The abstract is read at the publisher; the summary is JClub's.
BioLM-Score improves DEKOIS 2.0 screening endpoints over a matched geometry-only baseline, but affinity scoring and ranking do not improve uniformly under the same controlled comparisons.
- Why it matters: Protein–ligand scoring must balance affinity prediction, pose selection, and virtual screening, yet geometric likelihood models can generalize poorly when structural coverage is limited.
- What they did: The method combines frozen protein and molecular language-model embeddings with a residue–atom distance-likelihood framework using gated graph convolutions, comparing geometry-only, jointly trained, and two-stage fine-tuned variants.
- The result: Under a time-split protocol, scoring and ranking are lower than the geometric baseline while screening gains persist; on BDB2020+, affinity correlation versus GenScore narrows after LP-PDBBind-compatible retraining, and shuffling embeddings reduces the screening advantage.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.