SiteAF3

Structure predictors for biomolecular complexes normally start from nothing: given a receptor sequence and a ligand, they sample the whole system at once and decide for themselves where the ligand goes. In structure-based drug discovery that throws away information researchers usually already have — an experimental receptor structure, a known binding pocket, a set of hotspot residues from mutagenesis. It also produces a characteristic failure mode, where a ligand that binds an allosteric site is confidently docked into the orthosteric one instead.

SiteAF3 addresses this by restructuring how AlphaFold 3 runs its diffusion sampler. The receptor's atomic coordinates are held fixed throughout denoising, ligand atoms are initialized as Gaussian noise centered on the specified pocket, and a mask applied in the sequence-local attention block restricts coordinate updates to the ligand. Binding pocket and hotspot residue information can additionally be injected through the MSA representation, giving the network an explicit signal about which part of the surface matters. The result is site-specific folding: the ligand is folded into a chosen site rather than searched for across the whole receptor.

The method was developed by Haocheng Tang and Junmei Wang in the School of Pharmacy at the University of Pittsburgh, posted to bioRxiv in July 2025 and published in PNAS in October 2025. Unlike the open AlphaFold 3 reimplementations — Boltz-2, Chai-1, Protenix — SiteAF3 is not a separately trained model. It is an MIT-licensed patch layered over Google DeepMind's AlphaFold 3 codebase, shipping no weights of its own and running on the AlphaFold 3 parameters that users must obtain separately under DeepMind's restricted-access terms.

#Key Features

• Fixed-receptor conditional diffusion: Receptor coordinates stay constant while only ligand atoms are denoised, concentrating the sampler's capacity on ligand conformation and pose within a known structural context.
• Pocket and hotspot conditioning: Four selectable embedding modes — pocket MSA, hotspot MSA, AlphaFold 3 MSA, and pocket-masked AlphaFold 3 MSA — let users supply as much or as little prior site knowledge as they have.
• Coverage across ligand classes: The same procedure applies to protein–small molecule, protein–peptide, protein–dsDNA, and protein–RNA complexes, with helper scripts that classify chains by atom composition and residue ratios.
• Allosteric and orphan-protein cases: The largest gains appear where the baseline is weakest — allosteric binders and receptors with few structural homologs.
• Plug-in deployment: Installed by replacing one AlphaFold 3 model file and running a single inference script, so existing AlphaFold 3 environments carry over.

#Technical Details

On PoseBusters V2 (308 protein–small molecule complexes), SiteAF3 reaches an 81.2% success rate at the standard RMSD ≤ 2 Å threshold versus 73.1% for AlphaFold 3, with median RMSD improving from 0.55 Å to 0.38 Å. On the 558-complex FoldBench protein–ligand set, success rises from 62.0% to 71.6% and median RMSD from 1.20 Å to 0.66 Å under the pocket-masked MSA mode. Peptide complexes on PepPCBench improve from a median DockQ of 0.86 to 0.92; FoldBench protein–dsDNA from 0.59 to 0.71; and the AlphaFold 3 protein–RNA benchmark from 0.25 to 0.36. On a 14-complex allosteric binder set, average RMSD fell to under a third of AlphaFold 3's and median RMSD to about a fifth, with correct poses recovered across many of the 25 sampled predictions rather than one or two.

Because the receptor is supplied rather than predicted, runs can skip AlphaFold 3's genetic search stage, and the coordinate mask reduces the memory footprint of denoising, extending the practical token limit on a single workstation GPU. The trade-off is tight coupling to the upstream codebase: the repository pins a specific AlphaFold 3 commit and documents the call-site edits needed for newer versions.

#Applications

The natural users are structure-based drug discovery groups that already know the target site: predicting binding poses for a series of analogs against a crystallized receptor, modeling allosteric ligands that generic complex predictors misplace, prioritizing compounds in virtual screening, and building models of peptide or nucleic-acid binders at a defined interface. Because prior knowledge enters as pocket or hotspot residue lists, the method fits workflows where mutagenesis, crosslinking, or fragment screening has already localized an interaction but no complex structure exists.

#Impact

SiteAF3 is a targeted intervention rather than a new foundation model, and its value is in showing how much accuracy is available from conditioning an existing predictor on information experimentalists routinely hold. The code is MIT-licensed and openly available, but the workflow inherits AlphaFold 3's access restrictions, so it is not usable without separately approved weights, and its gains depend on the user supplying a correct site — a wrongly specified pocket constrains the sampler toward the wrong answer. Evaluation is in silico against redocking and complex benchmarks, without accompanying experimental validation.