BioLM-Score: Language-Prior Conditioned Probabilistic Geometric Potentials for Protein-Ligand Scoring

Abstract Motivation Protein-ligand scoring must address affinity prediction, pose selection, and virtual screening, but these objectives are not necessarily optimized together. Geometric likelihood models may generalize poorly when training-set structural coverage is limited. We test whether pretrained biomolecular representations improve scoring under a controlled geometric baseline. Results BioLM-Score integrates frozen protein and molecular language-model embeddings into a residue-atom distance-likelihood framework via gated graph convolutions. We evaluate geometry-only, jointly trained, and two-stage fine-tuned variants, using the jointly trained variant for the primary controlled comparisons. All three DEKOIS 2.0 screening endpoints improve over the matched geometric baseline with target-level paired support, but affinity scoring and ranking do not improve uniformly. Under a time-split protocol, scoring is significantly lower and ranking is also lower than the geometric baseline, while screening gains persist. On BDB2020+, the difference in affinity correlation between BioLM-Score and GenScore narrows after LP-PDBBind-compatible retraining. Shuffling protein embeddings at inference reduces the screening advantage. However, retrained shuffle controls recover much of the performance, so correct alignment is not established as necessary. BioLM-Score strengthens geometric scoring on screening tasks without retraining the language models. Availability and implementation Code, dataset-split identifiers, preprocessing and evaluation scripts, and language-model embedding-generation utilities are available at https://github.com/SZU-ADDG/BioLM-Score. An archival code snapshot and pretrained model weights are available at doi:10.5281/zenodo.21878818. Supplementary information Supplementary data are available at Bioinformatics online.

Authors

Institutions

Publication Details

Journal
Bioinformatics
Published
2026-11-14
DOI
https://doi.org/10.1093/bioinformatics/btag839
Primary Topic
Computational Drug Discovery Methods
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

BioLM-Score: Language-Prior Conditioned Probabilistic Geometric Potentials for Protein-Ligand Scoring

Zhangfan Yang, Ruibin Bai, Baoyun Chen, Jia Xu Wang et al.
2 links
Bioinformatics
Computational Drug Discovery Methods
article

BioLM-Score: Language-Prior Conditioned Probabilistic Geometric Potentials for Protein-Ligand Scoring

Zhangfan Yang, Ruibin Bai, Baoyun Chen, Jia Xu Wang, Dong Xu, Junkai Ji, Zexuan Zhu
article en

Abstract

Abstract Motivation Protein-ligand scoring must address affinity prediction, pose selection, and virtual screening, but these objectives are not necessarily optimized together. Geometric likelihood models may generalize poorly when training-set structural coverage is limited. We test whether pretrained biomolecular representations improve scoring under a controlled geometric baseline. Results BioLM-Score integrates frozen protein and molecular language-model embeddings into a residue-atom distance-likelihood framework via gated graph convolutions. We evaluate geometry-only, jointly trained, and two-stage fine-tuned variants, using the jointly trained variant for the primary controlled comparisons. All three DEKOIS 2.0 screening endpoints improve over the matched geometric baseline with target-level paired support, but affinity scoring and ranking do not improve uniformly. Under a time-split protocol, scoring is significantly lower and ranking is also lower than the geometric baseline, while screening gains persist. On BDB2020+, the difference in affinity correlation between BioLM-Score and GenScore narrows after LP-PDBBind-compatible retraining. Shuffling protein embeddings at inference reduces the screening advantage. However, retrained shuffle controls recover much of the performance, so correct alignment is not established as necessary. BioLM-Score strengthens geometric scoring on screening tasks without retraining the language models. Availability and implementation Code, dataset-split identifiers, preprocessing and evaluation scripts, and language-model embedding-generation utilities are available at https://github.com/SZU-ADDG/BioLM-Score. An archival code snapshot and pretrained model weights are available at doi:10.5281/zenodo.21878818. Supplementary information Supplementary data are available at Bioinformatics online.

Bioinformatics
University of Nottingham Ningbo China (CN), Shenzhen University (CN), Shenzhen Technology University (CN)
Computational Drug Discovery Methods
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon