ProtSSN

A substitution buried in a protein core often kills the enzyme not by breaking any motif a language model would recognise, but by disturbing packing a few angstroms away — a steric clash, a lost hydrogen bond, a helix that no longer resists unfolding when heated. A sequence-only scorer reads it as a mildly unusual token in an otherwise ordinary context and rates it accordingly; the assay reads it as dead. ProtSSN's answer is to hand the scorer the local geometry outright, as a graph over each residue's spatial neighbours, instead of hoping attention over the linear chain will recover it.

The model is a funnel of two encoders. A frozen ESM-2 650M language model reads the full amino acid chain and emits a per-residue embedding; those embeddings become node attributes on a k-nearest-neighbour graph built over C-alpha coordinates, which a six-layer roto-translation equivariant graph neural network then refines. Equivariance means a rotated or translated structure yields the same representation; only the graph half is trained; and no alignment is needed at training or inference.

ProtSSN was built by Yang Tan, Bingxin Zhou, Lirong Zheng, Guisheng Fan and Liang Hong at Shanghai Jiao Tong University, East China University of Science and Technology and the Shanghai AI Laboratory, posted to bioRxiv in December 2023 and published in eLife in 2025. It is also distributed as Venus-ProtSSN inside the group's Venus line, alongside ProSST and Venus-TIGER.

#Key Features

• Structure-aware scoring without alignments: variants are scored from a single wild-type structure and its sequence, with no homolog search, so shallow families are handled the same as well-studied ones.
• Denoising pretraining that imitates mutation: rather than masking tokens, training perturbs 5% of residues per epoch and re-encodes the mutated chain before asking the graph network to recover the original identities — a "mutate-then-recode" objective that puts the model in the position of a wet-lab mutagenesis experiment.
• Equivariance as an ablated design choice: swapping the equivariant layers for standard GCN or GAT convolutions measurably degrades prediction, which is what justifies carrying coordinates through the network rather than distances alone.
• Thermostability benchmarks with conditions attached: the authors released DTm and DDG, two single-site benchmarks grouped by protein-condition pair, because ProteinGym collapses stability assays into one class and discards the pH and temperature they were measured at.
• Nine released checkpoints, usable singly or as an ensemble: full-ensemble inference fits on a single 10 GB GPU, and the same checkpoints emit frozen embeddings or seed supervised fine-tuning.

#Technical Details

Pretraining runs once over CATH v4.3.0: 30,948 experimental structures filtered to under 40% sequence identity, with chains longer than 2,000 residues dropped. Node features come from a frozen ESM2-t33 prefix model; graphs use k nearest C-alpha neighbours for k in {10, 20, 30}, and the six equivariant layers carry 512, 768 or 1,280 hidden units, giving nine configurations of 148M, 160M and 184M trainable parameters respectively. Adam at a 1e-4 learning rate optimizes batches capped at 8,192 residue tokens. Deeper stacks of 12 or 18 layers did not help, and a 3B ESM-2 prefix scored worse than the 650M one.

On the paper's benchmarks the recommended k20_h512 checkpoint reaches Spearman correlations of 0.419 on DTm, 0.442 on DDG and 0.444 overall on ProteinGym v1; the nine-model ensemble reaches 0.425, 0.440 and 0.451. ESM-2 650M alone scores 0.330, 0.392 and 0.419 on the same three. Substituting ESMFold structures for AlphaFold 2 ones costs ProtSSN three to five times less accuracy than it costs the other structure-conditioned baselines tested.

#Applications

The practical use is triage before the bench: given a wild-type sequence and a structure from the AlphaFold Protein Structure Database or folded on demand, the model scores every single-site substitution and ranks the shortlist a directed-evolution campaign will actually build. That spans the properties enzyme engineers weigh together: catalysis and binding as measured in ProteinGym, and melting-temperature and folding-free-energy shifts in DTm and DDG. The checkpoints also export fused sequence-plus-structure embeddings for downstream classifiers and can be fine-tuned on labelled property data.

#Impact

ProtSSN demonstrates that a modest equivariant module on a frozen language model buys much of what structure conditioning offers, at a fraction of the cost of training a sequence-structure model jointly — and its DTm and DDG benchmarks give the field condition-controlled thermostability evaluations it previously lacked. The eLife assessment rates the findings valuable and the evidence compelling; SaProt edges it on ProteinGym v1 overall. Two constraints shape adoption: a structure is required, so this is not a sequence-only drop-in, and the code repository carries a CC BY-NC-ND 4.0 license ruling out commercial use and derivative works, even though the checkpoint repository is labelled Apache 2.0. The model is now folded into the group's VenusFactory2 platform.