HERMES Flashcards (IPD reading)

Visani et al. bioRxiv v4 2026, equivariant net for stability + antigen stabilization. TRAP cards = where the obvious summary is wrong. last 3 cards are figures to explain.

what does HERMES do in one sentence?

a fast SO(3) equivariant network that looks at the ball of atoms within 10 A of one residue and scores every possible amino acid there, which turns into a zero shot prediction of whether a mutation stabilizes or destabilizes, including which mutations lock a viral fusion protein in its prefusion shape

who wrote HERMES and where does it live?

Gian Marco Visani (first author, UW Allen School) with Zac Jones (UW Biochemistry + IPD, the only IPD affiliated author) and senior author Armita Nourmohammad. bioRxiv preprint, v4 January 2026, not journal published. code at github StatPhysBio/hermes

why does the local angle matter for HERMES?

Visani is also first author on the critique of MULTI-evolve, and Zac Jones is IPD. the methodological skeptic in the reading list is from down the hall

what exactly is HERMES's input?

a neighborhood: every atom within 10 A of the C-alpha of one focal residue, with the focal residue's identity masked. each atom carries element type (hydrogens included), partial charge and solvent accessible surface area

what is the holographic encoding in HERMES?

the atom point cloud is projected onto a 3D Zernike Fourier basis, spherical harmonics for the angular part times radial Zernike polynomials, keeping degrees up to l = 5 and the first 11 radial frequencies. rotations act simply in that basis, so the network never has to learn rotation invariance from data

what does SO(3) equivariant mean and why build it in?

rotate the input and the internal features rotate with it, so the answer does not change. a protein has no canonical orientation, so equivariance stops the network wasting capacity relearning the same pocket in a thousand orientations. implemented with e3nn primitives

how does HERMES score a mutation?

log p(mutant) minus log p(wild type) in the same structural context

what does HERMES get for free from scoring as a log ratio?

thermodynamic reversibility and path independence. Stability-Oracle enforces the same property with 19x thermodynamic permutation data augmentation

what is HERMES pretrained on?

masked amino acid prediction on ProteinNet CASP12 filtered at 30% identity: 10,957 train, 2,730 validation, 212 test structures, 10 epochs, ~40 min per epoch on one A40. each released model is an ensemble of 10 networks of ~3.5M parameters

name the three HERMES protocols

fixed (score both residues in the unchanged wild type neighborhood), relaxed (thread the mutation in and repack side chains within 12 A with PyRosetta FastRelax), and amortized (distill relaxed into fixed by fine tuning on relaxed predictions)

what is the rigid wild type problem?

if you score every mutation in the unchanged wild type structure, the model penalizes anything that changes side chain size, because the cavity is shaped like the wild type residue. it cannot see that carving out a void can help, or that neighbors could shift to fit something bulkier

how much does relaxation help, and what does it cost?

recall for stabilizing mutations goes from 0.27 (fixed) to 0.48 (relaxed), p < 0.01, with slight precision loss and higher F1. it costs about 66x more compute

what is amortization in HERMES and why is it the transferable idea?

run the slow relaxed protocol once on ~15k neighborhoods, only 0.5% of pretraining sites, then train the fast model to imitate its own relaxed answers. you get most of the benefit at fixed speed. the flexibility becomes implicit in the weights

where is amortization weak?

small to large substitutions. gains are comparable for large to small but much weaker small to large, which the authors blame on stabilizing small to large mutations being underrepresented in the fine tuning data

the HERMES antigen benchmark, in numbers

33 previously reported stabilizing mutations across five antigens: influenza HA 3, RSV-F 7, hMPV-F 11, DENV-E 8, SARS-CoV-2 spike 4. HERMES-amortized ranked 24 above wild type and 19 in the top 3 at their site

which mechanism does HERMES recover best and worst?

best is prolines, 7 of 8 stabilizing prolines recovered; all models beat BLOSUM62 on proline and electrostatic mutations. worst is cavity filling, and synergistic sets (RSV-F TriC, the DENV-E cation-pi pair) are nearly unrecoverable for every method

TRAP: how did HERMES compare to Rosetta on antigens?

