of two modules.

Module I: Engineering Proteins Without Previously Identified Mutation Sites

• Input: The wild-type amino acid sequence.
• Procedure:
  • Systematic Masking: Individually mask each amino acid position in the wild-type sequence.
  • PLM Inference: For each masked position, use the PLM (e.g., ESM-2) to calculate the likelihood (probability) of all possible single-residue substitutions.
  • Variant Ranking: Rank all possible single mutants based on the predicted likelihood, interpreting a higher likelihood as a proxy for higher fitness and stability.
  • Candidate Selection: Select the top-ranked variants (e.g., top 96) for experimental characterization to validate the predictions and identify beneficial mutations [17].

Module II: Engineering Proteins With Known Mutation Sites

• Input: The wild-type sequence and a set of pre-defined target sites (e.g., from structural analysis or prior experiments).
• Procedure:
  • Multi-site Masking: Simultaneously mask all the specified target sites in the wild-type sequence.
  • Combinatorial Sampling: Use the PLM to sample or predict high-likelihood amino acid combinations across the masked positions, generating a library of multi-mutant variants.
  • Library Construction: The proposed library is synthesized and tested using an automated biofoundry [17].

Workflow and System Architecture Visualization

Autonomous DBTL Cycle for Protein Engineering

The following diagram illustrates the closed-loop, autonomous Design-Build-Test-Learn (DBTL) cycle that integrates AI and laboratory automation.

Bayesian Optimization with a Multi-Output Gaussian Process

This diagram details the Bayesian Optimization process used by the SAMPLE agent to select sequences for testing.

The Scientist's Toolkit: Research Reagent Solutions

Table 2: Essential Research Reagents and Platforms for AI-Driven Protein Engineering

Item / Platform Name	Type	Primary Function in Workflow
ESM-2 (Evolutionary Scale Modeling) [17] [6]	Protein Language Model (PLM)	A transformer-based model trained on millions of protein sequences. Used for zero-shot prediction of high-fitness single and multi-mutants by learning evolutionary constraints [17] [6].
Gaussian Process (GP) Model [18]	Machine Learning Model	Serves as a probabilistic surrogate model in Bayesian Optimization. It models the protein fitness landscape, predicting the mean and uncertainty of fitness for unexplored sequences [18].
Automated Biofoundry (e.g., iBioFAB) [6]	Robotic Laboratory System	Integrated robotic system that automates the "Build" and "Test" phases of the DBTL cycle, including DNA assembly, transformation, protein expression, and functional assays with high reproducibility [6].
Cell-Free Protein Expression System [18]	Protein Synthesis Kit	Enables rapid in vitro protein synthesis without the need for cell culture, significantly speeding up the "Test" phase in platforms like SAMPLE [18].
Golden Gate Cloning [18]	DNA Assembly Method	A robust and efficient DNA assembly method used in automated workflows to construct protein variant genes from pre-synthesized DNA fragments [18].
ABZ-amine	ABZ-amine, CAS:80983-36-4, MF:C10H13N3S, MW:207.30 g/mol	Chemical Reagent
Diclofenac Amide-13C6	1-(2,6-Dichlorophenyl)-2-indolinone|CAS 15362-40-0	Research-use 1-(2,6-Dichlorophenyl)-2-indolinone, a key intermediate and impurity standard. This product is for research purposes only and is not intended for personal use.

The advent of autonomous protein engineering platforms, such as the Self-driving Autonomous Machine for Protein Landscape Exploration (SAMPLE), represents a paradigm shift in biological design [17] [14]. These systems leverage iterative Design-Build-Test-Learn (DBTL) cycles to navigate protein fitness landscapes with minimal human intervention. A critical physical component that enables this autonomy is the robotic biofoundry—an integrated facility that uses robotic automation and computational analytics to streamline and accelerate synthetic biology workflows [19] [20]. This application note details the core protocols and methodologies that underpin the automated gene synthesis, expression, and assaying processes within such biofoundries, providing a practical framework for their implementation in the context of advanced protein engineering research.

Automated Workflow Modules of a Robotic Biofoundry

The transition from a digital protein sequence to an empirically characterized variant is executed through a series of automated, interconnected modules. The following sections provide detailed protocols for these core operations.

Module 1: Automated DNA Construction and Clone Generation

Objective: To reliably and rapidly construct sequence-verified plasmid DNA for protein variant expression.

Background: Overcoming the DNA synthesis bottleneck is crucial for high-throughput campaigns. Traditional methods involving site-directed mutagenesis and sequence verification introduce significant delays [1]. The following protocol describes a high-fidelity assembly method that minimizes the need for intermediate sequencing.

Materials:

• Liquid Handling Robot: e.g., Integrated with the iBioFAB or similar platform [1].
• Thermal Cycler: On-deck or accessible via robotic arm.
• Microbial Transformation System: 96-well format for high-throughput.
• Plasmids: pSEVA-based backbone vectors, compatible with Golden Standard Modular Cloning (GS MoClo) or similar systems [21].
• Enzymes: High-fidelity DNA polymerase (e.g., for mutagenesis PCR), BsaI-HFv2, BbsI-HF, T4 DNA ligase [21].
• Oligonucleotides: PCR primers and synthetic gene fragments.

Procedure:

• Library Design & Primer Planning: The biofoundry's scheduling software receives a list of target variant sequences from the AI design module. Primers for PCR-based mutagenesis or DNA assembly are automatically designed and scheduled.
• Automated Mutagenesis PCR: In a 96-well plate, set up PCR reactions using a high-fidelity polymerase to amplify the plasmid template with the desired mutations. A typical reaction volume is 20 µL.
  • Robotic Program: The liquid handler dispenses water, buffer, dNTPs, template DNA, primers, and polymerase into the plate. The sealed plate is transferred to the thermal cycler for amplification [1].
• DpnI Digestion & Purification: Post-PCR, add DpnI enzyme to the reaction mix to digest the methylated template DNA. Incubate for 1 hour at 37°C. The robotic system then purifies the digested PCR product using a magnetic bead-based clean-up protocol [1].
• DNA Assembly: For multi-part assemblies (e.g., constructing multi-gene circuits), use a standardized assembly method like Golden Gate Assembly.
  • Reaction Setup: In a new 96-well plate, mix the purified PCR fragments (or synthesized oligos) with the appropriate backbone vector, BsaI or BbsI restriction enzyme, and T4 DNA ligase buffer and ligase.
  • Cycling Conditions: Program the thermal cycler for a Golden Gate reaction (e.g., 30 cycles of 37°C for 2 minutes and 16°C for 5 minutes, followed by a final 60°C for 10 minutes) [21].
• High-Throughput Transformation: Transform the assembly reaction into chemically competent E. coli cells (e.g., NEB 10-beta or BL21(DE3)) in a 96-well format.
  • Robotic Program: The liquid handler aliquots competent cells, adds the DNA assembly mix, and performs the heat-shock step. After outgrowth in SOC medium, the cells are plated onto selective LB-agar plates in 8-well Omnitrays [1].
• Colony Picking and Culture: A robotic colony picker selects individual colonies and inoculates them into deep-well 96-well plates containing liquid selective media. The plates are sealed with gas-permeable seals and incubated at 37°C with shaking for ~16 hours.
• Plasmid Purification: The liquid handler performs an alkaline lysis-based plasmid miniprep protocol on the overnight cultures to yield purified plasmid DNA, which is eluted in water or TE buffer [1]. A subset of clones may be sequenced for quality control, but the high fidelity of the assembly method (>95% accuracy) often allows this step to be bypassed in iterative cycles to save time [1].

Module 2: Automated Protein Expression and Purification

Objective: To express and purify target protein variants in a high-throughput, miniaturized format.

Background: Cell-free protein synthesis (CFPS) systems are particularly amenable to automation, enabling rapid expression freed from cell viability constraints [22]. For in vivo expression, auto-induction systems in microtiter plates streamline the process.

Materials:

• Cell-Free System: E. coli S30 extract, energy regeneration system (e.g., based on phosphoenolpyruvate or maltodextrin), amino acids, cofactors, and NTPs [22].
• Expression Host: E. coli BL21(DE3) or other suitable strains.
• Media: 2xM9 minimal medium for in vivo expression or Dynamite medium for high-density culture [21].
• Inducers: Isopropyl β-d-1-thiogalactopyranoside (IPTG), rhamnose, arabinose, or m-toluic acid, depending on the promoter system used [21].
• Purification Resins: Nickel-nitrilotriacetic acid (Ni-NTA) agarose or other affinity resins in 96-well filter plates.
• Liquid Handler: Capable of handling 96-well deep-well plates.

Procedure:

• A. Cell-Free Protein Synthesis (CFPS) Protocol:
  • Reaction Setup: On a chilled deck, the liquid handler prepares the CFPS master mix containing S30 extract, energy source, amino acids, salts, and cofactors in a 96-well plate.
  • Template Addition: The purified plasmid DNA or linear PCR template from Module 1 is added to the reaction mix.
  • Incubation: The plate is sealed and transferred to an incubator shaker at 30°C for 4-6 hours for protein expression [22]. The resulting lysate can often be used directly in functional assays.

• B. In Vivo Protein Expression Protocol:
  • Inoculation: The liquid handler uses the purified plasmid DNA from Module 1 to transform or transform and inoculate expression hosts in a deep-well 96-well plate containing selective media.
  • Growth and Induction: The plate is incubated at 37°C with shaking until the culture reaches mid-log phase (OD600 ~0.6-0.8). Induction is then performed by the automated addition of a specific inducer (e.g., 0.1-1.0 mM IPTG for the LacI/Trc system). For high-throughput, auto-induction media can be used to eliminate the need for manual induction timing [14].
  • Protein Production: Post-induction, the temperature is typically reduced to 25°C, and shaking continues for 16-20 hours for optimal protein expression [21].
  • Cell Harvest and Lysis: Cultures are centrifuged using a plate rotor, and the supernatant is discarded. The cell pellets are resuspended in lysis buffer and lysed by chemical or enzymatic methods (e.g., lysozyme) on the deck.
  • Automated Purification (SAPP Workflow):
    • Clarification: The lysate is centrifuged or filtered to remove cellular debris.
    • Affinity Chromatography: The clarified lysate is transferred to a 96-well filter plate pre-loaded with Ni-NTA resin. The liquid handler performs binding, washing, and elution steps using imidazole gradients.
    • Size-Exclusion Chromatography (SEC): The eluate is injected into a micro-scale SEC system. This step simultaneously purifies the protein and provides data on purity, oligomeric state, and dispersity [14]. An open-source software tool automates the analysis of thousands of SEC chromatograms.

Module 3: Automated High-Throughput Assaying

Objective: To quantitatively measure the fitness (e.g., enzymatic activity, binding affinity) of protein variants.

Background: Assays must be compatible with microtiter plate formats and yield quantitative, machine-readable data to feed the "Learn" phase of the DBTL cycle.

Materials:

• Microplate Reader: Capable of measuring absorbance, fluorescence, and/or luminescence.
• Assay Reagents: Substrates, cofactors, buffers specific to the enzyme being tested.
• Liquid Handler: For precise reagent dispensing.

Procedure:

• Assay Plate Preparation: The liquid handler transfers a small, standardized aliquot of the purified protein (from Module 2) or the CFPS reaction lysate into a 96-well or 384-well assay plate. For cell-based assays, crude cell lysates can be used directly after a lysis step [1].
• Reagent Dispensing: The assay reaction is initiated by the automated addition of the appropriate substrate and buffer.
  • Example for Phytase Activity: Initiate reaction by adding phytate substrate in a buffer at the desired pH (e.g., neutral pH for engineering pH robustness) [1].
  • Example for Methyltransferase Activity: Add substrate (e.g., ethyl iodide) and S-adenosyl-l-homocysteine (SAH) to measure alkyltransferase activity [1].
• Kinetic Measurement: The plate is immediately transferred to a microplate reader. The instrument continuously monitors the reaction (e.g., every 30-60 seconds for 10-30 minutes) by measuring the change in absorbance or fluorescence resulting from product formation.
• Data Processing: The raw kinetic data is automatically processed by integrated software. The initial reaction rates (V0) are calculated and normalized against protein concentration (determined by a parallel colorimetric assay like Bradford). The resulting fitness score (e.g., specific activity, improvement fold over wild-type) is compiled into a database for machine learning analysis [1] [23].

Performance Metrics and Benchmarking

The efficacy of an automated biofoundry is quantified by its throughput, speed, and success in engineering proteins. The table below summarizes performance data from recent campaigns.

