#34 · ML / AI methods

Stereochemistry-aware drug representations for affinity and side-effect prediction

The dominant thread is encoding molecular structure beyond connectivity and coupling it to a second modality. Two preprints use directed-bond graph neural network drug representations with explicit stereochemical or chirality awareness: one predicts drug-target affinity with sequence-based protein embeddings, the other predicts drug side-effect frequencies with PubMedBERT side-effect embeddings and reports matching state-of-the-art performance. The third item, MolCodon, takes a different route, encoding small molecules as deterministic three-character codon-based tokens and reporting 98.93% InChIKey round-trip accuracy for interpretable structural representation and similarity search. Its text does not describe stereochemical handling, so the shared thread is dominant rather than universal. What is new is explicit chirality handling in drug models combined with cross-modal fusion.