Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
AANet: Virtual Screening under Structural Uncertainty via Alignment and Aggregation
Authors: Wenyu Zhu, Jianhui Wang, Bowen Gao, Yinjun Jia, Haichuan Tan, Ya-Qin Zhang, Wei-Ying Ma, Yanyan Lan
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluated our method on a newly curated benchmark of apo structures, where it significantly outperforms state-of-the-art methods in blind apo setting, improving the early enrichment factor (EF1%) from 11.75 to 37.19. Notably, it also maintains strong performance on holo structures. These results demonstrate the promise of our approach in advancing firstin-class drug discovery, particularly in scenarios lacking experimentally resolved protein-ligand complexes. Our framework outperforms both physics-based and DL-based baselines, achieving near-holo performance on both predicted and experimental apo structures. |
| Researcher Affiliation | Academia | 1Institute for AI Industry Research, Tsinghua University, Beijing, China 2Beijing Frontier Research Center for Biological Structure, Tsinghua University, Beijing, China 3Beijing Academy of Artificial Intelligence (BAAI), Beijing, China 4Department of Computer Science and Technology, Tsinghua University, Beijing, China 5University of Electronic Science and Technology of China, Chengdu, China Correspondence to EMAIL |
| Pseudocode | Yes | Algorithm 1 Cavity Extraction Require: protein structure coordinates xn, co-crystallized ligand coordinates xm {C(s)}S s=1 detector(xn) each cavity: a set of alpha-sphere centers for s = 1 to S do R(s) c { r residues | x r, c C(s), x c d } residue-level pocket end for Return: candidate pockets {R(s) c }S s=1 Algorithm 2 Tri-Modal Contrastive Alignment Require: pocket encoder F, cavity encoder Fs, ligand encoder G, candidate cavities {P (s) c }S s=1, holo pocket Pl, positive Io U threshold τpos, negative Io U threshold τneg for each protein ligand complex (xn, l, Pl) do Io U(Pl, Pc) = |Pl Pc| |Pl Pc| compute overlap ratio between holo pocket and candidate Pc {P (s) c }S s=1 with Io U(Pl, Pc) τpos sample one positive cavity {P c } random half of P (s) c : Io U(Pl, P (s) c ) τneg sample hard negatives LCL = Lp,l F(Pl), G(l) + Lp,l Fs(Pc), G(l) + Lp,p Fs(Pc), F(Pl) Update parameters of F, Fs, G by minimizing LCL end for Return: pretrained pockt encoder F, pretrained cavity encoder Fs, pretrained ligand encoder G Algorithm 3 Training Cross-Attention Adapter for Cavity Aggregation |
| Open Source Code | Yes | Our implementation is publicly available at https: //github.com/Wiley-Z/AANet. |
| Open Datasets | Yes | We curated a benchmark from the DUD-E dataset [14] and LIT-PCBA [15]... Our model was initialized with Uni-Mol [30] and fine-tuned on the PDBBind 2020 general set [31]... For dynamic aggregation training, binding and activity data from Ch EMBL35 [32]... predicted apo structures from the Alpha Fold Protein Structure Database [33]... The COACH420 dataset for pocket identification was obtained from the P2Rank [28] repository... |
| Dataset Splits | Yes | We curated a benchmark from the DUD-E dataset [14] and LIT-PCBA [15]... A total of 38 targets with all three structure types were retained... Twelve targets were retained, and three were excluded... The COACH420 dataset for pocket identification was obtained from the P2Rank [28] repository and deduplicated against PDBbind... Our model was initialized with Uni-Mol [30] and fine-tuned on the PDBBind 2020 general set [31], with all entities overlapping with DUD-E or LIT-PCBA removed... All Uni Prot entries matching any PDB entity in DUD-E or LIT-PCBA were excluded to avoid data leakage. To further address potential leakage, we conducted an additional experiment by using MMseqs2 [43] to identify and remove any training proteins (from PDBbind and Ch EMBL) containing chains with 90% sequence identity to any DUD-E structure / target sequences. |
| Hardware Specification | Yes | Training Alignment phase 4 NVIDIA A100 (80 GB) 2 h per run Training Aggregation phase 4 NVIDIA A100 (80 GB) 6 h per run Model testing (single benchmark) 1 NVIDIA A100 (80 GB) 1 5 min per benchmark Docking 128-core CPU server 3 7 days per benchmark Baseline evaluation (per DL model) 1 NVIDIA A100 (80 GB) minutes 1 day per benchmark |
| Software Dependencies | Yes | All modules were from the Schrödinger Suite 2024-1 distribution. |
| Experiment Setup | Yes | Table S4: Training and evaluation hyperparameters used in this study. Data processing Max. No. ligand conformers 10 Min. RMSD among ligand conformers 1 Pocket radius 10 Io U for positives 0.5 Io U for negatives 0.1 Learning rate 0.001 Batch size 48 Random seed 1 Model hyperparameters Following Drug CLIP Cavity negative ratio 0.5 Max. epochs 200 Early stopping 10 / 5 Loss logit scale log(10) Loss logit bias 10 Adapter softmax temperature 5 Use fp16 True |