Table 1: Benchmarking Performance of Automated Protein Engineering Platforms

Platform / Study	Target Enzyme	Engineering Goal	Duration & Scale	Key Outcome
AI-Powered Platform [1]	Halide Methyltransferase (AtHMT)	Improve substrate preference & ethyltransferase activity	4 rounds, <500 variants	90-fold improvement in substrate preference; 16-fold improvement in target activity
AI-Powered Platform [1]	Phytase (YmPhytase)	Improve activity at neutral pH	4 rounds, <500 variants	26-fold improvement in activity at neutral pH
PLMeAE [17]	tRNA synthetase (pCNF-RS)	Improve enzyme activity	4 rounds, 10 days	Up to 2.4-fold improvement in enzyme activity
CAPE Challenge [23]	RhlA	Enhance catalytic activity for rhamnolipid production	2 rounds, ~1500 variants	Best mutant production 6.16x higher than wild-type

Essential Research Reagent Solutions

The following table catalogs key reagents and their functions that are fundamental to operating the automated workflows described herein.

Table 2: Key Research Reagent Solutions for Automated Biofoundries

Reagent / Material	Function / Application	Example Use Case
SEVA/GS MoClo Vectors [21]	Standardized, modular plasmid system for reliable DNA assembly	Facilitating rapid and combinatorial assembly of genetic circuits and expression cassettes
Cell-Free Protein Synthesis (CFPS) System [22]	In vitro transcription/translation system for rapid protein production	Bypassing cell culture for high-speed expression and screening of protein variants, including toxic proteins
Inducible Promoter Systems (e.g., RhaBAD, Trc, AraBAD) [21]	Provides precise, independent control over the expression of multiple genes	Enabling optimized co-expression of multi-enzyme cascades from a single plasmid
Semi-Automated Protein Production (SAPP) Workflow [14]	Integrated protocol for high-throughput protein purification and characterization	Delivering purified, characterized protein in 48 hours with minimal hands-on time
DMX DNA Synthesis Method [14]	Cost-effective method for constructing sequence-verified clones from oligo pools	Reducing the cost of DNA synthesis, the major bottleneck in large-scale campaigns, by 5-8 fold

Workflow Visualization

The following diagram illustrates the integrated, closed-loop workflow of an autonomous biofoundry for protein engineering.

Autonomous Biofoundry DBTL Cycle

This continuous cycle of Design-Build-Test-Learn, powered by AI and automated by robotics, enables the rapid and autonomous engineering of proteins with desired properties, dramatically accelerating the pace of biological innovation.

The advent of autonomous protein engineering platforms is revolutionizing the field of protein science, enabling the systematic and high-throughput exploration of protein sequence space. These integrated systems combine artificial intelligence (AI)-driven design with automated experimental workflows to execute rapid "Design-Build-Test-Learn" (DBTL) cycles. This article details specific application notes and protocols for engineering key protein properties—thermostability, catalytic activity, and substrate preference—within the context of such platforms. By providing quantitative results and standardized methodologies, we aim to equip researchers with the tools to leverage autonomous systems for accelerating the development of novel biocatalysts, therapeutics, and research reagents.

Application Notes

Enhancing Thermostability and Activity via Multimodal Inverse Folding

Application: The ABACUS-T model was employed to redesign several enzymes to significantly increase their thermostability without compromising, and in some cases enhancing, their native catalytic activity. This approach is particularly valuable for developing industrial enzymes that must operate under harsh conditions.

Key Results: The following table summarizes the quantitative outcomes of protein redesigns using the ABACUS-T platform.

Table 1: Summary of Protein Engineering Outcomes Using ABACUS-T [24]

Protein Engineered	Key Functional Enhancement	Change in Thermostability (ΔTm)	Number of Mutations Tested
Allose Binding Protein	17-fold higher binding affinity	≥ 10 °C	Dozens of simultaneous mutations
Endo-1,4-β-xylanase	Maintained or surpassed wild-type activity	≥ 10 °C	Dozens of simultaneous mutations
TEM β-lactamase	Maintained or surpassed wild-type activity	≥ 10 °C	Dozens of simultaneous mutations
OXA β-lactamase	Altered substrate selectivity	≥ 10 °C	Dozens of simultaneous mutations

Technical Insight: Traditional inverse folding models often produce highly stable but functionally inactive proteins because they prioritize structural stability over functional constraints. ABACUS-T addresses this by unifying several critical features in a single framework [24]:

• Atomic-level detail: Incorporates sidechain and ligand interactions.
• Evolutionary information: Integrates data from Multiple Sequence Alignments (MSA) to maintain functional residues.
• Conformational flexibility: Considers multiple backbone states to preserve functional dynamics.

This multimodal approach allows for the testing of highly mutated sequences (containing dozens of simultaneous mutations) with a high success rate, bypassing the need for extensive experimental screening.

Rapid Engineering of Protein Assemblies for Potency

Application: A semi-automated platform was used to engineer a potent multimeric viral neutralizer, demonstrating the power of high-throughput screening for optimizing complex protein assemblies.

Key Results:

Table 2: Outcomes of High-Throughput Multimer Engineering for Viral Neutralization [14]

Construct Type	Starting Monomer ICâ‚…â‚€	Best Engineered Dimer ICâ‚…â‚€	Best Engineered Trimer ICâ‚…â‚€	Commercial Antibody (MPE8) ICâ‚…â‚€
RSV Neutralizer	5.4 nM	40 pM	59 pM	156 pM

Technical Insight: The platform identified 19 correctly assembled multimer scaffolds from a library of 58 designs. The results highlight that geometric arrangement is critical to function; the optimal dimer and trimer constructs showed a dramatic >135-fold improvement in potency over the monomer and surpassed a leading commercial antibody. This success was made feasible by the platform's ability to rapidly screen a vast combinatorial space of protein assemblies [14].

Stabilization for DNA-Binding Specificity Retargeting

Application: Engineering the DNA-binding specificity of meganucleases requires accumulating numerous mutations on the protein scaffold, which often leads to destabilization and loss of function.

Key Results: A strategy of computationally "pre-stabilizing" the protein scaffold was used to counteract the destabilizing effects of mutations introduced to change DNA-binding specificity. This approach improved the recovery of active, fully retargeted enzymes with robust activity in vitro and in human genomic targets [25].

Technical Insight: Laboratory-directed evolution often fails to maintain the balance between stability and function that natural evolution enforces. By proactively engineering a more thermostable scaffold before initiating selections for new function, researchers can create a "fitter" starting point that is more tolerant to the destabilizing mutations that inevitably occur during the engineering process [25].

Experimental Protocols

Protocol: Semi-Automated Protein Production (SAPP) for High-Throughput Validation

This protocol enables a 48-hour turnaround from DNA to purified protein with minimal hands-on time, designed to integrate seamlessly with autonomous platforms [14].

1. Cloning (Sequencing-Free)

• Method: Use Golden Gate Assembly with a destination vector containing a ccdB "suicide gene."
• Rationale: This negative selection strategy yields ~90% cloning accuracy, eliminating the need for time-consuming colony picking and Sanger sequencing verification.
• Procedure: Perform the assembly reaction and transform into a suitable E. coli strain. The ccdB gene is lethal to non-recombinant cells, ensuring high background rejection.

2. Small-Scale Expression & Lysis

• Vessel: Perform all steps in a 96-well deep-well plate.
• Culture: Inoculate cultures in auto-induction media. This removes the requirement for manual monitoring and induction with IPTG.
• Growth: Incubate at appropriate temperature with shaking for 24 hours.
• Lysis: Following growth, lyse cells using a chemical lysis reagent (e.g., BugBuster Master Mix) on an orbital shaker.

3. Parallel Purification & Analysis

• Step 1: Affinity Purification. Using a programmable liquid handler, pass the lysates through a 96-well plate containing nickel-affinity resin for His-tagged proteins.
• Step 2: Size-Exclusion Chromatography (SEC). Subject the eluates to miniaturized SEC.
• Data Output: The SEC chromatogram provides simultaneous, quantitative data on:
  • Purity: Peak shape and homogeneity.
  • Oligomeric State: Apparent molecular weight.
  • Dispersity: Polydispersity index of the peak.
  • Yield: Integrated peak area.

4. Automated Data Analysis

• Tool: Use provided open-source software to automatically analyze thousands of SEC chromatograms.
• Output: Standardized data (e.g., purity score, oligomeric state, yield) is generated for direct feedback into the AI model.

Protocol: DMX for Low-Cost DNA Library Construction

The DMX protocol addresses the DNA synthesis bottleneck, reducing the per-design DNA construction cost by 5- to 8-fold [14].

1. Oligo Pool Synthesis

• Procedure: Order the library of gene variants as a complex pool of oligonucleotides.

2. Gene Assembly & Barcoding

• Assembly: Perform a one-pot isothermal assembly reaction to build full-length genes from the oligo pool.
• Barcoding: Within the cell lysate, use a novel isothermal barcoding reaction to tag each assembled gene variant with a unique DNA barcode.

3. Sequencing & Deconvolution

• Method: Use long-read nanopore sequencing on the barcoded lysate.
• Output: The sequencing data links each unique barcode to its corresponding full-length gene sequence, creating a map of sequence-verified clones.
• Recovery: This method typically recovers >75% of designed variants from a single pool.

Visualization of Workflows

ABACUS-T Multimodal Inverse Folding Architecture

SAPP High-Throughput Experimental Pipeline

The Scientist's Toolkit

Table 3: Essential Research Reagent Solutions for Autonomous Protein Engineering [8] [14]

Reagent / Material	Function in Workflow
ccDB Suicide Gene Vectors	Enables high-efficiency, sequencing-free cloning by selecting against non-recombinant background.
Auto-induction Media	Allows for high-throughput, parallel protein expression without manual induction steps.
96-Well Plate Nickel-Affinity Resin	Facilitates parallel purification of His-tagged proteins in a plate-based format compatible with liquid handlers.
Miniaturized SEC Columns	Provides simultaneous data on protein purity, oligomeric state, and yield from micro-volume samples.
Oligo Pools	Serves as the source material for building large libraries of gene variants cost-effectively.
Golden Gate Assembly Mix	A modular and highly efficient DNA assembly method used in both SAPP and DMX workflows.
Cyclobenzaprine N-oxide	Cyclobenzaprine N-oxide, CAS:6682-26-4, MF:C20H21NO, MW:291.4 g/mol
Nordoxepin hydrochloride	Nordoxepin hydrochloride, CAS:2887-91-4, MF:C18H20ClNO, MW:301.8 g/mol

Navigating Challenges and Enhancing Platform Performance

Overcoming Experimental Noise and Inconclusive Data with Automated Quality Control

The emergence of autonomous platforms like the Self-driving Autonomous Machines for Protein Landscape Exploration (SAMPLE) is revolutionizing protein engineering by performing fully autonomous design-build-test-learn cycles [9]. However, the performance of these systems critically depends on the quality and reliability of the experimental data they generate. Experimental noise and inconclusive results from poor-quality protein reagents can severely compromise the AI's learning process, leading to inefficient optimization or convergence on false optima [26] [9]. This application note details the implementation of robust, automated quality control (QC) protocols within autonomous protein engineering workflows to mitigate these challenges, with specific methodologies and datasets presented for direct implementation.

The Critical Role of QC in Autonomous Engineering

In traditional research, scientists use intuition to identify and discard anomalous data points. Autonomous systems lack this inherent judgment, making structured QC protocols essential. Without such safeguards, costly cycles can be wasted on characterizing poor-quality reagents, and the resulting noisy data can lead intelligent agents to build flawed models of the protein fitness landscape. One analysis attributed a staggering $10.4 billion worth of irreproducible research annually in the US alone to poor quality biological reagents and reference materials [26].

The SAMPLE platform exemplifies how to embed QC within an autonomous workflow. Its operation includes multiple automated checkpoints that validate successful gene assembly, expected enzyme reaction progress curves, and activity above background levels before data is passed to the learning agent [9]. This systematic approach to rejecting inconclusive data is a foundational element of its success, enabling the platform to engineer glycoside hydrolase enzymes with enhanced thermal tolerance efficiently [9].

Essential Quality Control Metrics for Protein Reagents

Implementing automated QC requires defining the critical quality attributes (CQAs) for protein reagents. The table below summarizes the minimal recommended QC tests, based on community-established guidelines, that should be integrated into an automated workflow [26] [27].