Rosetta performed on par with HERMES-amortized, and 17 of the 33 benchmark variants were originally found by Rosetta based screening, which biases the benchmark toward Rosetta. the argument for HERMES is cost and throughput, not accuracy

why doesn't stability fine tuning transfer to antigen stabilization?

Megascale style data is enriched for small compact domains, while antigens are large, multi domain and conformationally heterogeneous, and their stabilizing mutations often act through quaternary contacts or prefusion specific constraints

what is the locality tradeoff the authors name?

verbatim, locality is "a double-edged sword": local neighborhoods cut input complexity, remove size limits and scale well, but necessarily limit the model's ability to capture long range epistasis and allosteric coupling. they call HERMES a hypothesis generation tool

what is wild type preference bias in HERMES?

on Ssym, which has structures for both wild type and mutant, zero shot models predict forward effects better than reverse, driven by elevated log p(wild type). the authors suggest partial memorization from pretraining, and call it a fundamental tradeoff in current training paradigms

how does HERMES do on binding affinity?

on SKEMPI v2.0 (5,713 measurements over 331 complexes) zero shot HERMES-fixed reaches Spearman rho = 0.286, comparable to ProteinMPNN, while Rosetta and FoldX sit around 0.35. fine tuned it is competitive with RDE-Network and MIF-Network but trails Pythia-PPI

the inversion worth noticing in the binding results

for binding, HERMES-fixed correlates better than HERMES-amortized, the opposite of the stability result

what does the substitution matrix analysis show?

HERMES-fixed's learned preferences correlate most with steric properties, buried residue volume Spearman 0.66 and van der Waals volume 0.64. BLOSUM62, being evolutionary, preserves hydrophobicity instead. after stability fine tuning and restricted to core residues, HERMES shifts toward hydrophobicity too

benchmark leakage the authors flag themselves

the Megascale train split is not de-duplicated against T2837, and about 5% of T2837 proteins have >90% similar homologs in the Megascale training set

why is the antigen benchmark weak evidence even when it works?

33 mutations across 5 antigens is tiny, and it is recall only: the original papers rarely tested alternatives at the same site, so there is no measure of false positives or of novel proposals. it also scores only relative ranks at a site

what is Megascale?

the cDNA display proteolysis dataset of folding stabilities on small domains of 40-72 residues, split here ~216k train / ~28k test at 25% identity. it is the training fuel for basically every modern ML stability predictor

explain this figure: HERMES fig 3 (size change analysis)

a 3x3 grid: columns are small to large, similar sizes, large to small; rows are precision, recall and F1. the bars compare HERMES-fixed, relaxed, amortized, ProteinMPNN, Megascale fine tuned variants and ThermoMPNN. the story is the small to large recall column, where HERMES-fixed is nearly zero and relaxation lifts it, and the large to small column where relaxed and amortized gain most. the legend at the bottom right defines small GASCTPD, medium NVEQLI, large MHKFRYW

explain this figure: HERMES fig 7 (antigen stabilization benchmark)

panel A shows the four mechanism classes with real examples: electrostatic (V449D forming a new contact), proline (F817P capping a helix), cavity filling (F190 packing into a void, shown with mesh), and synergistic (W259 and R262 cation-pi across two helices). panel B counts correctly suggested stabilizing mutations per mechanism for BLOSUM62, Rosetta, ProteinMPNN, HERMES-amortized and HERMES-fixed+Megascale, hatched by rank (r_mt

explain this figure: HERMES fig 2 (stability prediction)

the stabilizing versus destabilizing classification and correlation metrics across the HERMES protocols and the baselines (ProteinMPNN, ThermoMPNN, Stability-Oracle, RaSP) on the standard splits. the takeaway is that fine tuned HERMES beats RaSP and matches Stability-Oracle and ThermoMPNN, and that pretraining is load bearing, since training only for stability performed poorly even after cutting capacity from 3.5M to 50k parameters