Deep Learning Models for Protein Function Prediction
EnzyKR and DeepLasso illustrate how task-specific neural architectures, grounded in physical constraints (activation barriers, topology) and trained on structure- or library-scale data, can deliver actionable predictions for enzyme selectivity and RiPP bioactivity in real design workflows.
EnzyKR. EnzyKR is a chirality-aware deep learning framework for stereoselective biocatalysis. It uses a classifier–regressor pipeline that first identifies the reactive binding conformer of a substrate–hydrolase complex and then predicts its activation free energy, with a structure-based encoding that captures chiral interactions. Trained on 204 docked substrate–enzyme complexes and tested on 20 held-out complexes, EnzyKR achieved R = 0.72 and MAE = 1.54 kcal·mol⁻¹ for barrier prediction; across 28 hydrolase-catalyzed kinetic-resolution reactions, it correctly predicted the favored enantiomer and outperformed DLKcat in 18/28 cases (64%). These results position EnzyKR as a practical model for selecting hydrolases that deliver high enantioselectivity in kinetic resolution.
DeepLasso. DeepLasso leverages massive sequence–activity data to predict the RNAP-inhibitory activity of lasso-peptide variants. Built from a >90,000-member ubonodin library, it combines a sequence encoder (CNN + BiLSTM + attention) with a topology encoder that tags ring, loop, and tail regions, feeding a classifier–regressor architecture to handle dropout vs. measurable variants and to score enrichment values. In 5-fold cross-validation, the classifier achieved 73%/63% hit rates on dropout/nondropout identification, and the regressor reached Pearson R = 0.80 (Spearman R = 0.77) on held-out variants. The screen also uncovered variants with sub-micromolar MICs against Burkholderia cenocepacia, underscoring the model’s utility for prioritizing potent designs across large mutational spaces.
Both works were led by Xinchun Ran.
The software code:
https://github.com/ChemBioHTP/LassoPred
Publications:
https://pubs.rsc.org/en/content/articlelanding/2023/sc/d3sc02752j
https://pubs.acs.org/doi/full/10.1021/acscentsci.2c01487