Adding assay context to GPCR bioactivity models improves accuracy and calibration, though gains shrink on unseen assays
- The authors extend proteochemometric models with assay descriptors, encoded either as categorical identifiers or as text embeddings of assay descriptions, so that experimental context is modelled alongside compound and protein features.
- They benchmarked on a curated ChEMBL-derived GPCR set under random, scaffold and assay-held-out splits, judging both point accuracy and probabilistic quality, retrospectively.
- Split choice drove difficulty most. Categorical identifiers helped consistently when assays recurred; on held-out assays the representations converged, with text embeddings holding a small, steady edge.
💡 Points Based on the abstract
Categorical assay IDs paying off when assays recur is close to leakage by design, so I discount that half of the result. What matters for a new campaign is the held-out-assay split, where the text-embedding edge is small and hardly justifies retooling. Whether any of it survives a prospective campaign is untested.