Model predicts how fast an enzyme works on a substrate, with uncertainty attached

The work

Technical · from the record

IECata predicts enzyme catalytic efficiency (kcat/Km) directly from an enzyme amino acid sequence and a substrate SMILES string, and returns an uncertainty estimate alongside each prediction. The authors assembled 11 815 kcat/Km entries from BRENDA and SABIO-RK for training and an out-of-domain set of 806 entries curated from the literature. Reported Pearson correlation was 0.778 on the in-domain test split and 0.663 on the out-of-domain set, against 0.652 for UniKP retrained on the same training data and 0.463 for the released UniKP model; under five-fold cross-validation on the whole dataset IECata reported higher R and PCC and lower MAE and RMSE than UniKP. Bilinear attention weights were compared with crystallographic binding-site residues across 45 sesquiterpene synthase structures, giving a mean overlap of 1.911 among the top 25 attention residues against 1.454 for randomly sampled pockets.

How AI was used

Entries pairing an enzyme sequence, a substrate SMILES string and a measured kcat/Km value were pulled from BRENDA and SABIO-RK, deduplicated against PubChem standard SMILES, filtered by unit, sequence length and SMILES length, log10 transformed and split 8:1:1. Each enzyme sequence was embedded per residue by the frozen ProtT5 protein language model, whose fixed parameters were used without further training, and the embedding was passed through a light attention module and a multilayer perceptron; each substrate was built as a molecular graph with atom-type, degree, hydrogen-count and chirality features and encoded by a graph convolutional network. A bilinear attention network formed a pairwise enzyme-substrate interaction map and a bilinear pooling layer produced a joint representation, which fed an evidential output layer emitting the four parameters of a normal-inverse-gamma distribution, from which the predicted value and its epistemic, aleatoric and total uncertainty were derived. The model was trained with a negative log-likelihood term plus an evidential regulariser whose coefficient was chosen by calibration, using both the 8:1:1 split and five-fold cross-validation. Ablations swapped the enzyme encoder for integer-plus-CNN and ProtT5-plus-CNN and the joint representation for one-side attention or linear concatenation, and UniKP was retrained on the same training data as a baseline. Attention weights were then read out over residues and substrate atoms for cocrystal structures and for a curated set of sesquiterpene synthase structures, and predictions were ranked by predicted value plus uncertainty over a mutant dataset to compute hit ratios.