Loading…
Protein-protein interface prediction that summarizes molecular surface patches with persistent homology descriptors, at 0.77 test AUC.
No providers recorded yet. Browse all providers
Identifying which patch of a protein's solvent-excluded surface is likely to bind a partner is a prerequisite for mapping druggable regions, designing inhibitors, and narrowing the search space of docking. The established approach represents the surface as a triangulated mesh, cuts it into local patches carrying geometric and physicochemical descriptors, and learns patch-level convolutions — the MaSIF family being the canonical example. These methods predict interfaces accurately, but surface parameterization, meshing, and feature mapping make preprocessing rather than training the dominant cost, and the learned feature extractors need substantial data to generalize.
TDA-NN, from the chemical engineering and statistics departments at the University of Wisconsin-Madison, replaces the learned surface convolutions with descriptors computed analytically from topology. Each local patch is treated as a 3D point cloud; a simplicial complex is built over it, and persistent homology summaries of that filtration become the patch's shape signature. Those summaries, augmented with local chemical features, feed a small feedforward classifier that scores each surface point as interface or non-interface. Because the descriptors are computed rather than learned, the pipeline is far cheaper to run and the model itself is tiny.
The entry sits alongside two related lines in the catalog: GeoGAT-site, which keeps learned attention but changes the surface unit from vertices to triangular faces, and TopoBind, which also uses persistent homology at interfaces but as one input stream to a much larger sequence-plus-structure model.
The combined topological and chemical descriptor vector is reduced by PCA to the number of components explaining more than 95% of cumulative variance — 20 components at a 9 Å radius — and passed to a feedforward network of five ReLU hidden layers with a sigmoid output, totaling 1,799 trainable parameters against 9,267 for the corresponding 9 Å MaSIF-site model. Training minimizes binary cross-entropy with the Adam optimizer over balanced samples of interface and non-interface patches, on 3,362 protein chains from the PDB split into 3,003 training and 359 test structures. Against MaSIF-site on the same data, TDA-NN reaches mean test AUC of 0.76 at 9 Å and 0.77 at 12 Å versus 0.84, while cutting preprocessing from roughly 27 s per protein to 5-8 s and total training from about 6 hours to 1-1.3 hours; training curves saturate after roughly 14-16 epochs. An ablation shows the two descriptor families are complementary but individually weaker: 0.73 test AUC for topology alone and 0.74-0.75 for chemistry alone.
The model is aimed at settings where interface scoring has to be run many times rather than once: sweeping large structure collections for interaction-prone or druggable surface regions, re-scoring conformational ensembles and molecular dynamics trajectories where a single structure is not representative, and prefiltering candidate binding regions before docking or binder design. The low preprocessing cost and small parameter count mean the whole workflow runs on modest hardware, which matters for groups without access to large GPU allocations.
TDA-NN is a preprint that has not yet been peer reviewed, and it is deliberately not a bid for best-in-class accuracy — it trails MaSIF-site by 0.07-0.08 AUC and is evaluated on one dataset and one task. Its contribution is a demonstration that analytically computed topological summaries of surface patches carry enough interface-relevant signal to support a classifier roughly five times smaller than the surface convolution network it is benchmarked against, at a fraction of the preprocessing cost. Source code, feature-extraction scripts, and data-preparation documentation are published in the Zavala group's repository; no trained checkpoint is distributed, and the repository carries no license file, so downstream reuse of the code rests on default copyright.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.