Lab-in-the-Loop Enzyme Engineering

Traditional protein engineering often follows a relatively linear process:
Design → Build → Test → Identify Hits → Start Again

A lab-in-the-loop approach creates a continuously learning system:
AI Design → Build → Test → Experimental Data → Model Update → Improved Design → Build → Test

Every experimental round provides information that changes what the model predicts and which variants should be tested next.
This allows the engineering campaign to become increasingly targeted as more experimental data becomes available.

AI models are powerful at exploring sequence space.
Laboratory experiments provide the ground truth.
Neoncorte Bio brings these two capabilities together.

AI ProvidesLarge-scale sequence exploration

• Mutation effect prediction
• Protein fitness prediction
• Variant prioritization
• Sequence-function modeling
• Multi-objective optimization
• Prediction of mutation combinations
• Experimental selection strategies

The Laboratory ProvidesExperimental measurements

• Real-world phenotype data
• Validation of computational predictions
• Information about unexpected sequence behavior
• New training data for subsequent model iterations

The combination is more powerful than either computational prediction or experimental screening alone.

1. Define the Engineering Objective
We begin with the desired protein phenotype.
Depending on the project, this could include:

• Higher catalytic activity
• Improved thermostability
• Increased solvent tolerance
• Improved substrate specificity
• Higher enantioselectivity
• Better expression
• Increased solubility
• Reduced aggregation
• Improved operational stability
• Altered cofactor specificity
• Higher space-time yield

Multiple objectives can be optimized simultaneously.

2. Analyze the Starting Sequence
We analyze the available protein sequence and project-specific data.
Depending on the project, this may include:

• Protein sequence
• Mutational data
• Activity measurements
• Stability measurements
• Expression data
• Structural information
• Homologous sequences
• Previous directed-evolution results
• Deep mutational scanning data

The objective is to extract as much information as possible before designing new variants.

3.Generate Computational Predictions
AI and machine-learning models explore the relevant sequence space.
Neoncorte Bio can evaluate:

• Single mutations
• Combinatorial mutations
• Higher-order variants
• Sequence-function relationships
• Fitness landscapes
• Mutation interactions
• Potentially beneficial regions of sequence space

Rather than testing every possible variant, the system prioritizes candidates with the strongest predicted potential.

4.Select Variants for Experimental Testing
The best computational candidates are not necessarily the only variants worth testing.
A lab-in-the-loop strategy can deliberately select variants that provide information as well as performance.
This can include:

• High-predicted-performance variants
• Diverse sequence variants
• Uncertain predictions
• Mutations testing specific hypotheses
• Variants designed to resolve model uncertainty

This helps the experimental campaign simultaneously optimize the protein and improve the model.

5. Build & Test in the Laboratory
Selected sequences are synthesized, expressed, and experimentally characterized.
Depending on the project, measurements may include:

• Enzyme activity
• Binding affinity
• Stability
• Melting temperature (Tm)
• Solubility
• Expression
• Selectivity
• Kinetic parameters
• Operational lifetime
• Process performance

The experimental results become the next layer of project-specific data.

6. Feed Experimental Results Back into the Model
This is the defining feature of the lab-in-the-loop methodology.
Experimental measurements are incorporated into the computational workflow.
The model can then learn:
Which mutations actually work?
Which predictions were wrong?
Which sequence regions are promising?
Which mutations interact positively or negatively?
Where is the model uncertain?
The next round of predictions is therefore based on a larger and more relevant dataset than the previous round.

7. Design the Next Round
The updated model identifies the next set of variants.
As the experimental dataset grows, the optimization process can become increasingly project-specific.
The cycle continues:
Predict → Test → Learn → Predict → Test → Learn
until the desired engineering target is reached.

Traditional experimental screening can become expensive because the search space grows exponentially with the number of mutations.
For a protein with hundreds of amino-acid positions, the number of possible variants quickly becomes enormous.
A laboratory cannot test all of them.
AI cannot perfectly predict all of them either.

Lab-in-the-loop engineering addresses the problem by combining computational scale with experimental truth.
AI explores the sequence space.
The laboratory provides reality.
The feedback loop continuously connects the two.

Real protein engineering projects rarely have only one objective.
An enzyme may need:
Higher activity + higher stability + better expression + greater selectivity
But improving one property can negatively affect another.
For example:

• Increasing activity can reduce stability.
• Increasing expression can increase aggregation.
• Improving substrate promiscuity can reduce selectivity.
• Stabilizing a protein can reduce catalytic efficiency.

Neoncorte Bio can use lab-in-the-loop optimization to learn these trade-offs directly from experimental data.
The result is not simply optimization of one measured property.
It is optimization of the protein toward the complete target phenotype.

Our approach integrates:

• AI-driven protein design
• Machine learning
• Protein language models
• Computational mutagenesis
• Virtual Deep Mutational Scanning
• Protein fitness landscape prediction
• Epistasis prediction
• Higher-order mutation prediction
• Active learning
• Bayesian optimization
• Multi-objective optimization
• Experimental validation
• Design-Build-Test-Learn cycles

The laboratory is not the final step in the workflow. It is part of the learning system.

• Enzyme engineering companies
• Protein engineering companies
• Industrial biotechnology companies
• Biocatalysis teams
• Enzyme manufacturers
• Synthetic biology companies
• Pharmaceutical companies
• Biotech startups
• CROs and CDMOs
• Academic research groups

It is especially valuable when a team already has experimental capabilities or access to wet-lab validation, but wants to make those experiments more targeted and data-efficient.