Table 1: Essential Quality Control Metrics for Protein Reagents

QC Category	Key Metric	Recommended Method(s)	Purpose in Autonomous Workflows
Identity	Protein Sequence	Mass Spectrometry (Intact Mass)	Confirms the correct protein is expressed and not a host contaminant [26].
Purity	Sample Homogeneity	SDS-PAGE, Capillary Electrophoresis	Detects contaminating proteins or proteolysis that could skew functional data [26].
Structural Integrity	Oligomeric State & Aggregation	Size Exclusion Chromatography (SEC), SEC-MALS, DLS	Identifies aggregates or incorrect oligomeric states that cause overestimation of active concentration [26] [27].
Concentration	Accurate Quantification	Spectrophotometry (A280), BCA/Bradford Assay	Ensures accurate dosing in functional assays; A260/A280 ratio checks for nucleic acid contamination [27].
Functional Stability	Folding & Thermal Stability	nanoDSF, Circular Dichroism (CD)	Assesses proper folding and provides a stability index (e.g., melting temperature) for batch consistency [27].

Extended QC for Specific Applications

For certain downstream applications, extended QC is essential. Proteins produced in E. coli for cell-based assays must be tested for endotoxins, typically using chromogenic LAL or recombinant Factor C (rFC) methods to achieve levels below 1 EU/mL [26] [27].

Protocol: Integrated QC for an Autonomous Protein Engineering Campaign

This protocol outlines a streamlined QC pipeline suitable for integration into platforms like the SAMPLE robotic system or the Illinois Biological Foundry (iBioFAB) [9] [1].

Automated Protein Expression and Purification

• Module 1: Gene Assembly and Verification. Utilize an automated, high-fidelity DNA assembly method (e.g., Golden Gate or HiFi-assembly-based mutagenesis). Verify successful assembly directly via PCR with a double-stranded DNA-binding dye like EvaGreen, eliminating the need for intermediate sequencing and maintaining continuity [9] [1].
• Module 2: Cell-Free Protein Expression. Employ a centralized robotic arm to transfer the verified expression cassette into a cell-free protein expression system. This step bypasses cell viability issues, accelerates protein production (∼3 hours), and is highly amenable to automation in 96- or 384-well formats [9].

Automated Quality Control Analysis

The following QC tests can be automated and performed in parallel on the expressed protein samples.

• QC 1: Concentration and Purity Check.

  • Method: Automated, non-destructive spectrophotometry.
  • Procedure: Transfer a 2 µL aliquot to a nanodrop spectrophotometer. Measure absorbance at 280 nm for concentration calculation and the A260/A280 ratio for nucleic acid contamination.
  • Automated Decision Point: Flag samples with concentration outside the expected range or with A260/A280 > 0.8 for further review [27].

• QC 2: Integrity and Oligomeric State Analysis.

  • Method: High-throughput Size Exclusion Chromatography (HT-SEC).
  • Procedure: Inject 10 µL of the protein sample onto an integrated UPLC system equipped with a size exclusion column. Monitor elution at 280 nm.
  • Automated Decision Point: Accept samples where the main peak constitutes >90% of the total chromatogram and its retention time corresponds to the expected oligomeric state. Samples showing significant aggregation or multiple peaks are failed [26] [27].

• QC 3: Thermal Stability Assessment.

  • Method: Nano Differential Scanning Fluorimetry (nanoDSF).
  • Procedure: Load a 10 µL protein sample into a nanoDSF capillary. Run a thermal ramp from 20°C to 95°C while monitoring intrinsic tryptophan and tyrosine fluorescence.
  • Automated Decision Point: Record the melting temperature (Tm). Flag samples with a Tm significantly lower than the stable wild-type control or historical benchmarks (e.g., ΔTm > 5°C) as potentially unstable [27].

Functional Assay and Data Integration

• Module 3: Functional Screening. Only samples passing the above QC checkpoints proceed to the functional assay (e.g., thermostability T50 measurement for enzymes). The SAMPLE platform, for instance, uses a colorimetric/fluorescent assay to measure activity over a temperature gradient [9].
• Module 4: Data Validation and Agent Feedback. The platform's software must perform a final validation of the functional assay data. This includes checking that enzyme reaction progress curves fit the expected sigmoidal model and that activity is above background. Only data from QC-passed proteins with valid functional readouts is then sent to the AI agent to update its model of the sequence-function landscape [9].

The following workflow diagram illustrates the integration of these automated QC checkpoints within a closed-loop autonomous system.

Figure 1: Autonomous engineering workflow with integrated QC checkpoints. This diagram illustrates the closed-loop design-build-test-learn cycle, highlighting the critical automated quality control (QC) pipeline. Only samples passing all QC checkpoints proceed to functional testing, ensuring the AI agent learns only from high-quality data.

Data Output and Analysis

Implementing this automated QC protocol in an autonomous platform provides quantitative data for both immediate decision-making and long-term analysis.

Table 2: Representative QC and Functional Data from an Autonomous Run This table simulates data output for four protein variants (V1-V4) during a single engineering cycle, demonstrating how QC metrics correlate with functional performance.

Variant	QC1: Conc. (mg/mL)	QC2: % Monomer (SEC)	QC3: Tm (°C)	QC Status	Functional Assay: T50 (°C)
V1	0.85	95	62.1	PASS	59.5 ± 0.8
V2	0.12	90	45.2	FAIL (Low Conc., Low Tm)	N/A
V3	0.78	60	58.5	FAIL (High Aggregation)	N/A
V4	0.91	97	65.3	PASS	63.1 ± 0.5

Analysis: The data shows that V2 and V3 were correctly flagged by the automated QC. Using V3's functional data would have been misleading, as the low T50 could be due to aggregation rather than an inherently unstable fold. The AI agent only learns from the high-quality data for V1 and V4, ensuring an accurate update of its model.

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Materials for Automated Protein QC Key reagents, instruments, and software required to establish the automated QC protocols described in this document.

Item	Function/Role	Example/Notes
Cell-Free Protein Expression System	Enables rapid, automated protein synthesis without cell culture.	T7-based expression reagents, suitable for 96-well formats [9].
Automated Liquid Handler	Precisely transfers reagents and samples across all modules.	Integrated with a central robotic arm (e.g., in the iBioFAB) [1].
Nano Spectrophotometer	Measures protein concentration and checks for nucleic acid contamination.	Provides non-destructive analysis with minimal sample consumption [27].
High-Throughput UPLC-SEC	Automatically analyzes sample homogeneity and oligomeric state.	Systems configured for 96-well plate sampling for high throughput [27].
nanoDSF Instrument	Automatically assesses protein folding and thermal stability.	Measures intrinsic fluorescence without dyes; uses minimal sample [27].
Protein LLMs & AI Agents	Designs variants and models the fitness landscape from validated data.	ESM-2 for variant design; Bayesian Optimization (e.g., Expected UCB) for search [9] [1].
HPCR	HPCR, MF:C52H40O12, MW:856.9g/mol	Chemical Reagent
GABAA receptor agent 1	2-(4-Chlorophenyl)-5-nitro-1H-benzimidazole	High-purity 2-(4-Chlorophenyl)-5-nitro-1H-benzimidazole for research. Explore its potential as a PARP inhibitor and anticancer agent. This product is for Research Use Only (RUO). Not for human or veterinary use.

Integrating robust, automated quality control is not an optional enhancement but a foundational requirement for reliable autonomous protein engineering. By implementing the detailed protocols and metrics outlined in this application note, research teams can equip platforms like SAMPLE and iBioFAB to self-correct for experimental noise, avoid dead-end experiments, and dramatically accelerate the discovery of robust, high-performance proteins. This systematic approach ensures that the AI-driven exploration of the protein fitness landscape is built upon a solid foundation of high-quality, reproducible data.

Strategies for Efficiently Sampling Vast Sequence Spaces

The exploration of vast protein sequence spaces represents a fundamental challenge in modern bioengineering and drug development. The number of possible sequences for even a short protein is astronomically large, making exhaustive experimental screening impossible [28]. Autonomous protein engineering platforms, such as the SAMPLE platform, are designed to overcome this challenge by intelligently navigating these expansive landscapes to discover variants with enhanced functions [6] [29]. Efficient sampling is therefore not merely an advantage but a necessity for success in projects ranging from therapeutic enzyme development [30] to the engineering of novel biocatalysts [31]. This Application Note details the core strategies and protocols that enable effective exploration within these spaces, framed within the context of autonomous research systems.

Core Sampling Strategies

Machine Learning-Guided Exploration

Machine learning (ML) has emerged as a powerful tool for predicting protein fitness, thereby reducing the number of variants that need to be experimentally characterized.

• Protein Language Models (PLMs): Models such as ESM-2 are trained on evolutionary-scale protein sequence databases and can perform "zero-shot" prediction of variant fitness, meaning they can propose promising mutants without any prior experimental data on the target protein. One study used this approach to design an initial library of 96 variants, 55-60% of which performed better than the wild-type enzyme [6] [29].
• Supervised Machine Learning: After an initial round of experimentation, the collected data can be used to train supervised ML models (e.g., Multi-Layer Perceptrons, Bayesian Optimization) to build a sequence-function model. This model then predicts which variants should be tested in the next cycle, creating an active learning loop [6] [29]. This strategy has been successfully used to improve enzyme activity by 2.4-fold within four rounds of evolution [29].

Library Design and Diversity Optimization

The design of the initial variant library is critical for exploring sequence space effectively.

• Combining Unsupervised Models: A high-quality starting library can be generated by combining predictions from a protein language model (ESM-2) with an epistasis model (EVmutation). This synergy maximizes both the diversity of the library and the likelihood of containing improved mutants, with one report showing over 50% of initial variants outperforming the wild-type for a given enzyme [6].
• Multi-Library Genetic Algorithms: For complex search spaces, such as those for therapeutic peptides, a multi-library approach using a genetic algorithm can be highly effective. This method partitions the vast sequence space into smaller, distinct sub-libraries that maximize intra-library and cross-library diversity. This ensures broad coverage while streamlining downstream synthesis and screening efforts [28].

Table 1: Key Strategies for Sampling Sequence Space

Strategy	Methodology	Key Advantage	Reported Outcome
Protein Language Models (PLMs)	Zero-shot fitness prediction using models like ESM-2 trained on evolutionary data.	Requires no prior experimental data; leverages evolutionary constraints.	>50% of initial 96-variant library showed improved fitness [6] [29].
Active Learning with Supervised ML	Iterative DBTL cycles using experimental data to train a fitness predictor (e.g., MLP).	Efficiently focuses resources on high-fitness regions of sequence space.	2.4-fold activity improvement in 4 rounds (10 days) [29].
Diversity-Optimized Library Design	Using genetic algorithms to design multiple, distinct sub-libraries.	Maximizes coverage of a vast search space while simplifying deconvolution.	Enables efficient exploration of NP-hard combinatorial peptide spaces [28].

Experimental Protocols

Protocol: Autonomous DBTL Cycle for Protein Engineering

This protocol outlines an iterative cycle for autonomous enzyme engineering, as implemented on the Illinois Biological Foundry (iBioFAB) and similar platforms [6] [29].

1. Design Phase

• Input: Wild-type protein sequence and a quantifiable fitness assay.
• Procedure:
  • For proteins without prior sites, use a PLM (e.g., ESM-2) to mask each residue and calculate the likelihood of all possible single-point mutations. Select the top 96 variants with the highest predicted fitness [29].
  • For proteins with known target sites, use a PLM to predict high-fitness multi-mutant combinations across the specified residues [29].
  • Alternatively, for subsequent cycles, use a trained supervised ML model to propose the next set of 96 variants based on accumulated experimental data.

2. Build Phase

• Method: Employ a high-fidelity (HiFi) assembly-based mutagenesis method on an automated biofoundry.
• Automated Modules:
  • Mutagenesis PCR and DpnI digestion.
  • DNA assembly and 96-well microbial transformations.
  • Colony picking and plasmid purification in a 96-well format.
  • Protein expression in a 96-deep well plate.
• Quality Control: Random sequencing confirms >95% correct assembly, eliminating the need for intermediate verification and enabling a continuous workflow [6].

3. Test Phase

• Assay: Execute an automated, high-throughput functional assay in a 96-well plate (e.g., ethyltransferase activity assay for AtHMT or phytase activity at neutral pH for YmPhytase) [6].
• Data Collection: Automatically record fitness data (e.g., enzyme activity) for each variant and associate it with the corresponding sequence.

4. Learn Phase

• Data Integration: Encode all tested variant sequences using the PLM, creating a numerical feature set.
• Model Training: Train a supervised ML model (e.g., a Multi-Layer Perceptron) on the experimental data to map sequence encodings to fitness values.
• New Proposal: Use the trained model to predict the fitness of a new virtual library of variants and select the top 96 candidates for the next DBTL cycle [29].

Workflow Visualization

The following diagram illustrates the closed-loop, autonomous workflow of the Design-Build-Test-Learn cycle.

