Researchers at the Arc Institute have developed a new machine learning framework that dramatically reduces the experimental burden of protein engineering, a process that has long been constrained by the sheer number of possible molecular combinations to test. The new approach, called MULTI-evolve, applies neural networks trained on datasets of roughly 200 variants to predict which combinations of mutations will enhance protein function most effectively. The work, published in Science, represents a fundamental shift in how computational prediction and laboratory testing can be integrated from the start of a protein engineering campaign.
Protein engineering typically requires two steps: identifying individual mutations that boost function, then combining them in ways that amplify their effects. The challenge lies in the astronomical search space. A protein of just 100 amino acids has 20^100 possible variants, more combinations than atoms in the observable universe. Traditional methods test hundreds of variants but explore only narrow regions of that space. Computational screening can broaden the search, yet most approaches still demand tens of thousands of measurements or 5 to 10 iterative rounds of testing and refinement.
The bottleneck, Arc Institute researchers realized, was not computational prediction alone. With newer protein language models now able to scan vast theoretical spaces, the real constraint shifted back to the laboratory: researchers could efficiently build and test only hundreds of variants in a single protein engineering campaign. The critical question became how to choose those hundreds most strategically to uncover substantially improved protein variants.
Why Quantity Alone Does Not Solve the Prediction Problem
Early attempts to train neural networks on single-mutant data failed to predict which multi-mutant combinations would work reliably. The reason was fundamental: models trained only on individual mutation effects lack information about how mutations interact with each other. Most large datasets of random variants are not useful either, because the vast majority of mutations do not enhance function. Testing thousands of random variants teaches models mostly about what does not work.
The Arc Institute team reframed the problem around quality over quantity. Instead of generating thousands of random variants, researchers first identified approximately 15 to 20 function-enhancing mutations using protein language models or direct experimental screening. Then they systematically tested all pairwise combinations of those beneficial mutations, generating roughly 100 to 200 measurements.
The critical insight was that every single measurement proved informative. Pairwise combinations reveal epistasis, the way mutations interact. A double mutant might perform better than the sum of its parts (synergy), worse than expected (antagonism), or exactly as predicted (additivity). These interaction patterns teach neural networks the rules governing how mutations combine.
Validation Across Diverse Protein Families
The team validated the approach computationally using 12 existing protein datasets from published studies. Neural networks trained only on single and double mutants could accurately predict complex multi-mutants containing 3 to 12 mutations across all 12 diverse protein families. The result held even when researchers reduced training data to just 10 percent of what was available.
This efficiency opens a practical path forward for laboratory work. Rather than running hundreds of rounds of iterative testing or screening thousands of random variants, protein engineers can now focus experimental effort on the most strategically informative combinations. The framework tightens the coupling between computational design and wet-lab validation from the project’s start, avoiding the waste of testing mutations that lack complementary interactions.
Integration of Machine Learning and Experimental Design
MULTI-evolve represents Arc Institute’s first lab-in-the-loop framework for biological design, reflecting the organization’s broader investment in AI-guided research. The approach does not replace experimental work; it focuses experimental resources where they matter most. The Science publication describes the framework as tightly integrating computational prediction and experimental design from the outset, moving away from a sequential model where computation happens first and experiments validate afterward.
The method assumes that identifying high-quality candidate mutations is achievable through existing tools or targeted screening. The real challenge, and the one MULTI-evolve addresses, is predicting which combinations of those candidates will work synergistically. By focusing neural network training on the informative pairwise space, researchers capture the epistatic relationships that govern multi-mutant behavior without needing massive datasets or extended iteration cycles.
The implications extend beyond academic protein engineering. Industrial protein design, enzyme optimization, and directed evolution efforts all face the same constraint: the laboratory is the bottleneck, not the computational models. A framework that reduces the number of variants researchers must test while improving prediction accuracy for the ones they do test could accelerate development timelines and lower costs across biotechnology and synthetic biology applications.
The work demonstrates that machine learning frameworks designed specifically for the problem at hand, not generic models applied broadly, deliver the most practical gains. By aligning model training to the structure of the problem (epistatic interactions between beneficial mutations), the team converted a data-hungry approach into one that learns effectively from focused, high-quality measurements. For protein engineering, that reorientation may prove as valuable as the computational method itself.
