Neural networks trained on atomic surroundings score how mutations change proteins

The work

Technical · from the record

HERMES is a family of SO(3)-equivariant all-atom neural networks that score amino acid substitutions from the atomic neighborhood within 10 Å of a residue. The networks were pre-trained to recover the identity of a masked residue on ProteinNet CASP12 structures, then fine-tuned either on predictions computed with Rosetta-relaxed mutant neighborhoods or on measured stability and binding effects. On the benchmark splits used by RaSP, Stability-Oracle and ThermoMPNN, the fine-tuned models matched Stability-Oracle and ThermoMPNN; on SKEMPI v2.0 the fine-tuned models were competitive with RDE-Network and MIF-Network and trailed Pythia-PPI. On 33 previously reported antigen-stabilizing mutations across five viral antigens, HERMES-amortized gave a better rank than the wild type to 24 of them, 19 of which ranked within the top three substitutions at their site; the paper also reports that pre-trained models assign elevated probabilities to wild-type residues in wild-type structural contexts.

How AI was used

Protein structures were pre-processed with either a PyRosetta or a Biopython pipeline to add hydrogens, partial charges and solvent-accessible surface areas, and each residue's surrounding atoms within 10 Å were projected onto a Zernike Fourier basis to form a holographic encoding, optionally with 0.50 Å Gaussian coordinate noise. SO(3)-equivariant layers followed by a multilayer perceptron map that encoding to 20 amino-acid logits, and the network was pre-trained on the masked focal-residue prediction task over ProteinNet CASP12 splits, with each model an ensemble of ten independently trained networks of about 3.5M parameters. Mutational effects were scored as the difference between mutant and wild-type logits, either on the wild-type structure alone (HERMES-fixed) or with the mutant term evaluated on a PyRosetta FastRelax side-chain-relaxed mutant neighborhood (HERMES-relaxed). HERMES-amortized was obtained by fine-tuning the network so that fixed-protocol scores regress onto HERMES-relaxed scores over a sampled subset of pre-training sites. Separately, the models were fine-tuned end-to-end under a Huber loss on Rosetta-derived and experimental ΔΔG datasets and on SKEMPI v2.0 binding labels with random, hold-out-protein and hold-out-type splits. The models were then run over all 20 substitutions at sites in multimeric viral antigen structures to rank candidate stabilizing mutations, alongside ProteinMPNN, ThermoMPNN, Rosetta and BLOSUM62 as comparisons, with ESMFold-predicted structures used in a separate robustness analysis.