Rank-guided learning plus robotics reprogram two enzymes 57-fold and 104-fold in five cycles

2026-08-05

REAP couples a ranking-plus-regression loss on protein-language-model features with a robotic closed loop, lifting P450 BM3 activity 57-fold and Sortase A 104-fold in five cycles.

What problem this solves

Directed evolution (iterating mutation and screening until an enzyme does what you want) has two chronic problems. First, the fitness landscape (the high-dimensional surface mapping each variant to its activity) is rugged and sparse: a round screens thousands of mostly-dead variants, and the mutations that actually help often sit far from the catalytic center, so active-site-focused search misses them. Second, pure wet-lab work is slow (a cycle takes weeks), while predictive models trained on the few measurements you have struggle under standard regression losses, which care about absolute values when, with sparse data, getting the ranking right matters more.

Method

REAP bolts three things into one loop: a protein language model for features, a hybrid loss called RankReg, and a fully automated wet-lab robot.

The model:

• ESM2 (a protein language model) makes zero-shot predictions to seed the initial training set. For the P450 BM3 case it picks 32 high-priority variants.
• PLM-RankReg is the core. Its loss has two parts: a pointwise regression term predicts absolute activity, and a pairwise ranking term enforces which variant beats which. The ranking term handles three cases: a hinge penalty on positive pairs forces the predicted gap past an adaptive margin, a symmetric hinge handles negative pairs, and a quadratic penalty keeps near-equal activities consistent. In sparse-data regimes the model learns the order first, then converges on accurate values.
• Inference uses an ensemble of 100 neural networks, reporting a mean μ and a standard deviation σ. Bayesian active learning picks variants that are either high-activity (exploit) or high-uncertainty (explore), spending the screening budget where the model knows least.

The robot is a four-stage pipeline: library construction (3 days), protein expression (2 days), lysis and reaction (1 day), functional assay (1 hour), all on a Tecan Fluent with no human in the loop. P450 BM3 products are quantified by Echo-MS; Sortase A by fluorescence dequenching. Pairwise Pearson correlations across 12 replicate runs exceed 0.98, and crude-lysate activity correlates with purified-protein activity at R≈0.96, so the high-throughput numbers are trustworthy.

Each round retrains the model, and the next round's suggestions climb from single mutations to combinatorial ones.

Results

P450 BM3 (a cytochrome P450 engineered here to hydroxylate (−)-deoxypodophyllotoxin, a precursor to the anticancer lignan podophyllotoxin):

• Starting from the FL#62 variant, REAP ran 5 cycles. Rounds 1–3 screened single mutants (93 in round 1), round 4 double mutants, round 5 quadruple mutants.
• The model nominated 27 sites; with 13 additional functionally relevant sites that is 40 in total. Network analysis pinned key mutations A330E, S81M, T180L, E207L, R498V, and quadruple variants carrying them ranked in the top three of round 5.
• Desired activity rose 57-fold over five cycles.

Sortase A (a Staphylococcus aureus transpeptidase used for protein conjugation):

• Same loop, up to 104-fold enhancement in desired activity.
• One telling find: D170Q is weaker than D185V and D185T as a single mutation, yet stronger in a quadruple background. This context-dependent epistasis (combined effect ≠ sum of single effects) is exactly what point-by-point screening would throw away, and combinatorial sampling caught it.

Model benchmarks (five-fold cross-validation, Spearman correlation for ranking):

It leads across the board on ranking. On extrapolation to combinatorial space, the edge over HuberLoss is not significant.

Why it matters

Ranking learning, caring about relative order rather than absolute value, is a mature idea in recommender systems, and it lands well on the sparse-data reality of directed evolution. Paired with automation, REAP compresses an engineering cycle to weeks and makes the model stronger each round. For anyone in synthetic biology or green chemistry (replacing metal catalysts with enzymes to make drug molecules), it is a working example of AI wired into the experiment itself rather than stopping at prediction.

Limitations

• Validated on only two enzymes, both relatively tractable systems. Generalization to enzymes with narrower substrate specificity or more rugged landscapes is unproven.
• The loop depends on a full automated wet-lab stack (robotic arms, Echo-MS) that most labs cannot reproduce, so generalizable currently glosses over a steep equipment barrier.
• The 57-fold and 104-fold figures are desired activity from high-throughput crude screens, not purified kinetic constants. The authors add kinetic characterization for representative variants, but the headline multiples are screening-scale.
• The initial training set relies on ESM2 zero-shot picks and needs a decent starting scaffold (FL#62) and a reliable protein language model; a fully cold start was not tested.
• On single-to-combinatorial extrapolation the margin over HuberLoss is not significant, so this line of improvement narrows against a more robust regression loss.

Terms

• Directed evolution: iterative mutation and screening that drives an enzyme toward a desired activity; the 2018 Chemistry Nobel direction.
• Fitness landscape: the high-dimensional surface mapping each variant to its activity; rugged means small mutations cause large swings.
• Protein language model (PLM, ESM2): self-supervised model trained on massive sequence databases; its outputs serve as features for function prediction and can zero-score variants.
• RankReg: the paper's hybrid loss, with a regression term for absolute values and a ranking term for relative order.
• Bayesian active learning: uses ensemble uncertainty (σ) to choose which experiment to run next, balancing exploit and explore.
• Epistasis: a mutation's effect depends on what else is mutated; the combined effect is not the sum of single effects.

Source

• Paper

What people are saying

• Major AI Breakthrough in Synthetic Biology and Green Chemistry ProxyLumina
• Nature Paper: AI and Robotic Platform Accelerates Automated Enzyme Engineering KevinKaichuang

All paper explainers