#17 · ML / AI methods

External knowledge and uncertainty in enzyme and drug-target prediction models

The common thread is augmenting learned biological representations with external knowledge or supervision while testing whether those gains hold under stringent evaluation. RAGCell uses LLM-built cell- and feature-level knowledge databases as supervision and aligns cell embeddings with text for single-cell analysis. Katalyst aligns enzyme sequences with protein text, substrates and products via contrastive learning for kcat prediction, and ProDTI combines biochemical similarity priors with positive-unlabeled learning for drug-target interaction. The LLM-in-the-loop reinforcement learning framework applies LLM advice and hybrid rewards to feature selection on high-dimensional bioinformatics data. Diagnostics temper these claims: class-incremental methods do not transfer reliably across five frozen protein encoders and two enzyme function datasets, similarity-driven leakage inflates enzyme-substrate models that fail to generalize across ligand chemistry, and RCP-DTA supplies adaptive, sample-specific conformal intervals for drug-target affinity.