Fig. 4: ML-guided directed evolution predicts highly active mutants with a lower screening burden than iterative site saturation mutagenesis.
From: Accelerated enzyme engineering by machine-learning guided cell-free expression
A, B Analysis of model fidelity with training sets built with smaller libraries than saturation mutagenesis, including reduced codon sets (NDT, NRT) and reduced amino acid alphabets based on BLOSUM50, quantified by spearman correlation (⍴) and NDCG. Comparing measured versus predicted activity on withheld ISM rounds is shown for models trained on the complete saturation mutagenesis dataset for both moclobemide (moc) (A), and metoclopramide (meto) (B). The experimentally validated percent conversion (n = 3; error bars indicate ± SD) of ML-predictions for moclobemide (C), metoclopramide (D), and cinchocaine (E) with the quadruple mutant from ISM (M4) colored gray. For cinchocaine, the ML model predictions did not include the highest performing mutant from ISM. Wild type McbA is colored dark gray. Percent conversion was measured by RP-HPLC in independent experiments. Source data are provided as a Source Data file.
