BioScore

Predicting a complex's three-dimensional structure and judging how tightly its components actually bind are different problems, and the second has not kept pace with the first. Scoring functions — the modules that assign a number to a docked pose or an affinity to a complex — remain almost entirely system-specific. Most deep-learning scoring functions are trained on protein–small molecule data and transfer poorly to protein–protein interfaces, nucleic acid complexes, cyclic peptides, or carbohydrates, and structure predictors are not a substitute: AlphaFold 3 ranks binding affinities poorly enough that it cannot reliably separate active from inactive compounds.

BioScore is a scoring function built to work across those systems from one set of weights. It was developed at Zhejiang University's College of Pharmaceutical Sciences with collaborators at Baidu, and released as a preprint in July 2025. The authors position it as a foundational scoring function: a single dual-scale geometric graph model, pretrained on mixed-system structural data and then fine-tuned per task, that supports affinity prediction, conformation ranking, and structure-based virtual screening for protein–ligand, protein–protein, protein–nucleic acid, nucleic acid–ligand, antibody–antigen, peptide–MHC, cyclic peptide, macrocycle, and carbohydrate complexes.

The work sits alongside interface representation models such as ATOMICA but targets the decision-making end of the pipeline rather than the encoder alone, and extends the same group's earlier protein–ligand scoring functions, RTMScore and GenScore, which could not predict absolute affinities or generalize beyond small-molecule ligands.

#Key Features

• Cross-system representation: Interface-masking encoding and distance-aware edge construction produce dual-scale atom- and block-level features, so proteins, nucleic acids, sugars, and small molecules are described in one vocabulary instead of separate receptor and ligand encoders.
• Dual-tower scoring: One output head derives a statistical potential from a mixture density network for docking and screening; a second maps block-pair representations to binding free energy for scoring and ranking, with an interaction-edge-count confidence term shared by both.
• Cross-domain pretraining: Pretraining on mixed protein–ligand and protein–protein structures before single-domain fine-tuning raised protein–protein affinity metrics by 43% on average and antibody–antigen Pearson correlation by 92%.
• Low-data transfer: Fine-tuned on only 276 protein–nucleic acid entries, it improves Pearson correlation over the best baseline by 71%, and transfers zero-shot to nucleic acid–ligand complexes, cyclic peptides, and non-peptidic macrocycles.
• A PPI scoring benchmark: The authors released a CASF-style protein–protein benchmark — 79 core complexes, 7,979 docking decoys, and 613,900 cross-docked screening conformations — to fill a gap in standardized PPI evaluation.

#Technical Details

The encoder is a Geometric Equivariant Transformer (GET), an E(3)-equivariant architecture with dual-level attention that jointly models atom-level and block-level interactions; the released checkpoints use three layers with a hidden size of 128, giving a compact model of roughly 12 MB per checkpoint. Inter-molecular edges are built at an 8 Å cutoff, with intra-molecular thresholds of 2 Å for small molecules and 10 Å for proteins and nucleic acids. Training data come from PDBbind v2020 — nearly 20,000 protein–ligand entries and about 1,700 protein–protein structures — supplemented by 272 SAbDab antibody–antigen complexes and 370 protein–carbohydrate complexes for the corresponding fine-tunes. Pretraining minimizes a negative log-likelihood over block-pair distance distributions; fine-tuning combines MSE, correlation, and MDN losses. Across 16 benchmark tasks and 70 baselines, BioScore ranks first on 9 of 13 metrics on the new PPI Benchmark, outperforms RTMScore and GenScore on CASF-2016 scoring, ranking, and screening while placing third on docking, and leads all three enrichment-factor metrics on DUD-E. On RNA–ligand affinity it trails only MM/GBSA, a far more expensive physics-based method.

#Applications

BioScore targets structure-based drug discovery beyond the classical small-molecule regime: selecting docked poses, ranking analog series during lead optimization, and enriching actives in virtual screens against protein–protein and protein–nucleic acid targets where dedicated scoring functions barely exist. Its handling of cyclic peptides, non-peptidic macrocycles, and carbohydrates is useful for modalities that sit between small molecules and biologics, and the antibody–antigen and peptide–MHC results extend it to immunology workflows. Because one pretrained checkpoint serves docking and screening while task-specific fine-tunes handle scoring and ranking, teams can adopt it without building a new model per target class.

#Impact

BioScore addresses an asymmetry in structural biology tooling: enormous investment in predicting structures, comparatively little in judging the interactions they imply. Its central empirical claim — that geometric patterns learned from abundant protein–ligand structures transfer to data-poor systems like protein–nucleic acid complexes — is the argument for treating scoring as a foundation-model problem. The work remains a preprint awaiting peer review, all evaluation is in silico with no prospective experimental validation, and absolute affinity accuracy still depends on fine-tuning against limited labeled data. Code and all pretrained and fine-tuned checkpoints are released under an MIT license, and the PPI Benchmark is distributed through Zenodo and Hugging Face.