The Scientist's Toolkit

Table 2: Essential Research Reagent Solutions for Autonomous Sequence Space Exploration

Tool / Reagent	Function / Description	Application in Workflow
Protein Language Model (ESM-2)	A transformer-based model that predicts amino acid likelihoods, enabling zero-shot fitness inference.	Design phase: Proposes high-likelihood single and multi-mutant variants for initial library generation [6] [29].
Automated Biofoundry (e.g., iBioFAB)	Integrated robotic system with liquid handlers, thermocyclers, and incubators for fully automated molecular biology.	Build & Test phases: Executes mutagenesis, transformation, protein expression, and assay in a continuous, high-throughput manner [6] [29].
HiFi Assembly Mutagenesis	A high-fidelity DNA assembly method that eliminates the need for intermediate sequence verification.	Build phase: Ensures robust and continuous library construction with >95% accuracy [6].
Supervised ML Model (e.g., MLP)	A machine learning model trained on experimental data to predict variant fitness from sequence encodings.	Learn & Design phases: Learns from experimental data to propose improved variants in subsequent DBTL cycles [29].

The integration of machine learning, strategic library design, and full laboratory automation creates a powerful framework for efficiently sampling vast protein sequence spaces. The outlined strategies and protocols demonstrate that through iterative, data-driven cycles, it is possible to achieve significant functional improvements—such as multi-fold increases in enzyme activity—within a matter of weeks, while characterizing only a minute fraction of the total possible sequence space. These approaches, central to autonomous platforms like SAMPLE, are paving the way for accelerated advancements in therapeutic development, biocatalysis, and synthetic biology.

Balancing Exploration and Exploitation in Bayesian Optimization

In Bayesian optimization (BO), the balance between exploring uncertain regions of the parameter space and exploiting areas known to yield high performance is fundamental to efficient optimization. This trade-off is managed through acquisition functions, which use the predictive mean and uncertainty from a probabilistic surrogate model, typically a Gaussian process (GP), to guide the selection of subsequent experiments [32]. The strategic balance is particularly crucial in autonomous protein engineering, where experimental resources are limited and each cycle of design, build, test, and learn (DBTL) is time-consuming and costly. By effectively navigating vast sequence spaces, BO enables researchers to identify promising protein variants with desired properties more rapidly than traditional methods.

Core Methodological Framework

Gaussian Process Surrogate Models

Gaussian processes serve as the foundation for Bayesian optimization by providing a probabilistic model of the unknown objective function. A GP is defined by a mean function, m(x), and a covariance kernel function, k(x,x'), which captures the similarity between data points [32]. The kernel function choice is critical, with common selections including the squared exponential (Radial Basis Function) and Matérn kernels [32]. The GP model generates both a prediction (mean) and an uncertainty estimate (standard deviation) for any point in the design space, forming the basis for the exploration-exploitation trade-off managed by acquisition functions.

Acquisition Functions for Balancing Trade-Offs

Acquisition functions mathematically formalize the balance between exploration and exploitation by leveraging the GP's predictive mean and uncertainty [32]. These functions determine the next experimental points to evaluate based on different strategies for balancing these competing objectives. The following table summarizes the primary acquisition function categories and their characteristics:

Table 1: Acquisition Functions for Balancing Exploration and Exploitation

Category	Representative Functions	Mechanism	Best Use Cases
Improvement-Based	Probability of Improvement (PI), Expected Improvement (EI)	Focus on probability or expectation of improving over current best value	When quickly converging to a known promising region is prioritized
Optimistic	Upper Confidence Bound (UCB)	Uses confidence interval upper bound: µ(x) + κσ(x)	When systematic exploration of uncertain regions is needed
Information-Based	Entropy Search, Predictive Entropy Search	Seek to reduce uncertainty about optimum location	When global optimization is critical and computational resources allow
Safe Exploration	Mean Deviation (MD) [33]	Penalizes uncertain regions: ρµ(x) - σ(x)	When evaluating poor or non-expressing proteins is costly or risky

The Upper Confidence Bound (UCB) function exemplifies the explicit balance between exploration and exploitation, mathematically represented as α(x) = µ(x) + κσ(x), where the tuning parameter κ controls the balance between mean prediction (µ) and uncertainty (σ) [32]. Similarly, the recently proposed Mean Deviation approach directly incorporates uncertainty as a penalty term, promoting safer exploration in regions where the model predictions are reliable [33].

Application in Autonomous Protein Engineering Platforms

Case Study: AI-Powered Enzyme Engineering Platform

Recent research demonstrates the successful implementation of Bayesian optimization within an autonomous platform for enzyme engineering. This platform integrated machine learning with biofoundry automation to engineer Arabidopsis thaliana halide methyltransferase (AtHMT) and Yersinia mollaretii phytase (YmPhytase) [1]. The autonomous system achieved remarkable results within four weeks: a 90-fold improvement in substrate preference and 16-fold improvement in ethyltransferase activity for AtHMT, along with a 26-fold improvement in neutral pH activity for YmPhytase [1]. This was accomplished while requiring construction and characterization of fewer than 500 variants for each enzyme, demonstrating the exceptional efficiency of Bayesian optimization in guiding experimental workflows.

Implementation Workflow for Protein Engineering

The protein engineering workflow follows an iterative DBTL cycle, with Bayesian optimization at its core:

Table 2: Bayesian Optimization in the Protein Engineering DBTL Cycle

Stage	Key Activities	BO Integration
Design	Generate initial variant library using protein LLM (ESM-2) and epistasis models	Initial space-filling design or prior knowledge excitation
Build	Automated mutagenesis, transformation, and plasmid preparation	Not applicable
Test	High-throughput characterization of protein function	Experimental observations with noise characterization
Learn	Train surrogate model (GP) on collected data	Update GP with new data, optimize acquisition function to select next variants

The following diagram illustrates the complete Bayesian optimization workflow within an autonomous protein engineering platform:

Experimental Protocols

Protocol: Bayesian Optimization for Enzyme Engineering

Objective: Optimize enzymatic properties (e.g., activity, specificity, stability) using autonomous Bayesian optimization.

Materials & Reagents:

• Target enzyme plasmid DNA
• Primers for site-directed mutagenesis or HiFi assembly
• Expression host cells (e.g., E. coli strains)
• Cell culture media and induction reagents
• Enzyme substrates and assay buffers
• Microtiter plates and robotic-compatible labware

Equipment:

• Automated liquid handling system
• PCR thermocycler
• Microplate spectrophotometer/fluorometer
• Colony picking robot
• Incubated shakers
• Integrated biofoundry (e.g., iBioFAB)

Procedure:

• Initial Library Design:

  • Generate 150-200 initial variants using unsupervised models (ESM-2 protein LLM and EVmutation epistasis model) [1]
  • Select variants using space-filling design (Sobol sequences) to cover sequence space

• Automated Library Construction:

  • Perform HiFi-assembly based mutagenesis in 96-well format
  • Conduct DpnI digestion to remove template DNA
  • Transform expression host using high-throughput microbial transformation
  • Plate on 8-well omnitray LB plates with selection antibiotic
  • Pick colonies using automated colony picker

• High-Throughput Characterization:

  • Express proteins in deep-well blocks with automated induction
  • Prepare crude cell lysates using lysis buffer addition and centrifugation
  • Perform enzyme activity assays in microtiter plates
  • Measure absorbance/fluorescence using plate readers
  • Normalize data to cell density or protein concentration

• Bayesian Optimization Cycle:

  • Train Gaussian process model on collected activity data
  • Optimize acquisition function (e.g., UCB, EI) to select next variant set
  • Design primers for next-generation variants
  • Iterate through steps 2-4 for 3-5 rounds or until performance targets met

• Validation:

  • Sequence final variants to confirm mutations
  • Purify and characterize best-performing enzymes using traditional biochemical assays
  • Compare to wild-type and intermediate variants

Protocol: Safe Bayesian Optimization for Antibody Engineering

Objective: Improve antibody binding affinity while maintaining expression and stability using safe exploration approach.

Special Considerations: This protocol uses the Mean Deviation-Tree-structured Parzen Estimator (MD-TPE) to avoid non-expressing variants by penalizing uncertain regions of sequence space [33].

Procedure:

• Training Data Collection:

  • Generate single and double mutants from parent antibody sequence
  • Measure binding affinity (e.g., SPR, ELISA) and expression yield
  • Create static dataset of sequence-function relationships

• Proxy Model Training:

  • Embed protein sequences using protein language model (ESM-2)
  • Train Gaussian process model on embedded sequences and binding measurements
  • Extract predictive mean μ(x) and uncertainty σ(x) functions

• MD-TPE Optimization:

  • Set risk tolerance parameter ρ based on experimental constraints (ρ < 1 for conservative search)
  • Calculate objective function: MD = ρμ(x) - σ(x)
  • Sample sequences with high MD scores using tree-structured Parzen estimator
  • Prioritize variants with high predicted affinity and low uncertainty

• Experimental Validation:

  • Express and characterize selected variants
  • Focus on expressed proteins with maintained stability
  • Iterate with updated model

The Scientist's Toolkit

Table 3: Essential Research Reagent Solutions for Bayesian Optimization in Protein Engineering

Reagent/Resource	Function	Implementation Example
Protein Language Models (ESM-2)	Predicts amino acid likelihoods and variant fitness based on sequence context	Initial library design and sequence embedding [1]
Gaussian Process Models	Surrogate function for predicting protein properties and quantifying uncertainty	Modeling sequence-activity relationships with uncertainty estimates [32] [33]
Epistasis Models (EVmutation)	Captures amino acid covariation patterns from multiple sequence alignments	Identifying functionally important residues for library design [1]
HiFi Assembly Mix	High-fidelity DNA assembly for mutagenesis with minimal errors	Automated variant construction without intermediate sequencing [1]
Automated Screening Assays	High-throughput measurement of enzyme activity or binding	Robotic implementation of biochemical assays in microtiter plates [1]
Acquisition Functions (UCB, EI, MD)	Mathematical criteria for balancing exploration and exploitation	Selecting next variant batches based on GP predictions [32] [33]

Technical Considerations and Optimization Strategies

Handling Experimental Noise and Bias

Bioprocess optimization presents unique challenges including positional biases in microtiter plates and batch-to-batch variability. To address these:

• Implement randomization of samples across plates and batches
• Use reference standards to characterize and account for measurement noise
• Consider heteroscedastic noise models when variance changes with design variables
• Include batch effects as covariates in the Gaussian process model [34]

Computational Implementation

The computational cost of Gaussian processes scales cubically (O(n³)) with the number of data points. For larger datasets (>1000 points):

• Implement sparse Gaussian process approximations
• Use inducing point methods to reduce computational complexity
• Consider alternative surrogate models (random forests, Bayesian neural networks) for very large datasets [32]

The following diagram illustrates the strategic decision process for selecting appropriate exploration-exploitation strategies based on experimental constraints:

Balancing exploration and exploitation in Bayesian optimization provides a powerful framework for autonomous protein engineering, enabling efficient navigation of vast sequence spaces while managing experimental constraints. The integration of Gaussian process surrogate models with carefully selected acquisition functions allows researchers to systematically trade off between investigating uncertain regions and refining promising solutions. Implementation in automated biofoundries demonstrates that these approaches can accelerate protein engineering campaigns by orders of magnitude, as evidenced by successful enzyme and antibody optimization case studies. As these methodologies continue to mature, they promise to further democratize and accelerate the development of novel biocatalysts and biotherapeutics.

The conceptualization of protein fitness landscapes has been profoundly shaped by Sewall Wright's metaphor of adaptive topography, where populations move from low to high fitness areas [35]. However, empirical evidence increasingly suggests that the evolution of quantitative traits in proteins is more consistent with navigation across high-dimensional "holey landscapes" rather than smooth, Gaussian landscapes [35]. These holey landscapes present unique challenges for protein engineers, as they feature extensive neutral plateaus punctuated by non-viable regions where specific trait combinations lead to dramatic fitness losses.

Multi-Output Gaussian Process (MOGP) models have emerged as powerful computational tools for addressing these challenges in autonomous protein engineering platforms. By simultaneously modeling multiple correlated protein properties and their complex interdependencies, MOGPs provide a robust statistical framework for navigating the discontinuous, high-dimensional fitness landscapes that characterize real protein systems. This Application Note details the implementation, performance, and integration of MOGP models within autonomous enzyme engineering platforms, providing researchers with standardized protocols for leveraging these advanced machine learning approaches.

Quantitative Performance Analysis of MOGP Frameworks

Recent benchmarking studies demonstrate that advanced MOGP implementations significantly outperform traditional single-output models and earlier MOGP alternatives across multiple performance metrics. The table below summarizes the comparative performance of state-of-the-art MOGP frameworks:

Table 1: Performance metrics of Graphical MOGP (GMOGP) versus alternative approaches across benchmark datasets

Model	Predictive Accuracy (R²)	Time Efficiency (s)	Memory Efficiency (MB)	Uncertainty Quantification
GMOGP (Proposed)	0.89 ± 0.05	124.3 ± 15.2	42.7 ± 3.1	Excellent
Linear Model of Coregionalization	0.76 ± 0.07	287.6 ± 24.8	78.9 ± 5.4	Good
Intrinsic Coregionalization Model	0.71 ± 0.08	198.5 ± 18.3	65.3 ± 4.2	Fair
Convolutional MOGP	0.82 ± 0.06	452.7 ± 36.1	123.5 ± 8.7	Good

The Graphical MOGP framework achieves Pareto optimal solutions through a distributed learning framework that simultaneously optimizes predictive performance, computational efficiency, and model flexibility [36]. This makes it particularly suited for autonomous protein engineering applications where rapid iteration is essential.

Applications in Autonomous Protein Engineering

Multi-Property Enzyme Optimization

Autonomous enzyme engineering platforms have successfully leveraged MOGPs to optimize multiple enzyme properties concurrently. In a recent case study, researchers engineered Arabidopsis thaliana halide methyltransferase (AtHMT) for improved substrate preference and catalytic activity using an integrated workflow that combined MOGPs with large language models [6]. The platform achieved a 90-fold improvement in substrate preference and 16-fold improvement in ethyltransferase activity in just four rounds of engineering while screening fewer than 500 variants [6].

The MOGP model was instrumental in modeling the correlated nature of substrate specificity and catalytic efficiency, identifying variants that balanced trade-offs between these properties. Similarly, for Yersinia mollaretii phytase (YmPhytase), MOGP-guided engineering yielded a variant with 26-fold improvement in activity at neutral pH, addressing the challenge of pH-dependent activity loss that limits application in animal feed [6].

Spatial Fitness Landscape Modeling

MOGPs have found significant application in modeling spatial variability across protein fitness landscapes. Researchers have employed these models to jointly analyze the spatial distribution of multiple cancer risk factors across geographical maps, with datasets comprising approximately 2000 spatial observations [37]. The Coregion regression approach in MOGPs provides a mathematical foundation equivalent to co-kriging methods from geostatistics, enabling efficient information transfer between correlated fitness objectives [37].

In mining applications, MOGPs have been used to model multiple mineral measurements at the same sites, demonstrating their utility in extracting maximal information from limited experimental samples [37]. This capability is particularly valuable in protein engineering where high-throughput experimental characterization remains costly and time-consuming.

Experimental Protocols

Protocol 1: MOGP Implementation for Protein Fitness Prediction

Purpose: To implement a Graphical Multioutput Gaussian Process (GMOGP) with attention mechanisms for predicting multiple protein fitness properties from sequence data.

Materials:

• Protein sequence data (variant libraries with multiple mutations)
• Experimentally measured fitness values for multiple properties (e.g., activity, stability, expression)
• Computational resources: Linux workstation with minimum 32GB RAM, GPU recommended
• Software: Python 3.8+, GPyTorch or GPflow libraries, custom GMOGP implementation [36]

Procedure:

• Data Preparation:
  • Encode protein sequences using biophysical features or learned embeddings from protein language models (e.g., ESM-2) [6]
  • Standardize fitness values for each property to zero mean and unit variance
  • Split data into training (80%), validation (10%), and test (10%) sets

• Model Configuration:

  • Initialize graphical structure defining relationships between output dimensions
  • Set attention mechanism parameters for weighting parent contributions
  • Configure kernel functions for each output dimension (Matérn 5/2 recommended)

• Model Training:

  • Optimize hyperparameters via distributed learning framework
  • Train for 1000 epochs with early stopping (patience=50)
  • Validate on holdout set to prevent overfitting

• Prediction and Uncertainty Quantification:

  • Generate predictions for candidate sequences
  • Extract posterior mean and variance for each fitness property
  • Calculate acquisition function values for experimental design

Troubleshooting:

• For memory issues with large datasets, implement inducing point approximations
• If convergence problems occur, adjust learning rate or kernel lengthscales
• Address identifiability issues in coregionalization matrices through regularization

Protocol 2: Autonomous DBTL Cycle with MOGP Guidance

Purpose: To execute an autonomous Design-Build-Test-Learn (DBTL) cycle for protein engineering using MOGP predictions to guide experimental efforts.

Materials:

• Illinois Biological Foundry for Advanced Biomanufacturing (iBioFAB) or equivalent automated platform [6]
• High-fidelity DNA assembly system (e.g., HiFi-assembly based mutagenesis)
• High-throughput functional assay capabilities
• Laboratory automation scheduling software (e.g., Thermo Momentum)

Procedure:

• Design Phase:
  • Generate initial variant library using combination of protein LLM (ESM-2) and epistasis model (EVmutation) [6]
  • Select 180 diverse variants covering sequence space while focusing on promising regions
  • Prioritize variants using MOGP predictions balanced with uncertainty sampling

• Build Phase:

  • Execute automated mutagenesis via HiFi-assembly method (95% accuracy) [6]
  • Perform microbial transformations in 96-well format
  • Conduct plasmid purification without intermediate sequence verification

• Test Phase:

  • Express proteins in automated fermentation system
  • Perform functional enzyme assays in high-throughput format
  • Measure multiple fitness properties simultaneously (e.g., activity, specificity, stability)

• Learn Phase:

  • Train MOGP on expanded dataset incorporating new experimental results
  • Identify promising regions of sequence space for next DBTL cycle
  • Select 96 variants for subsequent iteration based on MOGP predictions

Validation:

• Sequence random mutants to confirm targeted mutations (expect ~95% accuracy)
• Include wild-type controls in each experimental batch
• Replicate top performers to confirm functional improvements

Workflow Visualization

MOGP-Enhanced Autonomous Protein Engineering

Graphical MOGP Architecture with Attention

Research Reagent Solutions

Table 2: Essential research reagents and computational tools for MOGP-driven protein engineering

Category	Specific Solution	Function	Implementation Example
Computational Models	ESM-2 (Evolutionary Scale Modeling)	Protein language model for variant prioritization	Generate likelihood scores for amino acid substitutions [6]
	EVmutation	Epistasis model for identifying interacting mutations	Model higher-order interactions in fitness landscapes [6]
	Graphical MOGP	Multi-output regression with uncertainty quantification	Predict multiple enzyme properties simultaneously [36]
Experimental Platforms	iBioFAB (Illinois Biological Foundry)	Fully automated biofoundry for DBTL cycles	End-to-end automated protein engineering [6]
	HiFi-assembly Mutagenesis	High-fidelity DNA construction method	Generate variant libraries with ~95% accuracy [6]
	SAPP (Semi-Automated Protein Production)	High-throughput protein expression and purification	48-hour turnaround from DNA to purified protein [14]
Analysis Tools	DMX (DNA Multiplexing)	Cost-effective DNA variant construction	5-8x cost reduction for DNA synthesis [14]
	Automated SEC Analysis	High-throughput protein characterization	Simultaneous assessment of purity, yield, and oligomeric state [14]

Multi-Output Gaussian Process models represent a transformative approach for navigating the complex, holey fitness landscapes that characterize protein engineering challenges. By effectively modeling correlations between multiple protein properties and quantifying prediction uncertainty, MOGPs enable more efficient exploration of sequence space within autonomous engineering platforms. The integration of these advanced statistical models with automated experimental systems creates a powerful framework for addressing the persistent challenge of navigating high-dimensional fitness landscapes with limited experimental data.

As protein engineering continues to embrace autonomous methodologies, MOGPs will play an increasingly critical role in bridging the gap between computational design and experimental validation. The protocols and applications detailed in this document provide researchers with practical guidance for implementing these methods, accelerating the development of novel enzymes with tailored functions for biotechnology, medicine, and sustainable chemistry.

Proof and Impact: Validating Performance Against Traditional Methods

The field of protein engineering is undergoing a revolutionary shift, moving from traditional processes that could take years to modern autonomous platforms that deliver results in a matter of weeks. This dramatic acceleration is made possible by integrating artificial intelligence (AI) with fully automated robotic biofoundries, creating closed-loop systems that operate with minimal human intervention. These platforms are demonstrating unprecedented efficiency in navigating complex protein fitness landscapes, achieving engineering goals that were previously impractical or prohibitively time-consuming. The emergence of platforms like SAMPLE (Self-driving Autonomous Machines for Protein Landscape Exploration) represents a paradigm shift in how researchers approach protein optimization, enabling rapid exploration of sequence spaces that would be impossible to navigate manually [9]. This application note details the protocols, workflows, and performance metrics of these autonomous systems, providing researchers with a comprehensive benchmarking framework for evaluating next-generation protein engineering technologies.

Quantitative Performance Benchmarking

The performance differential between traditional and autonomous protein engineering methods is quantifiable across multiple dimensions, including timeline, throughput, and functional improvement. The table below summarizes key performance indicators from recent implementations of autonomous platforms.

Table 1: Benchmarking Autonomous Protein Engineering Platforms

Platform / Study	Engineering Goal	Timeframe	Variants Constructed & Screened	Key Functional Improvement
Generalized AI-Powered Platform [6]	Improve substrate preference & activity of AtHMT; Enhance neutral pH activity of YmPhytase	4 weeks	<500 variants per enzyme	90-fold improvement in substrate preference; 26-fold improvement in neutral pH activity
PLMeAE Platform [29]	Improve activity of tRNA synthetase	10 days	384 variants over 4 rounds	2.4-fold improvement in enzyme activity
SAMPLE Platform [9]	Enhance thermal tolerance of glycoside hydrolase	20 rounds (continuous operation)	3 sequences per round	>12°C improvement in thermostability
Autonomous Lab (ANL) [38]	Optimize medium conditions for glutamic acid production	Continuous operation	Multiple parallel experiments	Improved cell growth parameters

The data demonstrates that autonomous platforms consistently achieve significant engineering goals within weeks, a timeline that contrasts sharply with traditional directed evolution campaigns that often require multiple years to complete [29] [9]. This acceleration is coupled with remarkable efficiency in library design, as these systems typically require construction and testing of only hundreds to low-thousands of variants, a small fraction of the sequence space that must be explored empirically with traditional methods.

Core Experimental Protocols for Autonomous Engineering

The transformative speed and efficiency of autonomous protein engineering are enabled by standardized, automated protocols that execute the Design-Build-Test-Learn (DBTL) cycle without human intervention.

The Generalized Autonomous DBTL Workflow

The following diagram illustrates the continuous, automated workflow that forms the backbone of platforms like SAMPLE and the Illinois Biological Foundry (iBioFAB).

Diagram 1: Autonomous DBTL Workflow for Protein Engineering

Detailed Module Protocols

Module 1: AI-Driven Protein Design

Objective: To generate a high-quality, diverse library of protein variants for experimental testing using computational models.

Reagents & Specifications:

• Input Protein Sequence: Wild-type or parent sequence in FASTA format.
• Computational Models: Protein Language Models (PLMs) such as ESM-2 [6] [29] and epistasis models like EVmutation [6].
• Hardware: High-performance computing cluster.

Procedure:

• Sequence Encoding: The input protein sequence is tokenized into its constituent amino acids, represented by their one-letter codes [39].
• Zero-Shot Prediction (Module I): For proteins without predefined mutation sites, the PLM performs a zero-shot analysis. Each amino acid position is virtually masked, and the model calculates the likelihood of all possible substitutions at that site, ranking them based on predicted fitness [29].
• Targeted Design (Module II): If mutation sites are known (e.g., from prior experiments or structural analysis), the PLM samples multi-mutant combinations specifically at these residues to predict high-fitness variants [29].
• Library Finalization: The top-ranked variants (e.g., 96-180 candidates) are selected for the first round of experimental testing [6] [29].

Module 2: Automated Build Pipeline on the Biofoundry

Objective: To physically construct the DNA sequences encoding the designed protein variants and express them as proteins, fully automated using robotic systems.

Reagents & Specifications:

• DNA Construction: Utilizes high-fidelity (HiFi) assembly PCR or Golden Gate cloning with pre-synthesized DNA fragments [6] [9].
• Expression System: Cell-free protein expression systems or microbial transformations in 96-well plates [6] [9].
• Core Equipment: Liquid handlers (e.g., Opentrons OT-2), thermocyclers, robotic arms (e.g., Brooks PF400), and incubators [38] [29].

