GeoGAT-site

Knowing which patch of a protein's surface will bind a partner is a prerequisite for designing inhibitors, engineering binders, and narrowing the search space of docking. Geometric deep learning approaches represent the solvent-excluded surface as a triangulated mesh and learn from local "surface fingerprints" — patches carrying shape and physicochemical descriptors. The established form of this idea, MaSIF, centers one patch on every mesh vertex, which produces on the order of 402 million patches across a few tens of thousands of structures and makes preprocessing, rather than training, the dominant cost.

GeoGAT-site, from Harbin Institute of Technology and collaborators in Harbin and at KAUST, changes the unit of representation from vertices to triangular faces. Each face carries the mean of its three vertices' normal vector and physicochemical attributes, and a 9 Å neighborhood of faces becomes a graph whose edges encode inter-face distance and the cosine angle between face normals. A graph attention network then multiplies its ordinary feature-based attention score by an explicit geometric term, so the model weights neighbors by how close and how co-oriented their surface elements are rather than inferring that from features alone. Aggregating over faces cuts patch-generation time by 72.51% — a 3.64-fold speedup — relative to the vertex-centered pipeline.

The result is a single fixed checkpoint applied to unseen structures without per-target retraining, and it sits alongside sequence-based interface predictors such as SPPIDER and PSIVER and protein-language-model approaches like PLMDA-PPI, while complementing full-complex predictors such as AlphaFold-Multimer by supplying interface constraints to docking. The work is a July 2025 bioRxiv preprint and has not been peer reviewed.

#Key Features

• Face-centered surface fingerprints: Patches are centered on triangular mesh faces rather than vertices, yielding 150 million fingerprints from over 20,000 structures instead of the ~402 million vertex-centered patches a comparable corpus would produce.
• Geometry-modulated attention: Attention coefficients are the product of a standard feature-based score and a geometric term combining a Gaussian distance decay with the cosine of the angle between face normals, making the model sensitive to concave pockets versus planar contacts.
• Chemistry and geometry in one node vector: Each node concatenates a surface normal with electrostatic potential, hydrogen-bond strength, and hydrophobicity, averaged across the face's vertices.
• Coverage across interaction classes: Performance is reported separately for enzyme-substrate, receptor-ligand, and antibody-antigen complexes rather than as a single pooled number.
• Docking-ready output: Per-patch interface probabilities can be used to constrain the translational and rotational search of docking tools.

#Technical Details

Structures from the PDB are protonated with PDB2PQR under the AMBER force field at pH 7.0, triangulated with MSMS (probe radius 3.0 Å, density 1.0), and annotated with APBS electrostatics, HBPLUS hydrogen bonding, Kyte-Doolittle hydrophobicity, and FreeSASA interface labels (ΔSASA > 0.1 Ų), giving a positive-to-negative ratio near 1:10. The network is two GeoGATConv layers with 6 input channels, 128 hidden channels, 64 output channels, and four attention heads, followed by global mean pooling and a linear classifier; it is trained with weighted cross-entropy and Adam (learning rate 5×10⁻⁴, weight decay 5×10⁻⁴). The 150-million-fingerprint corpus is split 90M/30M/30M for training, validation, and internal testing.

On an independent set of 167 complexes (~1 million patches, filtered to <30% sequence identity and TM-score <0.5 against training), GeoGAT-site reaches ROC AUC 0.89 versus 0.845 for MaSIF-site, 0.65 for SPPIDER, and 0.63 for PSIVER, with 0.90 on enzyme-substrate, 0.88 on receptor-ligand, and 0.87 on antibody-antigen subsets. Ablations attribute much of this to the edge terms: the full model scores 0.8956 on the independent set against 0.8166 without edge attributes, 0.7683 with chemical features only, and 0.7134 with geometric features only.

#Applications

The model targets structure-driven workflows where a binding surface must be localized before anything else can proceed: prioritizing hotspot regions for protein-protein interaction inhibitors, choosing which surface patches to mutate when engineering a binder or an antibody, restricting docking search space in tools such as ZDock, and screening candidate interaction partners at scale. Because it reads only surface geometry and physicochemistry, it applies to protein families with no close homolog of known complex structure — where sequence-based predictors degrade most.

#Impact

GeoGAT-site's contribution is less a new learning objective than a representation-level argument: on triangulated molecular surfaces, faces are a cheaper and equally informative unit than vertices, and the geometry that vertex-centered methods encode through geodesic convolutions can instead be pushed into attention edge weights. Real limitations remain. Training uses bound (holo) structures, so accuracy on unbound conformations is untested; averaging over each face smooths vertex-level chemical detail and may hurt small interfaces; and evaluation is entirely computational, with no experimental validation of predicted sites. The preprint states that code and dataset are publicly available but gives no repository or download location, and no public release has appeared, so the model cannot currently be run by others.