Protein stability predictor that fuses ESM-2 sequence embeddings into the ThermoMPNN structure network to score ΔΔG for single-point mutations.
No providers recorded yet. Browse all providers
A single amino acid substitution can be geometrically unremarkable and still be catastrophic. The side chain fits, the packing is unchanged, nothing clashes — and the variant unfolds at a lower temperature anyway, because the position has been conserved across every homolog for reasons the local environment does not show.
Stability predictors come at this from two sides. Inverse folding models such as ProteinMPNN encode a residue's spatial neighbourhood, and ThermoMPNN, from the Kuhlman lab, adds a light regression head on those frozen embeddings to predict ΔΔG. Protein language models such as ESM-2 encode the opposite view, the evolutionary record of what has been tolerated at that position across millions of sequences, with no structure at all. SA-MPNN, from the High Performance Computing Center at the National Vaccine and Serum Institute and the state key vaccine laboratory of China National Biotec Group, puts both into one network and lets a self-attention layer decide how they interact.
The fusion is the contribution. Concatenating two embeddings and feeding the result to an MLP is the obvious move, and the authors compared several such strategies; what they kept instead treats the structural tokens and the sequence tokens as one long token sequence and runs a transformer encoder over the whole thing, so a structural token can attend to the language model's view of any residue and vice versa. Published in the Journal of Chemical Information and Modeling in 2026, the model reports what the authors call modest but consistent gains over ThermoMPNN across benchmarks, clearest in correlation and screening-oriented metrics — the regime that matters when candidates are being ranked for the bench.
The structural branch uses the pretrained ProteinMPNN v_48_020 encoder; the hidden states of its first two message-passing layers join the residue embedding to give a 384-dimensional token per position. The sequence branch defaults to ESM-2 150M (esm2_t30_150M_UR50D), whose 640-dimensional representations are linearly projected to the same width — ESM-2 8M and 650M and ESM-MSA-1b were also evaluated. Both token sets are layer-normalized, concatenated into a length-2L sequence, and passed through an 8-head, 2-layer transformer encoder. The fused token at the mutated position goes through a light-attention module and a small MLP to 21 amino-acid logits, and ΔΔG is the mutant-minus-wild-type logit difference, which cancels position-specific bias.
Training uses the Megascale dataset of Tsuboyama et al. — hundreds of thousands of proteolysis-derived folding stability measurements — for up to 50 epochs at a learning rate of 3e-4 with early stopping. Evaluation spans held-out Megascale, the homologue-free FireProt split, S669, S2648, S783 and SSYM, plus the melting-temperature sets S571 and S4346. In the AtBgl1A case study, the top 20 predicted variants were synthesized, 13 expressed, and 5 showed improved thermostability on an UNcle platform; the best, GC20, melted at 76.98 °C, 5.97 °C above wild type.
The target use is enzyme and biologic thermostabilization, where the practical question is which 20 of several thousand possible single-point mutants to clone and assay. Because the checkpoint is fixed and inference needs only a structure and a mutation list, it can be pointed at a new protein with no retraining — the AtBgl1A campaign is that workflow carried through to melting curves. The same ranking suits vaccine antigen and industrial enzyme work, where shelf life and process temperature are set by the least stable component.
SA-MPNN belongs to a growing group of stability predictors that build on pretrained structural encoders and language models rather than training from scratch, alongside catalog entries such as HERMES, ProStab and UniStab. Its contribution is narrow and checkable: given the same ThermoMPNN backbone and training set, attention-based fusion of language model features buys a real if limited improvement, concentrated in ranking quality. The prospective β-glucosidase experiment has the longest shelf life — 5 improved variants from 13 expressed is a hit rate a protein engineer can plan around, and it is reported forward rather than retrospectively.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.