AANet

Structure-based virtual screening ranks large compound libraries against a protein target by scoring how well each molecule fits a binding pocket. Almost every method in this family — physics-based docking and deep-learning scoring alike — was developed and benchmarked on holo structures, meaning crystal structures solved with a ligand already bound and the pocket therefore already identified. That assumption breaks in the setting where virtual screening matters most: a first-in-class target with no experimentally resolved complex, where the only structure available is an unliganded (apo) crystal form or a prediction from AlphaFold 2, and the pocket boundaries are unknown.

AANet, from the Institute for AI Industry Research at Tsinghua University with collaborators at BAAI and the University of Electronic Science and Technology of China, addresses that gap directly. Rather than requiring a single correct pocket, it enumerates candidate cavities detected geometrically on the input structure and learns to reason over the whole set. Two ideas do the work: a tri-modal contrastive alignment that pulls ligand, holo-pocket, and detected-cavity representations into a shared space, and a cross-attention adapter that aggregates evidence across candidate cavities conditioned on the query ligand.

Because the aggregation module tolerates pocket ambiguity, AANet can be trained on activity data that carry no structural annotation at all — bioactivity records mapped onto predicted structures — dramatically widening the usable training signal. The model is pretrained once and applied across benchmark targets without per-target fine-tuning. It was released in June 2025 and accepted to NeurIPS 2025.

#Key Features

• Tri-modal contrastive alignment: Ligand, holo pocket, and geometrically detected cavity representations are aligned with pairwise-sigmoid losses and hard-negative mining over non-binding cavities, making the scorer tolerant of pocket localization error.
• Cross-attention cavity aggregation: An adapter dynamically weights candidate cavity embeddings conditioned on the ligand, so the model does not need a committed pocket annotation at either training or inference time.
• Learns from pocket-agnostic activity data: Aggregation lets ChEMBL bioactivity records be paired with predicted structures, unlocking multi-target training data that pocket-conditioned scorers cannot consume.
• Robust to the pocket detector: Fpocket supplies the primary cavity proposals, with performance validated against PocketFinder, SURFNET, and LIGSITE.
• Holo performance is retained: Gains in the apo and blind settings do not come at the cost of accuracy on conventional ligand-bound structures.

#Technical Details

AANet initializes its molecular and pocket encoders from Uni-Mol and trains in two phases. Phase one performs tri-modal contrastive alignment on the PDBbind 2020 general set; phase two trains the cross-attention aggregation adapter on ChEMBL35 activity records mapped to predicted structures. Targets appearing in the evaluation benchmarks are excluded from both phases to prevent leakage.

On the 38-target DUD-E benchmark the model reaches an EF1% of 40.85 on holo structures, against 33.70 for DrugCLIP and 17.25 for Glide. The gap widens as pocket information degrades: on experimentally determined apo structures with annotated pockets AANet scores 38.03 (DrugCLIP 21.36, Glide 7.74), and in the fully blind setting on predicted structures it reaches 37.19 versus 12.05 for DrugCLIP. On the harder 12-target LIT-PCBA benchmark it scores 7.54 EF1% on holo and 5.40 in the blind predicted-structure setting, where DrugCLIP falls to 0.88. Code is released under Apache 2.0; the trained checkpoint is distributed separately under CC BY-NC 4.0, restricting weight reuse to non-commercial work.

#Applications

The intended use is early-stage hit finding against targets that lack a solved complex — novel or poorly characterized proteins where the only structural input is an apo crystal form or a predicted model. Computational chemists can screen libraries against such targets without first committing to a pocket definition or running a separate site-prediction step, and the same pretrained model applies across targets without retraining. The training recipe is also of interest independently: it shows how to fold large bioactivity archives into a structure-based scorer despite their missing structural annotations.

#Impact

AANet reframes pocket uncertainty as something a screening model should absorb rather than something upstream tooling must eliminate, and the size of its blind-apo margin makes the case that holo-only benchmarks have been overstating how well structure-based scorers transfer to real discovery campaigns. As a 2025 NeurIPS paper with a modest public codebase, its adoption is still early, and the reported enrichment gains come entirely from retrospective benchmarks rather than prospective screens with experimental confirmation. The non-commercial weight license also limits industrial deployment, though the Apache-2.0 code permits retraining.