Research map / Probabilistic and scientific modeling / Molecules and proteins
AANet: Virtual Screening under Structural Uncertainty via Alignment and Aggregation
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 AANet, a tri-modal contrastive alignment plus cross-attention cavity aggregation framework for virtual screening on apo and AlphaFold2-predicted structures without known pocket annotations, evaluated on a new DUD-E/LIT-PCBA apo benchmark.
Problem
Structure-based virtual screening degrades sharply on apo and predicted protein structures because binding pockets must be inferred without ligand-bound (holo) annotations.
Approach
Align ligand, holo pocket, and geometrically detected cavities with pairwise-sigmoid contrastive losses (with hard negative cavities and enlarged pockets), then aggregate candidate cavities via a ligand-conditioned cross-attention adapter trained on pocket-agnostic activity data.
Key results
- On DUD-E blind apo-exp, AANet improves EF1% from 11.75 (DrugCLIP) to 37.19.
- On DUD-E holo, AANet reaches BEDROC 0.6365 and EF1% 40.85, versus DrugCLIP 0.5157 and 33.70.
- On COACH420 top-1 pocket identification at 1 Å DCA, AANet achieves 0.1140 vs 0.0535 for DrugCLIP and 0.0728 for TankBind.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add cavity augmentation to contrastive pocket-molecule learning (ablation) | AANet alignment phase · DUD-E apo-exp and apo-pred (blind) | BEDROC: 0.0838 -> 0.2007; 0.0822 -> 0.2341 | Helped |
| add hard negative sampling of non-binding cavities (ablation) | AANet alignment phase with cavity augmentation · DUD-E apo-exp and apo-pred (blind) | BEDROC: 0.2007 -> 0.4016; 0.2341 -> 0.4097 | Helped |
| enlarge the pocket (ablation) | AANet alignment phase with cavity augmentation and negative sampling · DUD-E apo-exp and apo-pred (blind) | BEDROC: 0.4016 -> 0.4708; 0.4097 -> 0.4585 | Helped |
| replace argmax adapter with cross-attention adapter (ablation) | AANet aggregation phase · DUD-E apo-exp and apo-pred (blind) | BEDROC: 0.4639 -> 0.5550; 0.4620 -> 0.5996 | Helped |
| add pocket-label supervision to the attention adapter (ablation) | AANet aggregation phase · DUD-E apo-exp and apo-pred (blind) | BEDROC: 0.5550 -> 0.5764; 0.5996 -> 0.6232 | Helped |
| use AANet instead of the strongest baseline in blind apo-exp screening | DUD-E · DUD-E apo-exp (blind) | EF1%: 18.52 -> 37.19 | Helped |
| use AANet instead of the strongest baseline in blind apo-pred screening | DUD-E · DUD-E apo-pred (blind) | EF1%: 17.36 -> 40.85 | Helped |
| use AANet instead of the strongest baseline in blind apo-exp screening | LIT-PCBA · LIT-PCBA apo-exp (blind) | EF1%: 3.42 -> 3.92 | Helped |
| use AANet instead of the strongest baseline in blind apo-pred screening | LIT-PCBA · LIT-PCBA apo-pred (blind) | EF1%: 2.90 -> 5.40 | Helped |
Surprising findings
- Docking-free DrugCLIP shows a sharp drop in blind settings (EF1% 33.70 on holo to 11.75 on apo-exp blind), while TankBind stays flat due to training on detected pockets and a large binding radius.
- Enlarging the pocket radius from 6 Å to 10 Å with atom down-sampling further improved apo blind BEDROC in the ablation.
Limitations
- Performance on predicted apo structures slightly declines with alternative pocket detectors (e.g., PocketFinder, SURFNET, LIGSITE), relying on detector quality.
- Docking-based baselines were not evaluated in the blind setting due to computational cost, and LIT-PCBA docking used a single holo structure.
- Some targets with very low IoU between detected cavities and holo pockets show no clear correlation with performance.
Experimental setup
Datasets: PDBBind 2020 general set ChEMBL35 DUD-E LIT-PCBA COACH420 AlphaFold Protein Structure Database
Benchmarks: DUD-E LIT-PCBA COACH420
Metrics: BEDROC EF1% AUROC top-n DCA pocket success rate
Baselines: Glide-SP RTMScore EquiScore TankBind DrugCLIP PRANK
Builds on: DrugCLIP Uni-Mol sigmoid loss for language-image pre-training Fpocket
Closest papers by meaning
- CausalBind: Causal Modeling and Learning for Protein-Molecule Virtual Screening (NeurIPS 2026 · similarity 0.86)
- Drug-TTA: Test-Time Adaptation for Drug Virtual Screening via Multi-task Meta-Auxiliary Learning (ICML 2025 · similarity 0.86)
- 3DMolFormer: A Dual-channel Framework for Structure-based Drug Discovery (ICLR 2025 · similarity 0.86)
- Integrating Protein Dynamics into Structure-Based Drug Design via Full-Atom Stochastic Flows (ICLR 2025 · similarity 0.86)
- DeltaDock: A Unified Framework for Accurate, Efficient, and Physically Reliable Molecular Docking (NeurIPS 2024 · similarity 0.86)
- M²N: A Progressive Macro-to-Micro 3D Modeling Scheme for Unveiling Drug-Target Affinity (AAAI 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.