Research map / Label-efficient and robust learning / Self-supervised and contrastive

Protein-ligand binding representation learning from fine-grained interactions

ICLR 2024 · Label-efficient and robust learning › Self-supervised and contrastive · approach: Self-supervision in supervised pipelines · 4 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes BindNet, a self-supervised pre-training method that learns protein-ligand binding representations via a Transformer interaction module over frozen Uni-Mol pocket and ligand encoders, using atomic distance map prediction and masked ligand reconstruction objectives.

Problem

Deep learning models for protein-ligand binding tasks (affinity, virtual screening, docking) generalize poorly due to limited labeled data, and existing pre-training methods encode proteins and ligands separately without capturing inter-molecular interaction patterns.

Approach

Pre-train a Transformer-based interaction module on top of frozen pocket and ligand encoders to predict the bound complex structure from the primary (unbound) ligand, using inter-molecular atom-pair distance map regression (ADP) and masked ligand representation reconstruction (MLR) in representation space rather than atom-token space.

Key results

• On LBA 30%, BindNet achieves RMSE 1.340, Pearson 0.632 versus Uni-Mol's 1.434 and 0.565; on LBA 60%, RMSE 1.230 vs. 1.343 for the best prior (ProNet).
• On LEP, BindNet reaches AUROC 0.882 and AUPRC 0.870, versus Uni-Mol's 0.823 and 0.787.
• On CASF-2016 docking, BindNet attains 45.26% at 1.0 Å and 89.12% at 3.0 Å, versus Uni-Mol's 43.16% and 87.02%.

Measured effects reported in the paper

Surprising findings

• Supervised methods DrugVQA and AttentionSiteDTI achieve the best DUD-E AUC but drop to near-random AUC (~0.5) on the AD dataset, suggesting decoy bias.
• Supervised 3D models such as Atom3D-GNN degrade sharply out-of-distribution (LBA RMSE rises from 1.601 to 12.475 from CC to CD setting).

Limitations

• Relies on frozen Uni-Mol encoders, so the pocket and ligand representations are not themselves updated during pre-training.
• Docking uses a post-processing gradient-descent step for pose generation, and the 1.0 Å success rate remains below Smina.
• DUD-E results are weaker than supervised DrugVQA and AttentionSiteDTI on AUC and some RE metrics.

Experimental setup

Datasets: BioLiP PDBBind DUD-E AD dataset CASF-2016 Atom3D LEP

Benchmarks: PDBBind LBA 30% PDBBind LBA 60% Atom3D LEP DUD-E AD dataset CASF-2016 docking

Metrics: RMSE Pearson Spearman AUROC AUPRC AUC RE RMSD success rate

Baselines: Uni-Mol GeoSSL CoSP SMT-DTA DeepAffinity ProNet Holoprot AttentionSiteDTI DrugVQA AutoDock Vina Smina

Builds on: Uni-Mol Transformer

Closest papers by meaning

• Self-supervised Pocket Pretraining via Protein Fragment-Surroundings Alignment (ICLR 2024 · similarity 0.88)
• DecoyDB: A Dataset for Graph Contrastive Learning in Protein-Ligand Binding Affinity Prediction (NeurIPS 2025 · similarity 0.87)
• A Systematic Evaluation of Co-folding Model Representations for Small-Molecule Learning (NeurIPS 2026 · similarity 0.85)
• UniSite: The First Cross-Structure Dataset and Learning Framework for End-to-End Ligand Binding Site Detection (NeurIPS 2025 · similarity 0.84)
• CausalBind: Causal Modeling and Learning for Protein-Molecule Virtual Screening (NeurIPS 2026 · similarity 0.84)
• Knowledge Enhanced Representation Learning for Drug Discovery (AAAI 2024 · similarity 0.84)