Procedure:

• Automated Mutagenesis: The biofoundry's liquid handlers prepare mutagenesis PCR reactions based on a digital worklist from the Design phase. DpnI digestion is performed to eliminate template DNA [6].
• DNA Assembly: For platforms like SAMPLE, a Golden Gate assembly reaction is robotically assembled using predefined DNA fragments to build the full gene [9]. The iBioFAB platform employs a HiFi-assembly method that eliminates the need for intermediate sequence verification, ensuring a continuous workflow [6].
• Protein Expression: The assembled DNA is transferred directly into a cell-free expression system [9] or used to transform microbial cells (e.g., E. coli) in a 96-well format. After an incubation period, cells are lysed to release the expressed protein variants [6].

Module 3: High-Throughput Testing and Assaying

Objective: To quantitatively measure the fitness (e.g., activity, stability) of each expressed protein variant.

Reagents & Specifications:

• Assay Reagents: Substrate-specific chromogenic or fluorogenic reagents for enzymatic assays.
• Equipment: Microplate readers (e.g., SpectraMax iD3), centrifuges (e.g., BioNex HiG), and LC-MS/MS systems for metabolite quantification [6] [38].

Procedure:

• Sample Preparation: Crude cell lysates containing the expressed variants are clarified, often using an automated centrifugation step [6] [38].
• Functional Assay: An aliquot of each variant is dispensed into a 96-well assay plate. The reaction is initiated by adding the relevant substrate.
  • Thermostability Assay (SAMPLE Protocol): Enzyme activity is measured at a gradient of temperatures. The T50 value (temperature at which 50% of activity is lost) is calculated by fitting the activity-temperature data to a sigmoid curve [9].
  • Activity Assay: For enzymes like phytase or methyltransferase, the consumption of substrate or generation of product is monitored spectrophotometrically over time [6].
• Data Quality Control: The platform automatically performs exception handling, checking for successful DNA assembly (via dyes like EvaGreen), valid reaction progress curves, and activity above background levels [9].

Module 4: Machine Learning for Learning and Redesign

Objective: To use experimental data to refine the AI agent's understanding of the sequence-function relationship and propose improved variants for the next cycle.

Reagents & Specifications:

• Software: Machine learning models, typically Bayesian optimization (BO) with Gaussian Processes (GP) or supervised models like Multi-Layer Perceptrons (MLPs) [29] [9].

Procedure:

• Data Integration: The quantitative fitness data from the Test phase is linked to the corresponding protein sequences.
• Model Training: A supervised ML model (e.g., an MLP) is trained to map sequence encodings (provided by the PLM) to the experimental fitness data, creating a custom fitness predictor [29]. Alternatively, a Bayesian optimization agent uses a GP to model the landscape, often incorporating a classifier to predict whether a sequence is functional before predicting its fitness [9].
• Variant Proposal: The trained model scores a vast number of in silico variants. Optimization algorithms like Upper Confidence Bound (UCB) are used to select the next set of variants to test, balancing exploration of uncertain regions with exploitation of known high-fitness areas [9].
• Cycle Iteration: The new list of proposed variants is sent automatically to the Build module, initiating the next DBTL cycle.

The Scientist's Toolkit: Research Reagent Solutions

Successful implementation of an autonomous protein engineering pipeline relies on a suite of specialized reagents and computational tools. The following table catalogs the essential components.

Table 2: Key Research Reagents and Solutions for Autonomous Protein Engineering

Item Name	Function / Application	Specifications & Notes
ESM-2 (Evolutionary Scale Modeling)	A protein language model used for zero-shot prediction of variant fitness and sequence representation learning.	A transformer-based model; provides a likelihood score for amino acid substitutions that can be interpreted as fitness [6] [29].
Cell-Free Protein Expression System	A machinery for in vitro transcription and translation, bypassing the need for living cells.	Enables rapid protein expression (∼3 hours in SAMPLE platform); ideal for high-throughput screening and expressing potentially toxic proteins [9].
Golden Gate Assembly Mix	A modular DNA assembly method using Type IIS restriction enzymes.	Used in SAMPLE to combinatorially assemble pre-synthesized DNA fragments into full genes [9].
HiFi Assembly Mix	A high-fidelity DNA assembly method for mutagenesis.	Used in the iBioFAB platform for error-prone PCR and site-directed mutagenesis, achieving ~95% accuracy without intermediate verification [6].
EvaGreen Dye	A fluorescent dye that binds double-stranded DNA.	Used for quality control in automated pipelines to verify successful gene assembly and PCR amplification [9].
Gaussian Process (GP) Model	A Bayesian machine learning model for modeling sequence-function landscapes.	The core of the SAMPLE agent; models both continuous fitness and a binary classification of protein activity/inactivity [9].

Autonomous protein engineering platforms have decisively shifted the benchmark for project timelines from years to weeks. This step-change in efficiency is underpinned by tightly integrated DBTL cycles, where AI agents and robotic biofoundries operate in a closed loop. The standardized protocols for AI-driven design, automated DNA construction, high-throughput assays, and iterative machine learning provide a robust framework for tackling a wide array of protein engineering challenges. As these platforms continue to evolve, they promise to further accelerate the development of novel enzymes and therapeutics, fundamentally reshaping the landscape of biotechnology research and development.

The advent of autonomous protein engineering platforms represents a paradigm shift in synthetic biology, moving the field from a bespoke, specialist-dependent craft to a scalable, data-driven science. These platforms integrate artificial intelligence (AI), robotics, and biofoundry automation to execute iterative Design-Build-Test-Learn (DBTL) cycles with minimal human intervention [6] [5]. For researchers and drug development professionals, the value of these systems is ultimately quantified by their ability to generate enzymes with enhanced functional properties—specifically, activity and stability. This Application Note documents specific, quantitative improvements in enzyme performance achieved through cutting-edge autonomous and machine learning-driven platforms, providing detailed protocols and data to inform future research and development efforts.

Documented Quantitative Improvements

Recent studies demonstrate the efficacy of autonomous platforms in achieving significant enhancements in enzyme performance over remarkably short timeframes. The quantitative data below summarize key successes.

Table 1: Documented Improvements in Enzyme Activity via an Autonomous AI Platform

Enzyme	Engineering Goal	Key Improvement	Experimental Scale	Timeframe
Arabidopsis thaliana Halide Methyltransferase (AtHMT)	Improve ethyltransferase activity & substrate preference	~16-fold increase in ethyltransferase activity; ~90-fold shift in substrate preference [6]	4 rounds; <500 variants screened [6]	4 weeks [6]
Yersinia mollaretii Phytase (YmPhytase)	Enhance activity at neutral pH	~26-fold higher specific activity at neutral pH [6]	4 rounds; <500 variants screened [6]	4 weeks [6]

Table 2: Documented Improvements in Enzyme Stability via ML-Guided and Short-Loop Engineering

Enzyme	Engineering Strategy	Key Improvement (Stability)	Key Improvement (Activity)
Lactate Dehydrogenase (Pediococcus pentosaceus)	Short-loop engineering [40]	Half-life 9.5x higher than wild-type [40]	Not Specified
Urate Oxidase (Aspergillus flavus)	Short-loop engineering [40]	Half-life 3.11x higher than wild-type [40]	Not Specified
Protein-Glutaminase (PG)	Machine Learning (iCASE strategy) [41]	Slightly increased thermal stability [41]	Specific activity up to 1.82-fold higher [41]
Xylanase (XY)	Machine Learning (iCASE strategy) [41]	Melting temperature (Tm) increased by 2.4 °C [41]	Specific activity 3.39-fold higher [41]

Experimental Protocols for Autonomous Enzyme Engineering

Protocol: Generalized Autonomous Engineering DBTL Cycle

This protocol outlines the core workflow for autonomous enzyme engineering as implemented on the Illinois Biological Foundry for Advanced Biomanufacturing (iBioFAB) [6].

I. Design Phase

• Input Requirements: The process requires only two inputs: the wild-type protein sequence and a quantifiable assay to measure fitness (e.g., enzymatic activity under specific conditions) [6].
• Initial Library Design: For the first cycle, use unsupervised models to design a diverse and high-quality variant library.
  • Protein Language Model: Utilize a pretrained transformer model like ESM-2 to predict the likelihood of amino acids at specific positions, interpreting the likelihood as variant fitness [6].
  • Epistasis Model: Simultaneously, use a model like EVmutation, which focuses on local homologs of the target protein, to account for residue-residue interactions [6].
  • Curation: Combine the outputs of both models to generate a list of ~180 single-point mutants for initial experimental testing [6].
• Iterative Design: In subsequent cycles, employ a supervised machine learning model (a "low-N" regression model) trained on the experimentally characterized variants from previous rounds. This model predicts the fitness of new variants, often by combining beneficial single mutations into higher-order combinations [6].

II. Build Phase This phase is fully automated on the biofoundry.

• High-Fidelity Mutagenesis: Perform site-directed mutagenesis using an optimized HiFi-assembly-based method. This method eliminates the need for intermediate sequence verification, achieving approximately 95% accuracy and enabling a continuous workflow [6].
• Plasmid Construction & Verification: The robotic system automates cloning, plasmid transformation, and colony picking. Randomly selected mutants can be sequenced to confirm accuracy [6].
• Protein Expression: Automate cell culture, protein expression induction, and cell harvesting in a 96-well format [6].

III. Test Phase

• Sample Preparation: The platform automates crude cell lysis to prepare samples for functional assay [6].
• Fitness Assay: Perform an automated, high-throughput enzymatic assay. The specific assay is defined by the user's fitness goal (e.g., for AtHMT, alkyltransferase activity was measured) [6]. The assay data for each variant is collected and stored in a centralized database for the Learn phase.

IV. Learn Phase

• Data Analysis: The collected fitness data is used to retrain and refine the supervised machine learning model, improving its predictive power for the next DBTL cycle [6].
• Hypothesis Generation: The AI system autonomously analyzes the results to propose the next set of variants to test in the subsequent cycle, closing the loop [6].

The following workflow diagram illustrates this integrated, autonomous process:

Protocol: Machine Learning-Guided iCASE Strategy

For environments without full biofoundry automation, the computational iCASE (isothermal compressibility-assisted dynamic squeezing index perturbation engineering) strategy provides a powerful alternative for balancing stability and activity [41].

I. Target Selection and Analysis

• Identify High-Fluctuation Regions: Calculate the isothermal compressibility (βT) across the enzyme's structure to identify flexible regions (e.g., specific loops, α-helices) that are critical for dynamics [41].
• Map Active Site: Perform molecular docking to identify residues involved in substrate binding and catalysis [41].

II. In Silico Mutant Screening

• Calculate Dynamic Squeezing Index (DSI): Compute the DSI for residues in the high-fluctuation regions. Residues with a DSI > 0.8 (top 20%) are selected as candidate mutation sites [41].
• Predict Energetic Effects: Use a computational tool like Rosetta to predict the change in folding free energy (ΔΔG) upon mutation for the candidate residues. Filter out mutations with highly destabilizing ΔΔG values [41].
• Final Candidate Selection: Select 10-15 single-point mutants for experimental characterization based on a combination of high DSI and favorable ΔΔG [41].

III. Experimental Validation and Combination

• Screen Single Mutants: Express and purify the selected single-point mutants. Measure specific activity and thermal stability (e.g., via melting temperature, Tm, or half-life) [41].
• Combine Beneficial Mutations: Combine positive mutations from the single-point screen to construct higher-order (double, triple) mutants. Characterize these combined variants to identify synergistic effects that simultaneously enhance both activity and stability [41].

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Reagents and Platforms for Autonomous Enzyme Engineering

Item / Platform	Function / Application	Reference
Illinois Biofoundry (iBioFAB)	A fully automated robotic platform that executes end-to-end molecular biology workflows, from DNA construction to cell culture and assay.	[6]
Protein Language Models (e.g., ESM-2)	Unsupervised AI models that predict beneficial mutations from evolutionary patterns in protein sequences, used for initial library design.	[6]
Epistasis Models (e.g., EVmutation)	Computational models that identify interacting residues to guide the design of combinatorial libraries.	[6]
Low-N Machine Learning Models	Supervised regression models trained on small datasets (N<500) from initial rounds to predict variant fitness in subsequent cycles.	[6]
HiFi-Assembly Mutagenesis	A high-fidelity DNA assembly method that eliminates the need for intermediate sequencing, enabling continuous automated workflows.	[6]
Sodium Alginate-Modified Rice Husk Beads	A robust and reusable immobilization support for enzymes, enhancing pH, temperature, and storage stability for application in bioremediation.	[42]

Workflow Visualization: Autonomous vs. ML-Guided Strategies

The following diagram contrasts the fully autonomous platform with a computation-heavy strategy like iCASE, highlighting different pathways to quantitative success.

