Traditional models of chemistry struggle to account for enzyme-catalysed reactions; VenusRX learns about reactants, products and catalytic function all at once.
Reaction prediction models learn mostly from transformations of molecules, whereas an enzymatic reaction depends on both the structure of the molecules and the function of the enzyme that catalyzes it. A team of nine researchers on arXiv proposes to treat the problem as a common space linking reactants, products and EC numbers (the International Classification of Enzymes). It publishes VenusRX-Bench, a test bed that merges several biochemical bases with standardized curation and breakdowns that avoid leakage between training and test, on three tasks: direct prediction, one-step retrosynthesis and EC number prediction.
The tests reveal a clear gap between chemistry and enzymology, related to the lack of data, the role of the catalytic context and the size of the biomolecules. To bridge this gap, the authors trained VenusRX, a T5-type sequence-to-sequence model, first on millions of reactions generated by templates, then on real biochemical reactions. It obtains the best or comparable results on most tasks; providing the EC number systematically improves predictions, and the learned representations in turn make it possible to guess this number.
Read the full article on the original source
Go to the source →