Fig. 1: Graphinity architecture and synthetic dataset preparation.

a, The EGNN deep learning models are trained on graphs of three-dimensional protein structure coordinates. The graphs are built from atoms in the neighborhood of the mutated site, conceptually illustrated with circles (nodes, atoms) and connecting lines (edges, interactions). The antibody is shown in purple, the antigen in blue and inter-binding partner edges in orange. Our model architecture consists of three E(n) EGC layers²³, followed by a linear layer. For ΔΔG prediction, the embeddings generated from the E(n) EGC layers for the WT and mutant complex are subtracted from one another before passing through the linear layer. b, The synthetic ΔΔG datasets were generated from structurally resolved complexes from SAbDab^26,27. We mutated interface residues and predicted ΔΔG values using FoldX¹¹ and Rosetta Flex ddG¹². c, An example of ΔΔG data for a complex. PDB: 1XGP⁵⁶; antibody in purple, antigen in blue; affinity values from SKEMPI 2.0³¹. Mut, mutant.