Protein engineering is a cornerstone of modern biotechnology, enabling the development of enzymes, therapeutics, and biosensors with enhanced properties. For decades, the field has been dominated by two primary methodologies: rational design and directed evolution. Rational design employs precise, knowledge-driven modifications to protein sequences, while directed evolution mimics natural selection through iterative rounds of mutagenesis and screening [43] [44]. Although these methods have proven successful, they often face limitations in efficiency, throughput, and the requirement for extensive domain expertise.

The emergence of autonomous protein engineering platforms represents a paradigm shift, integrating artificial intelligence (AI), robotic automation, and advanced computational models to create self-driving laboratories. This analysis provides a comparative assessment of these methodologies, focusing on their operational frameworks, performance metrics, and practical applications. The content is framed within the context of the SAMPLE (Self-driving Autonomous Machines for Protein Landscape Exploration) research initiative, which aims to fully automate the process of protein design and testing [44].

Comparative Workflow Analysis

The fundamental distinction between traditional and autonomous methods lies in their operational workflows and the degree of human intervention required.

Traditional Protein Engineering Workflows

• Rational Design: This approach functions like architectural planning. It requires deep prior knowledge of protein structure and function to introduce specific, targeted mutations. It relies heavily on computational models to predict how these modifications will affect protein performance. While its precision is an advantage, its success is constrained by the completeness of existing structural data [44] [45].
• Directed Evolution: This method mimics natural evolution in a laboratory setting. It involves creating large libraries of random protein variants, typically via techniques like error-prone PCR, and then screening or selecting for individuals with desirable traits. This process is iterative, with multiple rounds of mutation and selection. Its main strength is that it does not require prior structural knowledge, but it can be resource-intensive and time-consuming [43] [44].

The Autonomous Platform Workflow

Autonomous platforms, such as the one exemplified by the Illinois Biological Foundry (iBioFAB), integrate and automate the classic Design-Build-Test-Learn (DBTL) cycle. The process is continuous and requires minimal human intervention [6]:

• Design: Machine learning models, including protein Large Language Models (LLMs) like ESM-2 and epistasis models, are used to design a diverse and high-quality initial library of protein variants [6].
• Build: A fully automated robotic system executes molecular biology protocols, including mutagenesis PCR, DNA assembly, transformation, and plasmid purification, to construct the physical variant library [6].
• Test: The platform conducts high-throughput protein expression and functional assays, automatically collecting fitness data on each variant [6].
• Learn: The experimental data is used to retrain and refine the machine learning models, which then propose a new and improved set of variants for the next cycle, closing the loop [6].

This integrated feedback loop allows for a more efficient exploration of the protein sequence space compared to traditional methods.

Performance Metrics and Quantitative Comparison

The superiority of autonomous platforms is demonstrated by direct comparisons of key performance indicators, such as engineering efficiency, timeline, and library size requirements.

Table 1: Quantitative Comparison of Protein Engineering Methodologies

Metric	Rational Design	Directed Evolution	Autonomous Platforms
Typical Timeline	Weeks to months [44]	Months to years [43]	~4 weeks for 4 rounds [6]
Library Size per Round	Small (Targeted variants) [44]	Very Large (10^3 - 10^6 variants) [43]	Focused (<500 variants per round) [6]
Human Intervention	High (Expert-driven design) [45]	High (Screening & selection) [43]	Minimal (Fully autonomous operation) [6]
Required Prior Knowledge	High (Structure & mechanism) [44]	Low	Low (Only sequence & fitness metric) [6]
Reported Improvement Fold	Varies widely	Varies widely	16-fold to 90-fold (in substrate preference) [6]

Further data underscores the performance of specific algorithms. The DeepDE algorithm, which uses supervised learning on ~1,000 mutants and explores triple-mutation spaces, achieved a 74.3-fold increase in GFP activity over four rounds, far surpassing the benchmark superfolder GFP [46]. Another autonomous platform, focusing on iterative cycles, successfully engineered a halide methyltransferase (AtHMT) for a 90-fold improvement in substrate preference and a phytase (YmPhytase) with a 26-fold improvement in activity at neutral pH. This was accomplished in four rounds over 4 weeks, while requiring the construction and characterization of fewer than 500 variants for each enzyme [6].

Detailed Experimental Protocols

To illustrate the practical implementation of these methodologies, below are detailed protocols for a classic directed evolution campaign and a modern autonomous workflow.

Protocol 1: Directed Evolution via Error-Prone PCR

This protocol outlines a standard manual directed evolution pipeline for improving enzyme activity [43] [47].

• Step 1: Library Generation via Error-Prone PCR

  • Set up a 50 µL PCR reaction containing: template DNA (10-100 ng), primers flanking the gene of interest, standard PCR buffer, dNTPs, MgCl2 (supplemented to 1-5 mM to increase error rate), and 0.5-1 unit of Taq polymerase (which lacks proofreading activity).
  • Run PCR with the following cycling conditions: initial denaturation at 95°C for 2 min; 25-30 cycles of 95°C for 30 sec, 55°C for 30 sec, 72°C for 1 min/kb; final extension at 72°C for 5 min.
  • Purify the PCR product and clone it into an expression vector using standard molecular biology techniques (e.g., restriction digestion/ligation or homologous recombination).

• Step 2: Transformation and Screening

  • Transform the ligated plasmid library into a competent expression host (e.g., E. coli).
  • Plate transformed cells onto agar plates with antibiotic selection and incubate overnight.
  • Manually pick hundreds to thousands of individual colonies into 96- or 384-well plates containing culture medium.
  • Induce protein expression and perform cell lysis, either chemically or enzymatically.
  • Assay the lysates for the desired enzymatic activity. This may involve colorimetric/fluorimetric assays in plates or, for more complex analyses, using HPLC/MS.

• Step 3: Hit Identification and Iteration

  • Identify variants (hits) that show improved activity over the wild type.
  • Sequence the genes of the best hits to identify beneficial mutations.
  • Use the best variant as the new template for the next round of error-prone PCR.
  • Repeat steps 1-3 for 3-5 rounds or until a satisfactory improvement is achieved.

Protocol 2: Autonomous Engineering on an Integrated Biofoundry

This protocol describes the end-to-end automated workflow as implemented on the iBioFAB for the engineering of AtHMT and YmPhytase [6].

• Step 1: Computational Design of Initial Library

  • Input: Provide the wild-type protein sequence and a defined fitness objective (e.g., improved activity at neutral pH).
  • Process: A combination of a protein LLM (ESM-2) and an epistasis model (EVmutation) is used to generate a list of ~180 single-point mutations predicted to be beneficial.
  • Output: A finalized list of variants for construction.

• Step 2: Automated Library Construction (iBioFAB)

  • The robotic system executes a high-fidelity (HiFi) assembly-based mutagenesis method to generate variant plasmids, eliminating the need for intermediate sequencing.
  • The workflow is modular, with automated steps for mutagenesis PCR, DpnI digestion, assembly, transformation into E. coli, colony picking, and plasmid purification.
  • A central robotic arm schedules and integrates all instruments (liquid handlers, thermocyclers, incubators).

• Step 3: High-Throughput Characterization

  • The platform inoculates cultures in deep-well plates for protein expression.
  • After growth and induction, cells are lysed, and crude lysates are transferred to assay plates.
  • A plate reader performs the functional enzyme assay, and data is automatically recorded in a centralized database.

• Step 4: Machine Learning-Guided Iteration

  • The assay data for all variants is used to train a low-N machine learning model (e.g., ridge regression) to map sequence to function.
  • The trained model predicts the fitness of a vast number of potential higher-order mutants (e.g., combinations of the initial single mutants).
  • The top predicted variants (typically double or triple mutants) are selected for the next round of construction and testing.
  • This DBTL cycle repeats autonomously for 3-5 rounds.

The Scientist's Toolkit: Key Research Reagent Solutions

The successful implementation of these protein engineering strategies relies on a suite of specialized reagents, computational tools, and hardware.

Table 2: Essential Research Reagents and Tools for Protein Engineering

Item Name	Function/Description	Application Context
ESM-2 (LLM)	A large language model trained on protein sequences; predicts amino acid likelihoods and variant fitness from sequence context [6].	Autonomous Design: Used for generating intelligent initial variant libraries.
OrthoRep System	An in vivo continuous evolution system with a dedicated error-prone polymerase for targeted mutagenesis in yeast [48].	Directed Evolution: Enables rapid generation of mutant libraries within the host.
Error-Prone PCR Kit	A optimized reagent mix (e.g., with Mn2+) to introduce random mutations during PCR amplification [43].	Directed Evolution: Standard method for creating random mutant libraries.
ProteinMPNN	A neural network for designing amino acid sequences given a protein backbone structure; aids in de novo design and refining protein properties [49].	Rational & Autonomous Design: For de novo protein design or fixing structural issues.
iBioFAB / AutoEvoLab	A fully integrated robotic biofoundry capable of automating molecular biology and screening workflows with minimal human intervention [6] [48].	Autonomous Platforms: The physical hardware for the "Build" and "Test" phases.
AlphaFold2 / RoseTTAFold	AI-powered tools for highly accurate protein structure prediction from amino acid sequences [49].	Rational Design: Provides critical 3D structural data for targeted mutagenesis.

Workflow Visualization

The following diagrams illustrate the core logical workflows for the traditional and autonomous protein engineering cycles, highlighting key differences in integration and iteration.

Traditional DBTL Cycle

The traditional Design-Build-Test-Learn cycle is characterized by significant manual intervention and discrete, often disconnected, phases. This can lead to bottlenecks and slower iteration times.

Autonomous Engineering Cycle

The autonomous cycle is characterized by a tightly integrated, automated loop where machine learning directly uses experimental results to inform the next design round, drastically accelerating the optimization process.

The comparative analysis presented herein clearly demarcates the operational and performance boundaries between traditional protein engineering methods and emerging autonomous platforms. While rational design and directed evolution remain powerful and widely used tools, autonomous platforms like SAMPLE offer a transformative approach by integrating AI and robotics into a closed-loop system. The quantitative data demonstrates their ability to achieve significant functional improvements in proteins on a timescale of weeks, rather than months or years, and with remarkably focused library sizes.

The future of protein engineering lies in the synergistic combination of these methodologies. The exploratory power of directed evolution and the precision of rational design can be powerfully amplified when embedded within an autonomous, data-generating platform. As these self-driving laboratories become more accessible and their underlying AI models continue to mature, they are poised to dramatically accelerate the pace of discovery and application in therapeutic development, industrial biocatalysis, and basic biological research.

In the field of autonomous protein engineering, the efficiency of the Design-Build-Test-Learn (DBTL) cycle is paramount. Two critical factors directly govern the success and speed of this iterative process: the scale and quality of the mutant libraries designed (Library Size) and the capacity to experimentally screen these libraries (Experimental Throughput). The emergence of integrated AI-platforms has dramatically enhanced both capabilities, enabling the exploration of vast sequence spaces with unprecedented precision and speed. This protocol examines the resource advantage conferred by modern autonomous systems, detailing the methodologies that allow researchers to leverage large library sizes and high experimental throughput for accelerated enzyme engineering. We frame this within the context of a generalized autonomous platform that has demonstrated the ability to achieve significant enzyme improvements in condensed timeframes [6].

Background

Traditional protein engineering campaigns, whether employing rational design or directed evolution, often face a fundamental trade-off between the comprehensiveness of sequence space exploration and practical experimental constraints. The size of the initial mutant library and the scale of testing were historically limited by resource availability, making exhaustive searches impractical.

The advent of autonomous protein engineering platforms has disrupted this paradigm. These systems integrate machine learning (ML), large language models (LLMs), and biofoundry automation to create a closed-loop DBTL cycle that operates with minimal human intervention [6] [5]. The core advantage lies in the platform's ability to intelligently design high-quality, diverse initial libraries using pre-trained models, and then to experimentally test these libraries at a scale and speed unattainable through manual methods. This synergy between computational power and robotic automation is redefining the resource landscape of protein engineering.

Key Principles and Quantitative Benchmarks

The performance of autonomous platforms can be quantified by their efficiency in navigating the protein fitness landscape. The following data illustrates the quantitative benchmarks achieved by a state-of-the-art system.

Table 1: Performance Metrics of an Autonomous Platform in Engineering Two Distinct Enzymes [6]

Enzyme	Engineering Goal	Library Variants Screened	Timeframe	Key Improvement
Arabidopsis thaliana halide methyltransferase (AtHMT)	Improve ethyltransferase activity & substrate preference	< 500	4 rounds / 4 weeks	16-fold increase in ethyltransferase activity; 90-fold shift in substrate preference
Yersinia mollaretii phytase (YmPhytase)	Enhance activity at neutral pH	< 500	4 rounds / 4 weeks	26-fold higher specific activity at neutral pH

The data in Table 1 demonstrates the profound efficiency of autonomous systems. By constructing and characterizing fewer than 500 variants for each enzyme, the platform achieved substantial functional improvements within a single month [6]. This efficiency is a direct result of the strategic integration of large-scale in silico library design with high-throughput experimental validation.

