Fig. 6. Using case studies that were in fact experimentally implemented prior to our model, we demonstrate our model understands that directed evolution of enzymes (1) optimizes their catalytic activity towards new substrates,2,40,41 (2) alters their cofactor dependence,42 (3) inverts their enantioselectivity,43 and (4) makes them catalyze new chemical reactions.44 Using generated negative examples, we illustrate types of reaction proposals that the model is likely to avoid. In (5A) and (5B), desired vs. proposed transformations significantly overlap in reaction chemistry, but they accommodate drastically different substrates. Both transformations in (5C) have similar substrates (i.e. acyl-CoA), but catalyze different reactions. In (5D), neither substrate nor reaction chemistry overlap between the native- and desired-transformations. The model intentionally discourages making drastic changes to the native transformation to yield the desired transformation, sometimes resulting in false negatives similar to (1D).1.
