AuroBind

Only about 3-4% of human protein-coding genes currently have a drug candidate against them, and structure-based virtual screening has long promised to widen that fraction. The obstacle is that the co-folding models now dominant in structure prediction were never trained to answer the question a screen actually asks. They predict what a complex looks like; they do not predict whether a compound binds, or how tightly. Confidence scores are a weak substitute, and running a full diffusion-based predictor over tens of millions of compounds is computationally out of reach.

AuroBind, from a multi-institution collaboration led by Shanghai Jiao Tong University, closes that gap by fine-tuning an atomic-level co-folding model on million-scale chemogenomic data so that it emits a ligand-bound complex structure and a scalar binding-fitness score in a single forward pass. The architecture retains the general layout of AlphaFold 3 and adds three extensions: a chemogenomic embedding module with a fitness head that predicts per-residue and global scores, a self-distillation stage that refines structural accuracy on high-confidence complexes, and a distilled student network, AuroFast, for library-scale triage. Inputs are a raw protein sequence and a ligand SMILES string, with no predefined binding pocket required.

What separates AuroBind from most virtual screening methods is that its claims rest on a prospective wet-lab campaign rather than retrospective benchmarks alone. The authors screened a 30-million-compound library against ten therapeutically relevant targets and tested roughly 30-50 compounds per target experimentally.

#Key Features

• Joint structure and fitness prediction: A single forward pass yields both an atomic-resolution protein-ligand complex and a scalar potency estimate, removing the usual split between a pose predictor and a separate scoring function.
• Preference-optimized fitness alignment: After an initial mean-squared-error objective, the fitness head is refined with direct preference optimization, using structural confidence as soft weights so that noisy or ambiguous binding measurements contribute less.
• Self-distillation on high-confidence complexes: Predicted structures over BindingDB were filtered to 230,000 high-confidence, high-fitness complexes and fed back as additional training data.
• AuroFast student model: A distilled lightweight network screens 20-30 million compounds per target in hours at roughly 100,000x the throughput of the full model, feeding its top candidates back to AuroBind for rescoring.
• Pocket-free generalization to orphan targets: Screens require no co-crystal structure, no known actives, and no defined binding site, which is what allowed hits to be found for the orphan GPCRs GPR151 and GPR160.

#Technical Details

Training proceeds in two stages. The first is supervised structure prediction on roughly 100,000 protein-ligand complexes curated from the PDB, followed by the self-distillation pass over BindingDB. The second fine-tunes on approximately 1.27 million chemogenomic pairs assembled from ChEMBL and PubChem under ExCAPE-DB cleaning rules, spanning about 1,300 protein domains with scalar pXC50 labels; a fitness head sits after the Pairformer trunk. Full training consumed 30 days on 96 NVIDIA H800 GPUs.

On the ConPLex splits for DAVIS and BindingDB, a zero-shot AuroBind variant reached AUPR of 0.22 and 0.24, against 0.08 and 0.13 for AlphaFold 3 confidence used as a proxy; the fine-tuned model reached 0.61 and 0.70. On LIT-PCBA, covering over 2.6 million protein-ligand pairs across 15 targets, AuroFast achieved an enrichment factor at 1% of 7.58, 37.6% above the strongest baseline. Structural fidelity survived the functional fine-tuning: on PoseBusters V1 and V2 the model reached 79.1% and 81.7% success, ahead of AlphaFold 3 (78.0% and 81.0%) and of Protenix by 6.0 and 6.8 points.

#Applications

The intended workflow is on-demand hit discovery: point AuroFast at a purchasable library of tens of millions of compounds, rescore the survivors with AuroBind, and filter for drug-likeness, solubility, and novelty before ordering. A full screen against one target completes within 24 hours on two H800 GPUs, putting ultra-large library screening within reach of an academic cluster. Because no pocket definition or known ligand is needed, the approach extends to understudied targets that classical docking cannot address, and a hosted server allows use without a local install.

#Impact

Across ten targets spanning receptor tyrosine kinases, serine kinases, an epigenetic regulator, and GPCRs, experimental hit rates ran from 7% for HDAC3 to 69% for GSK3-alpha at 10 uM, exceeding 20% for five targets. The strongest compound was a TrkB binder with an IC50 of 220 pM, and three sub-nanomolar hits were confirmed; 25% of the 283 tested compounds had Tanimoto similarity below 0.30 to any previously reported active, indicating genuinely new chemotypes rather than rediscovered scaffolds. Functional assays confirmed GPR160 modulation in liver and prostate cancer models. The work remains a preprint, the code and checkpoints carry a CC BY-NC-SA 4.0 license that bars commercial use, and weights are distributed through Google Drive rather than a model hub. The authors also note that performance on highly dynamic or disordered targets has not yet been established.