The principle of library size advantage is further supported by research in virtual ligand screening. A direct comparison of docking a 1.7 billion-molecule library versus a 99 million-molecule library against the model enzyme AmpC β-lactamase revealed that the larger library yielded a two-fold improvement in hit rates, discovered more new scaffolds, and produced compounds with improved potency [50]. This confirms that larger virtual libraries, when coupled with effective screening, directly enhance key outcomes. Furthermore, the study highlighted that testing only dozens of molecules, a common practice, leads to highly variable results; convergence in hit rates and affinities required testing several hundred molecules [50].

Experimental Protocols

Protocol 1: Automated Construction and Characterization of Protein Variants on a Biofoundry

This protocol details the automated workflow for the "Build" and "Test" phases within an autonomous enzyme engineering cycle, as implemented on the Illinois Biological Foundry for Advanced Biomanufacturing (iBioFAB) [6].

Key Research Reagent Solutions:

• HiFi Assembly Mix: For high-fidelity DNA assembly, crucial for eliminating intermediate sequence verification.
• DpnI Restriction Enzyme: Digests methylated template DNA post-mutagenesis PCR.
• LB Plates (Omnitray): For automated plating of microbial transformations.
• Cell Lysis Reagent: For crude cell lysate preparation in a 96-well format.
• Assay-Specific Substrates & Buffers: For functional enzyme assays tailored to the fitness objective (e.g., alkyl halides for AtHMT, phytic acid for YmPhytase).

Procedure:

• Module 1: Mutagenesis PCR. The biofoundry's robotic liquid handling system prepares PCR reactions using primers designed in silico to introduce the desired mutations. A high-fidelity DNA polymerase is used to minimize errors.
• Module 2: DpnI Digestion. The robotic system adds DpnI enzyme to the PCR product to digest the template plasmid, enriching for the newly synthesized mutant DNA.
• Module 3: DNA Assembly. The digested product undergoes a HiFi-assembly based mutagenesis method, which was specifically developed for this platform to achieve ~95% accuracy, thereby removing the need for sequence verification and ensuring workflow continuity [6].
• Module 4: Microbial Transformation. The assembled plasmid is transformed into competent E. coli cells in a 96-well format.
• Module 5: Colony Picking. The robotic arm picks individual colonies from the transformation plates and inoculates them into deep-well plates containing expression media.
• Module 6: Protein Expression. The system induces protein expression in the cultured cells.
• Module 7: Functional Assay. The platform performs automated crude cell lysate removal and conducts the functional enzyme assays. The assay data (fitness measurements) is automatically recorded for the subsequent "Learn" phase.

All instruments are scheduled via integrated software (e.g., Thermo Momentum) and orchestrated by a central robotic arm, enabling end-to-end automation without human intervention [6].

Protocol 2: Intelligent Design of Protein Variants Using Protein LLM and ML

This protocol covers the computational "Design" phase for generating initial and subsequent mutant libraries.

Key Research Reagent Solutions:

• Input Protein Sequence: The wild-type amino acid sequence of the target enzyme.
• Pre-trained Protein Language Model (e.g., ESM-2): A transformer model trained on global protein sequences to predict the likelihood of amino acids at specific positions based on sequence context [6].
• Epistasis Model (e.g., EVmutation): A statistical model that identifies co-evolved residues and patterns in multiple sequence alignments of protein homologs [6].
• Low-N Machine Learning Model: A regression model capable of learning from small datasets (N < 500) to predict variant fitness from sequence.

Procedure:

• Initial Library Design: a. Provide the wild-type protein sequence as input to the platform. b. The system uses the protein LLM (ESM-2) and the epistasis model (EVmutation) in combination to generate a list of candidate single-point mutations. c. The models rank these mutations based on predicted fitness, maximizing the diversity and quality of the initial library. d. For a proof-of-concept campaign, approximately 180 variants are selected for the first round of experimental screening [6].
• Iterative Library Design: a. Following the first round of experimentation, the experimentally measured fitness data for the ~180 variants is collected. b. This data is used to train a supervised low-N machine learning model, which learns the relationship between sequence composition and fitness for the specific enzyme. c. The trained model is used to predict the fitness of a vast in silico library of higher-order mutants (combinations of the best single mutations). d. The top-predicted variants (e.g., ~150-200) are selected for the next "Build" and "Test" cycle. This process repeats, with the model being re-trained with new data after each round, guiding the search toward optimal regions of the fitness landscape.

Workflow Visualization

The following diagram illustrates the integrated, closed-loop DBTL cycle of the autonomous enzyme engineering platform.

Diagram 1: Autonomous DBTL cycle for enzyme engineering.

The Scientist's Toolkit

Table 2: Essential Research Reagent Solutions for Autonomous Enzyme Engineering

Item	Function / Application in Protocol
Illinois Biological Foundry (iBioFAB)	A fully automated biofoundry that integrates robotic instruments to execute the entire Build-Test workflow, from DNA construction to functional assays [6].
Protein Language Model (ESM-2)	An unsupervised AI model used in the Design phase to intelligently propose beneficial mutations based on evolutionary patterns learned from millions of protein sequences [6].
Epistasis Model (EVmutation)	A computational model used alongside the LLM to analyze co-evolution in protein families, helping to design a high-quality initial variant library [6].
Low-N Machine Learning Model	A supervised learning algorithm trained on the small datasets (N < 500) generated each cycle to predict fitness and guide subsequent library design [6].
HiFi DNA Assembly Method	A high-fidelity DNA construction technique that achieves ~95% accuracy, enabling continuous automated workflow by eliminating the need for intermediate sequence verification [6].

Conclusion

Autonomous protein engineering platforms represent a fundamental shift in biological design, merging AI's predictive power with the tireless precision of robotics to create a potent, accelerated discovery engine. The validation of platforms like SAMPLE and others demonstrates their ability to reliably engineer proteins with enhanced properties—such as thermostability and catalytic activity—in a fraction of the time required by traditional methods. The key takeaways are the critical importance of the closed-loop DBTL cycle, the efficiency of AI-guided search strategies in navigating complex fitness landscapes, and the robustness of automated experimental workflows. For the future of biomedical and clinical research, this technology promises to dramatically shorten the timeline for developing novel therapeutics, including antibodies, enzymes, and targeted therapies, paving the way for more personalized and effective treatments. The next frontier will involve expanding these platforms to tackle even more complex protein systems and integrating them more deeply into the drug development pipeline, ultimately ushering in an era of personalized and on-demand protein therapeutics.

References

• 1. A generalized platform for artificial intelligence-powered ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12215622/]
• 2. Computational scoring and experimental evaluation of ... [https://www.nature.com/articles/s41587-024-02214-2]
• 3. New “evolution engine ” creates super- proteins 100... | ScienceDaily [https://www.sciencedaily.com/releases/2025/08/250807233038.htm]
• 4. Biophysics-based protein language models for ... [https://www.nature.com/articles/s41592-025-02776-2]
• 5. The Autonomous Lab: AI Platforms Redefining Enzyme ... [https://www.ailurus.bio/post/the-autonomous-lab-ai-platforms-redefining-enzyme-engineering]
• 6. A generalized platform for artificial intelligence-powered ... [https://www.nature.com/articles/s41467-025-61209-y]
• 7. Self-driving lab: AI and automated biology combine to improve ... [https://news.illinois.edu/self-driving-lab-ai-and-automated-biology-combine-to-improve-enzymes/]
• 8. An industrial automated laboratory for programmable ... [https://www.nature.com/articles/s44286-025-00303-w]
• 9. Self-driving laboratories to autonomously navigate the ... [https://www.nature.com/articles/s44286-023-00002-4]
• 10. 'Self-driving lab' speeds protein discovery process | Wisconsin ... [https://energy.wisc.edu/news/self-driving-lab-speeds-protein-discovery-process]
• 11. Biophysical fitness landscape design traps viral evolution [https://pmc.ncbi.nlm.nih.gov/articles/PMC11996392/]
• 12. Evaluation of machine learning-assisted directed evolution ... [https://www.sciencedirect.com/science/article/abs/pii/S2405471225002200]
• 13. Software Overview [https://2025.igem.wiki/tsukuba/software-overview]
• 14. Breaking the Bottleneck in AI-Driven Protein Engineering [https://www.ailurus.bio/post/breaking-the-bottleneck-in-ai-driven-protein-engineering]
• 15. combining machine learning and rapid cell-free testing [https://www.nature.com/articles/s41467-025-65281-2]
• 16. Abstraction hierarchy to define biofoundry workflows and ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12214803/]
• 17. Integrating protein language models and automatic ... [https://www.nature.com/articles/s41467-025-56751-8]
• 18. Self-driving laboratories to autonomously navigate the ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10926838/]
• 19. Frontiers in biofoundry: opportunities and challenges [https://www.frontiersin.org/journals/synthetic-biology/articles/10.3389/fsybi.2025.1630026/full]
• 20. Engineering biological systems using automated biofoundries [https://pmc.ncbi.nlm.nih.gov/articles/PMC5544601/]
• 21. Modular plasmid design for autonomous multi-protein ... [https://jbioleng.biomedcentral.com/articles/10.1186/s13036-025-00483-2]
• 22. Automated and Programmable Cell-Free Systems for Scalable ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12463565/]
• 23. Critical Assessment of Protein Engineering (CAPE) [https://pmc.ncbi.nlm.nih.gov/articles/PMC11574941/]
• 24. Enhancing functional proteins through multimodal inverse ... [https://www.nature.com/articles/s41467-025-65175-3]
• 25. Optimization of protein thermostability and exploitation of ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7347439/]
• 26. Quality control of protein reagents for the improvement ... [https://www.nature.com/articles/s41467-021-23167-z]
• 27. Protein Characterization and Quality Control [https://selvita.com/drug-discovery/small-molecules/expertise-areas/protein-sciences/protein-characterization-and-quality-control]
• 28. A multi-library approach to parallelised sequence space ... [https://www.sciencedirect.com/science/article/abs/pii/S0010482525015604]
• 29. Integrating protein language models and automatic biofoundry ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11814318/]
• 30. Tools for computational design and high-throughput ... [https://www.sciencedirect.com/science/article/abs/pii/S0169409X22000333]
• 31. Connecting chemical and protein sequence space to ... [https://www.nature.com/articles/s41586-025-09519-5]
• 32. Bayesian Optimization in Bioprocess Engineering—Where Do ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12067035/]
• 33. Safe model based optimization balancing exploration and ... [https://www.nature.com/articles/s41598-025-12568-5]
• 34. A Guide to Bayesian Optimization in Bioprocess Engineering [https://arxiv.org/html/2508.10642v1]
• 35. Drift on holey landscapes as a dominant evolutionary ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10756301/]
• 36. Graphical Multioutput Gaussian Process with Attention [https://openreview.net/forum?id=6N8TW504aa]
• 37. What are the applications of Multi-outputs Gaussian processes ... [https://discourse.pymc.io/t/what-are-the-applications-of-multi-outputs-gaussian-processes-mogps-in-your-work/9571]
• 38. Development of the autonomous lab system to support ... [https://www.nature.com/articles/s41598-025-89069-y]
• 39. Protein engineering in the deep learning era - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC11685842/]
• 40. Short-loop engineering strategy for enhancing enzyme ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11982487/]
• 41. Tailoring industrial enzymes for thermostability and activity ... [https://www.nature.com/articles/s41467-025-55944-5]
• 42. Enhanced enzyme activity and stability through immobilization ... [https://jbioleng.biomedcentral.com/articles/10.1186/s13036-025-00546-4]
• 43. A primer to directed evolution: current methodologies and ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10074555/]
• 44. Insights Into Protein Engineering: Methods and Applications [https://www.the-scientist.com/insights-into-protein-engineering-methods-and-applications-72286]
• 45. Rational Design vs Directed Evolution [https://synapse.patsnap.com/article/protein-engineering-showdown-rational-design-vs-directed-evolution]
• 46. An iterative deep learning-guided algorithm for directed ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12415074/]
• 47. Comparison of five protein engineering strategies to ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5736438/]
• 48. Automated Lab Accelerates Programmable Protein Evolution [https://bioengineer.org/automated-lab-accelerates-programmable-protein-evolution/]
• 49. Advances in AI for Protein Structure Prediction [https://pmc.ncbi.nlm.nih.gov/articles/PMC10968151/]
• 50. The impact of Library Size and Scale of Testing on Virtual ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11257449/]