Research map / Detection and segmentation / CNN features

UniSite: The First Cross-Structure Dataset and Learning Framework for End-to-End Ligand Binding Site Detection

NeurIPS 2025 · Detection and segmentation › CNN features · approach: Efficient DETR query design · 2 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.

Introduces UniSite-DS, a UniProt-centric ligand binding site dataset, and UniSite (1D/3D), an end-to-end set-prediction transformer for binding site detection, plus an IoU-based Average Precision metric.

Problem

Ligand binding site detection for proteins is evaluated and trained on PDB-centric datasets, which miss binding sites across multiple structures of the same protein; existing pipelines are discontinuous (segmentation then clustering) and DCC/DCA metrics allow double-counting and ignore shape/residue

Approach

Aggregate all ligand binding sites per UniProt ID into a dataset, cast binding site detection as set prediction with DETR-style queries and Hungarian bijective matching so overlapping sites are predicted directly, and evaluate with residue-level IoU-based COCO-style Average Precision.

Key results

• UniSite-3D reaches AP0.3 of 0.5603 on UniSite-DS vs 0.4469 for GrASP, and AP0.5 of 0.3835 vs 0.2848.
• On HOLO4K-sc, UniSite-3D attains AP0.3 of 0.7091 vs 0.6668 for GrASP; UniSite-1D (sequence only) reaches 0.6867.
• On COACH420, UniSite-3D achieves AP0.3 of 0.7196 vs 0.7150 for GrASP, while GrASP keeps the best DCC top-n (0.4851 vs 0.4702 for UniSite-3D).

Measured effects reported in the paper

Surprising findings

• The sequence-only UniSite-1D outperforms all structure-based baselines on UniSite-DS and HOLO4K-sc AP metrics.
• About 20% of proteins suffer double counting under DCC/DCA evaluation, and ground-truth residue centroids yield nonzero DCC (mean 2.15 Å) and DCA (1.57 Å).

Limitations

• Manual curation is required to remove unreasonable UniSite-DS entries.
• Training on UniProt-centric data and testing on PDB-centric benchmarks creates a train-test gap.
• VN-EGNN baseline residues had to be reconstructed via fixed-radius (9 Å) selection, so AP for it depends on this choice.

Experimental setup

Datasets: UniSite-DS HOLO4K COACH420 scPDB PDBbind

Benchmarks: UniSite-DS HOLO4K-sc COACH420

Metrics: AP@0.3 (IoU) AP@0.5 (IoU) DCC top-n success rate DCA top-n success rate recall at IoU thresholds

Baselines: Fpocket Fpocket-rescore P2Rank DeepPocket GrASP VN-EGNN

Builds on: DETR GearNet-Edge ESM-2 P2Rank Fpocket

Closest papers by meaning

• Towards precision protein-ligand affinity prediction benchmark: A Complete and Modification-Aware DAVIS Dataset (NeurIPS 2025 · similarity 0.87)
• DecoyDB: A Dataset for Graph Contrastive Learning in Protein-Ligand Binding Affinity Prediction (NeurIPS 2025 · similarity 0.87)
• UniMatch: Universal Matching from Atom to Task for Few-Shot Drug Discovery (ICLR 2025 · similarity 0.87)
• DeltaDock: A Unified Framework for Accurate, Efficient, and Physically Reliable Molecular Docking (NeurIPS 2024 · similarity 0.86)
• CBGBench: Fill in the Blank of Protein-Molecule Complex Binding Graph (ICLR 2025 · similarity 0.86)
• AANet: Virtual Screening under Structural Uncertainty via Alignment and Aggregation (NeurIPS 2025 · similarity 0.85)