A simple linear alignment of pretrained embeddings offers interpretable kinase drug-target prediction without heavy deep architectures
- A linear model projects ChemBERTa molecule embeddings and ESM-2 protein embeddings into one shared space, where predicted affinity is simply the cosine similarity between drug and target.
- It was pretrained on a large curated set of interactions and evaluated on the KIBA and Davis kinase benchmarks, where the authors report it surpassing existing methods.
- The learned space reportedly recovers kinome phylogeny without supervision and retrieves known targets of approved drugs, though validation appears retrospective with no new experiments.
💡 Points Based on the abstract
Two frozen embeddings and a cosine is a refreshingly simple baseline, which is exactly why the split matters. Random splits let near-identical kinase inhibitors straddle train and test, and the pretraining corpus may overlap the benchmark pairs. Before believing the margin over deep models, I'd score it on my own scaffolds.