Structure-based mutational effect prediction from local atomic environments, scoring how substitutions change protein stability and binding affinity.
No providers recorded yet. Browse all providers
HERMES is a family of structure-based models that predict how an amino acid substitution changes a protein's behaviour, working entirely from the atoms packed around the mutated residue. Rather than reading a sequence or an alignment, it takes a 10 Å all-atom neighbourhood, masks the focal residue, and asks which of the 20 amino acids belongs there. The log-likelihood ratio between the mutant and wild-type amino acid becomes a mutational-effect score, which the authors show tracks changes in thermodynamic folding stability and protein-protein binding affinity without any supervision on those measurements.
The model was developed by Gian Marco Visani, Armita Nourmohammad, and colleagues at the University of Washington Paul G. Allen School of Computer Science and Engineering, with collaborators at the Institute for Protein Design, Fred Hutchinson Cancer Center, Yale University, and MIT. It extends the group's earlier H-CNN architecture with a roughly 2.75x inference speedup and, more consequentially, an architecture that can be fine-tuned end to end for arbitrary phenotypes rather than through a bolted-on regression head.
HERMES occupies the structure-based corner of the mutational-effect landscape, alongside inverse-folding models such as ProteinMPNN and the stability-specialised RaSP, Stability-Oracle, and ThermoMPNN. Its distinguishing move is treating side-chain packing explicitly: three inference protocols differ only in how much the local environment is allowed to rearrange around the substitution.
HERMES pretrains on masked amino-acid recovery over ProteinNet's CASP12 set, filtered at 30% sequence identity. Each released model is an ensemble of 10 independently trained networks of 3.5M parameters each, with variants trained both on clean coordinates and with 0.5 Å Gaussian coordinate noise; the noised variants score better zero-shot. Because Rosetta side-chain relaxation runs roughly 66x slower than a fixed forward pass, HERMES-amortized is distilled from relaxed predictions on about 15,000 neighbourhoods, 0.5% of the pretraining set.
On zero-shot stabilising-versus-destabilising classification, packing awareness matters: HERMES-relaxed reaches recall 0.48 against 0.27 for HERMES-fixed, with the largest gains on substitutions between residues of very different size. Fine-tuned on the Stability-Oracle cDNA117k and ThermoMPNN Megascale splits, HERMES outperforms RaSP and matches Stability-Oracle and ThermoMPNN, and remains accurate when given ESMFold structures instead of crystallographic ones. On SKEMPI v2.0 binding effects it reaches Spearman 0.286 zero-shot, comparable to ProteinMPNN but behind Rosetta and FoldX at roughly 0.35; fine-tuned on SKEMPI it reaches Pearson 0.471 under a random split and 0.435 under a homology-aware split that separates interaction classes.
HERMES suits any workflow that needs to triage a large substitution space quickly: saturation-mutagenesis library design, stability engineering of enzymes and biologics, and interpretation of missense variants where a structure is available. The authors demonstrate a vaccine-design pipeline that screens all substitutions with HERMES-amortized and filters the top hits with Rosetta; in the pre-fusion core workflow, 7 of the top 10 suggested mutations were Rosetta-validated, against 4 for ProteinMPNN and 2 at random. A companion study applies the pretrained model to T-cell receptor recognition, predicting TCR-pMHC binding and T-cell activity at up to 0.72 correlation with experiment and designing immunogenic peptides that activated T-cells at success rates up to 50%.
The model's practical appeal is throughput: exhaustive saturation mutagenesis of a whole antigen is cheap enough to run on a single GPU, a scale at which Rosetta screens become impractical. The MIT-licensed repository carries several of the trained ensembles directly in trained_models/ — each seed a multi-megabyte PyTorch checkpoint shipped with its hyperparameters, training log, and loss traces — with the remaining models on Zenodo. A second MIT repository, hermes_experiments, holds the paper's evaluation code, the T2837, Ssym, Megascale, SKEMPI, RaSP, and antigen benchmark tables, and the released prediction matrices, so the reported comparisons can be re-derived rather than taken on trust. The fine-tuning path is documented, so the same architecture can be retargeted to any phenotype for which mutational data exist. Several limitations are worth stating plainly. Scores are computed one site at a time, so synergistic multi-mutation designs are outside the model's reach, and nearby mutated residues bias each other's predictions. Zero-shot binding prediction trails physics-based energy functions. The pretrained models also carry a bias toward size-conserving substitutions, which the amortized and fine-tuned variants substantially reduce. The work remains a preprint and has not been peer reviewed.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.