GatorAffinity

GatorAffinity is a geometric deep learning scoring function for protein-ligand binding affinity prediction, developed in the Li Lab at the University of Florida and released as a bioRxiv preprint in late 2025. Predicting binding affinity from three-dimensional protein-ligand complexes is a fundamental task in computational drug discovery, but progress has been constrained by data scarcity: the widely used PDBbind dataset contains fewer than 20,000 experimental structures with annotated affinities, even as a vast number of measured affinities go unused because they lack accompanying structures.

GatorAffinity attacks this bottleneck by generating structural data at scale. The authors curate more than 450,000 synthetic protein-ligand complexes annotated with Kd and Ki values using the Boltz-1 structure prediction model, then augment these with over one million synthetic IC50-annotated complexes drawn from the recently released SAIR database. Pretraining on this large synthetic corpus and fine-tuning on high-quality experimental PDBbind structures yields a model that generalizes markedly better than prior approaches, positioning synthetic structure generation as a practical route around the experimental data ceiling.

#Key Features

• Large-scale synthetic pretraining: The model is pretrained on over 1.45 million synthetic complexes—450,000+ Boltz-1-generated Kd/Ki structures plus 1M+ IC50 complexes from SAIR—far exceeding the scale of experimental datasets.
• ATOMICA backbone: GatorAffinity builds on the ATOMICA all-atom geometric architecture, which delivers strong performance improvements even with limited experimental pretraining structures.
• Experimental fine-tuning: A final fine-tuning stage on high-quality PDBbind (LP-PDBbind) structures adapts the pretrained model to measured binding affinities for its best accuracy.
• Leak-proof benchmarking: On a leak-proof evaluation designed to prevent train-test contamination, GatorAffinity significantly outperforms state-of-the-art affinity prediction methods in accuracy and generalizability.
• Open model and dataset: The pretrained model and the large-scale synthetic dataset GatorAffinity-DB are released, with MIT-licensed code and a ready-to-run inference script.

#Technical Details

GatorAffinity is a geometric deep learning model that uses the ATOMICA backbone to represent all-atom protein-ligand complexes and predict binding affinity. Pretraining proceeds on the combined synthetic corpus—Boltz-1-derived Kd/Ki complexes and SAIR-derived IC50 complexes—before fine-tuning on experimental structures from the leak-proof LP-PDBbind split. The repository ships pretrained checkpoints (a base model pretrained on the IC50+Kd+Ki data and an experimentally fine-tuned best-performance variant) together with train.py and inference.py scripts. Source code is released under the MIT license, while the distributed model checkpoints carry a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 license. Evaluation on the leak-proof benchmark shows consistent gains over competing scoring functions, which the authors attribute to the scale and diversity of the synthetic pretraining data.

#Applications

GatorAffinity is intended for structure-based drug design, virtual screening, and lead optimization, where accurate ranking of protein-ligand binding affinities accelerates candidate prioritization. By releasing both the model and the GatorAffinity-DB synthetic dataset, the authors provide a reproducible foundation that computational chemists and machine learning researchers can use to score complexes, screen compound libraries, and build on the synthetic-data strategy for related affinity and docking tasks.

#Impact

GatorAffinity demonstrates that augmenting scarce experimental structures with large volumes of AI-generated synthetic complexes can meaningfully improve binding affinity prediction while maintaining reliability on held-out data. This reframes structure prediction models such as Boltz-1 as engines for training-data generation, not just endpoints, and offers a scalable template for other structure-based prediction problems. As a preprint awaiting peer review, its advantages rest on the reported leak-proof benchmark, and the non-commercial license on the released checkpoints constrains direct commercial deployment.