Loading…
Antibody-antigen binding affinity predictor that scores the interface as a graph over frozen ESM-2 embeddings, jointly regressing log10(KD) and ΔG.
No providers recorded yet. Browse all providers
The dissociation constant of an antibody–antigen complex and its binding free energy are two readings of a single physical event, tied to one another by ΔG = RT ln K_D. Most affinity predictors nonetheless fit a separate regressor for each, so the two numbers returned for the same complex are free to drift out of thermodynamic register, and neither head gets to borrow the signal the other one learned. Phys-AbGAT treats that coupling as part of the training objective rather than as a post-hoc conversion.
The model represents the paratope–epitope interface as a spatial graph and reads both quantities off it at once. Residues that fall within 16.5 Å across the antigen–heavy and antigen–light contact surfaces become nodes, each carrying a frozen ESM-2 650M residue embedding alongside a one-hot chain identity and its Cartesian coordinates. Contacting residue pairs become edges annotated with a radial-basis expansion of their separation and a tag recording which antibody chain the contact belongs to. Two graph attention layers propagate over that interface before attentional and mean pooling collapse it to a graph-level vector, from which two regression heads emit log10(K_D) and ΔG under a loss whose third term penalises disagreement between the predicted and the observed ratio of those targets.
Phys-AbGAT was developed at Nankai University with collaborators at Nanchang University and the Tianjin Centers for Disease Control and Prevention, and published in Bioinformatics in September 2026.
Each node vector is 1286 values wide: 1280 from ESM-2 650M, three for a one-hot heavy/light/ antigen chain label, and three coordinates. Edge features begin as a distance plus a two-way source tag, expand through 16 radial basis functions, and are projected into the 128-dimensional hidden space that the network runs in. Two GATv2 layers with four attention heads, graph-size normalisation, residual connections and 0.2 dropout update the node states; attentional aggregation and global mean pooling are concatenated into a 256-dimensional graph embedding that feeds two 256→64→1 regression heads. The composite loss weights the ΔG term at 0.4, the log10(K_D) term at 0.6 and the physics term at 0.2. Training used 632 affinity-annotated complexes split 506 for training and 126 for validation, with a held-out set of 42 complexes for evaluation. On that independent benchmark the model reached Pearson correlations of 0.569 for log10(K_D) and 0.573 for ΔG. Matched-set comparisons against four baselines gave lower RMSE and higher correlation in every case, but only the margins over MVSF-AB and PPA-Pred2 survived Holm correction.
The intended use is ranking and triage in therapeutic antibody engineering: given a solved or modelled antibody–antigen complex, the model returns both quantities in one pass, without a docking run or a free-energy simulation. Because it requires only a heavy chain and an antigen chain, it covers conventional IgG-derived fragments and single-domain nanobodies alike. The attention maps add a second output — a shortlist of interface residues the model weighted heavily — which can steer alanine scanning or mutagenesis at the paratope.
Phys-AbGAT is a small, honestly reported model rather than a scale story, and its contribution is the multi-task framing: it demonstrates that enforcing a thermodynamic relationship between two correlated targets is a usable regulariser when affinity-labelled structures number in the hundreds. The authors are explicit that two of their four baseline comparisons do not reach significance after multiple-testing correction, and correlations near 0.57 on 42 complexes leave substantial unexplained variance. Practical limits follow from the interface it exposes: one heavy chain, one optional light chain and one antigen chain per complex, and a three-dimensional structure as a precondition, which rules out the sequence-only screening that repertoire-scale triage often needs. Code, data splits and the trained checkpoint are released under the MIT License.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.