AI model trained on protein folding predicts how mutations affect binding

The work

Technical · from the record

DGMPNN is a message-passing graph neural network that was pre-trained on an inverse-protein-folding task over PDB protein assemblies and then transfer-learned, with an attention-based graph pooling readout, to predict ln(KD) for two-chain protein complexes, from which ΔΔG between a mutant and the wild type is computed. Under three repeats of 10-fold cross-validation the text reports average Pearson correlations of 0.875 on S1131 and 0.790 on S4169 for ΔΔG, and the model was compared with energy-based and earlier machine-learning predictors including FoldX. A version trained on SARS-CoV-2 BA.1 RBD deep mutational scanning data was then used to score multi-site mutation combinations across the receptor binding motif, and nine selected variants were expressed and assayed; across the six variants that behaved as monomers, predicted and BLI-derived ΔΔG gave a Spearman coefficient of 0.886 and a Pearson coefficient of 0.726.

How AI was used

Protein complexes were converted into directed residue graphs whose edges encode Euclidean distances across 25 backbone and idealised Cβ atom pairs, expanded with 16 Gaussian kernels, plus one-hot relative sequence positions and an intra- versus inter-chain flag, with node features initialised to zeros. A three-layer encoder and three-layer decoder message-passing network, operating over the 48 nearest neighbours by Cα-Cα distance, was pre-trained autoregressively on PDB assemblies deposited as of 2 August 2021 and clustered at 30% identity, recovering masked native residue identities under a teacher-forcing schedule with homologous chains masked. For the downstream task, an additional encoder layer produced multi-head residue-level attention scores conditioned on encoder edge features and decoder node features; the reweighted node features were summed, concatenated with global pooling, and passed to a two-layer MLP trained with mean squared error against experimental ln(KD) from SKEMPI-derived and DMS-derived datasets, with pooling and readout parameters initialised from a Xavier uniform distribution. Evaluation used 10-fold cross-validation, leave-one-complex-out splits, a pre-training set purged of 1R0R homologues, and an external set of published affinity measurements. The fine-tuned model was then run over enumerated multi-site mutation combinations in 21 five-residue patches of the BA.1 receptor binding motif and over a rule-filtered combinatorial library, retaining variants whose predicted ΔG did not exceed that of the wild type; attention scores were also inspected to locate residues contributing to the interface, and nine model-selected variants were taken forward to recombinant expression, BLI and ELISA.