Skip to content

Machine-Learning-Guided Cell-Free Enzyme Engineering

Summary

Landwehr et al. (2025) built a design-build-test-learn (DBTL) platform that pairs cell-free DNA assembly and cell-free gene expression (avoiding cloning and cell culture) with augmented ridge-regression machine learning. It was used to rapidly map sequence-function landscapes and specialize the promiscuous amide synthetase McbA toward nine distinct small-molecule pharmaceutical targets. Across 1,217 evaluated enzyme variants and 10,953 unique reactions, ML-predicted variants achieved 1.6- to 42-fold improved activity over the parent enzyme, with some campaigns completed in about one week.[1]

Method

The workflow is explicitly staged: (1) identify non-native reactions a generalist enzyme (McbA) can already catalyze promiscuously and prioritize pharmaceutically valuable ones; (2) run a site-saturation "hot-spot scan" (here, 64 residues) to identify positions that materially affect activity for a given target reaction; (3) train an augmented ridge-regression model — combining a site-specific amino acid encoding with a zero-shot variant-effect prediction (e.g. from EVmutation or ESM-1b) — on the resulting single-mutant data; (4) use the model to rank higher-order combination mutants without needing to test them all experimentally.[1]

Model quality was evaluated with normalized discounted cumulative gain (NDCG), chosen because the practical goal is correctly ranking high-fitness variants rather than predicting exact activity values; augmented models (single-mutant assay data + zero-shot prediction) outperformed ridge regression on assay data alone. Using the complete 77-variant site-saturation training set outperformed common reduced-library strategies (NDT/NRT codon sets, alanine/glycine/proline/cysteine scanning, BLOSUM-grouped encodings), because those reduced sets under-sample the many near-zero-activity variants needed to avoid "holes" in the training data.[1]

Results

  • Across all nine engineered McbA variants (spanning metoclopramide, moclobemide, cinchocaine, procainamide, declopramide and others), activity improved 1.6- to 42-fold over wild-type McbA; the moclobemide variant reached 96% conversion (a 42-fold increase in catalytic efficiency) and was scaled to milligram quantities.[1]
  • In each of the nine campaigns, the ML-predicted quadruple mutant outperformed a naive combination of the four best individual single-mutants — evidence the augmented model captures genuine (non-additive) epistasis rather than just ranking single mutations.[1]
  • Across the full study, 2,856 McbA variants were characterized (1,217 of which fed the ML models), spanning 1,100 possible amide products and 12,584 substrate-pair/mutant reactions; 19 unique residue positions were found to significantly affect biocatalysis, with a distinct hot-spot residue set per target reaction — a given mutation (e.g. V177S) that helped for one product did not generalize across chemically similar substrates, indicating that hot spots are reaction-specific rather than substrate-fragment-specific.[1]
  • Six of the nine engineering campaigns were run simultaneously, each completed within about one week, exploiting the low per-reaction cost of cell-free expression (cents per 10-µL reaction).[1]

Synthesis: This case study illustrates the "cheap-first, expensive-only-on-survivors" funnel argued for in Predicting Catalytic Competence of Enzyme-Ligand Complexes — but applied entirely at the assay/ML layer, without structural or QM/MM steps. It is a useful counterexample showing that for enzymes with tractable expression and assay throughput, an empirical DBTL loop can out-compete a purely computational pipeline in wall-clock time, even though it provides no mechanistic explanation for why a given hot-spot residue matters.

See Also

Citations

[1] Landwehr, G.M., Bogart, J.W., Magalhaes, C., Hammarlund, E.G., Karim, A.S., Jewett, M.C. (2025). Accelerated enzyme engineering by machine-learning guided cell-free expression. Nature Communications, 16, 767. Supports: all method and results claims above. Location: Abstract; Results (model performance evaluation; McbA engineering campaigns); Discussion.