2026-08-11
A Caltech team uses active learning (a multi-task ensemble with expected hypervolume improvement) to guide protoglobin directed evolution; just 655 sequences yield a better variant for every parent reaction and catalysis of 5 reactions the parent cannot do.
Directed evolution is the workhorse method for engineering enzymes, and the Arnold lab won the 2018 Nobel Prize in Chemistry for it. Its pain point is the starting point: you need a protein with detectable activity for the target reaction before you can optimize anything. Finding that starting point means screening huge variant libraries, with a low hit rate.
The Arnold group had earlier built ALDE (active learning-assisted directed evolution): train a machine-learning model, update it as experiments come in, and let it propose which mutations to test next. This paper extends it to multi-reaction optimization, calling it mr-ALDE (multi-reaction ALDE). A single model ingests function data across many reactions and predicts mutations that are more active toward both carbene and nitrene transfer, aiming for a "broadly functional" enzyme library to seed later screening.
The starting protein is PromPgb, a thermostable protoglobin that already catalyzes both carbene and nitrene transfer. That fact is the precondition for everything that follows.
Training data comes from screening active-site double mutants (152, that is 8×19) and triple mutants (264) against 8 Initial Training Reactions (ITRs), yielding over 4,000 sequence-to-function pairs and more than 500 multi-mutation variants.
The model is an ensemble of multi-task supervised learners that takes a sequence encoding and predicts two objectives at once: mean carbene activity and mean nitrene activity. To pick the next variants to test it uses the expected hypervolume improvement (EHVI) acquisition function, which trades off "high predicted score" against "high uncertainty" and pushes along the multi-objective Pareto front.
Two rounds are run. Round 1 proposes 63 variants, tested against 14 new reactions plus the ITRs, 22 reactions in total (13 carbene, 9 nitrene). Those results feed back into the model, and Round 2 proposes 64 more, tested against those 22 plus 4 more, 26 reactions in total. Across the two rounds, 127 model-predicted variants were tested; counting the training library, 655 sequences in total.
For every reaction where PromPgb produced detectable product, at least one model-predicted variant was more active than the parent. Crucially, the gains also showed up on reactions mechanistically distinct from the training set, evidence that the sequence features the model learned do generalize.
The clearest case is reaction N6 (intramolecular C(sp³)−H amination of an alkyl azide): the parent gives only trace product, while variant FYIFMMFQ reaches 3% yield. This variant appeared in Round 1, when the model had seen only ITR data, even though N6 requires both azide activation and an intramolecular cyclization, steps absent from the training set.
The headline numbers show how much the model mattered. Only about 33% of training-library variants beat parent yield on any ITR, whereas nearly 70% of the predicted-library variants beat parent on at least one reaction. Other specifics: on reaction C13, variant VVFFMAFN gives a sixfold improvement over PromPgb; for trans-5, a cyclopropane precursor of the antithrombotic drug ticagrelor, variant VYIAIVFQ delivers a 9:1 preference for the trans product; variant TAIFMVFQ inverts selectivity toward the Z product on reaction C8. All of this came from testing 655 sequences.
Of the 10 reactions the parent cannot catalyze, 5 were catalyzed by variants in these libraries. The tradeoff: on the "native" reaction N1, the best mutant reached only one fifth of PromPgb's yield, showing a balance between broad promiscuous activity and native function.
The expensive part of directed evolution is wet-lab screening. A model that picks worthwhile mutations shrinks that burden. One model-guided campaign here tested only 655 sequences and still produced a broadly functional catalyst set, a real saving for medicinal-chemistry and fine-chemical groups looking for enzyme starting points.
It is also a validated case of active learning doing protein engineering better than random screening, useful reference for anyone in MLDE (machine-learning-directed evolution) and self-driving labs.
The authors are explicit. The study covers a single enzyme class (protoglobin), and the starting point was "particularly well positioned": thermostable, already catalyzing both carbene and nitrene transfer. Without such a parent, mr-ALDE is hard to run. The lab also had deep prior knowledge of the ApePgb family, which let it pick active-site residues that influence non-native reactivity and carry epistatic effects. Whether the approach transfers to less-studied systems is unverified, and the authors say it is still unclear how to optimally design training reaction sets, libraries, and model parameters.
The caveat after reading: this is a carefully chosen scaffold plus heavy biochemical prior. Machine learning contributed, but experimental design and domain knowledge matter just as much, and whether the gains transfer to harder, less-characterized enzymes stays open. The yields on the "new" reactions are also modest (3% for N6); these are detectable activities, not production-ready catalysts.