UniStab

Insert a residue into a protein and the backbone has to move; delete one and it moves again. Most stability predictors cannot represent that. They score a substitution against a fixed wild-type backbone, or they add up the effects of individual substitutions and call the sum the effect of the combination — an assumption that neither survives an indel nor captures the epistasis making two mutations together behave unlike either alone. The alternative, generating an explicit mutant structure per candidate, is far too slow to screen with.

UniStab, from Hong Tan, Shenggeng Lin and Yi Xiong at Shanghai Jiao Tong University and the Shanghai Innovation Institute, takes a third route: it reads the latent representations a folding model builds on the way to a structure and never decodes them into coordinates. Wild-type and mutant sequences each pass through the folding trunk of ESMFold — 48 Evoformer-style blocks that jointly refine a per-residue sequence representation, seeded from ESM-2 embeddings, and a pairwise one — and the trunk output becomes the feature. Because those representations were optimized in service of structure prediction, they already encode the geometric constraints and residue couplings that a backbone rearrangement disturbs.

The pretrained weights stay frozen; only low-rank adaptation (LoRA) updates inside the trunk and a small readout on top are trained. The result is one checkpoint covering single-point substitutions, multi-point substitutions and indels in a single sequence-to-ΔΔG pass. Published in Chemical Science in 2026, it extends the same group's earlier ProStab, which scores single substitutions against a fixed backbone.

#Key Features

• One model across every mutation type: Single-point, multi-point and indel variants go through the same network, instead of being split across predictors that each cover one regime.
• Implicit structure, no coordinate generation: The trunk's sequence and pair representations are consumed as features, so backbone plasticity is modelled without predicting a mutant structure per candidate.
• Parameter-efficient adaptation: ESM2-3B and the ESMFold trunk stay frozen; LoRA updates (rank 8, α 16) on the trunk's projection and MLP layers supply the task-specific capacity, plus a three-block fusion module and MLP head.
• Attention that localizes the perturbation: Wild-type-versus-mutant attention difference maps concentrate on the mutated positions and on neighbours whose heavy-atom contacts change.
• Beam search for stabilizing design: The ΔΔG score doubles as a design objective — UniStab–Beam proposed three double-substitution protein-glutaminase variants, each backed by 100 ns of molecular dynamics.

#Technical Details

Training used MegaScale, the cDNA-display proteolysis corpus of over 776,000 measurements, filtered to 470,995 mutations and split at the level of 25%-identity sequence clusters into 373,325 training, 43,070 validation and 54,600 test samples, minimizing MAE with AdamW at learning rate 1 × 10⁻⁴. On the MegaScale test split, Spearman correlation reaches 0.780 on indels, 0.668 on single-point mutations and 0.595 on double mutations, where over 95% of variants are destabilizing. Against external benchmarks the model ranks first on three of four: it leads on MegaDouble (0.595 versus 0.584 for SPURS) and on the 339-sample PTMUL multi-mutant set (0.713 versus 0.600 for MutateEverything), while ThermoMPNN-D — an inverse-folding method built on ProteinMPNN with explicit inter-residue distance features — keeps a slight edge on PTMUL doubles at 0.566 against 0.555.

Feature-space distance between wild-type and mutant embeddings grows with mutation count and correlates with local RMSD from mutant structures predicted by Chai-1 — Spearman 0.209 within an 8 Å radius, against 0.029 for a sequence-only ESM2-3B baseline. Ablations show the fusion module is load-bearing: removing it drops single-mutant Spearman from 0.668 to 0.466, and halving the LoRA rank drops indels from 0.780 to 0.629.

#Applications

The immediate use is triage in protein engineering — ranking candidate variants of an enzyme or biologic before anything is expressed. Coverage of indels and higher-order combinations matters because campaigns rarely stop at one substitution, and the beam-search mode turns ranking into proposal: the protein-glutaminase run converged on L1M + A129C, L1M + S133I and L1M + S167C, all showing reduced backbone fluctuation and higher retention of native contacts in simulation. Alongside each score, the attention maps name the contacts a candidate is predicted to disturb.

#Impact

UniStab's contribution is the demonstration that a folding model's intermediate representations work as a general-purpose structural feature space, not merely as a step toward coordinates — a pattern that transfers to any task where backbone response matters but atomic output does not. The authors state the limits plainly: MegaScale ground truth is largely confined to roughly −2 to +5 kcal/mol, and below −2 every method tested, UniStab included, shows negative rank correlation on PTMUL; latent features stay coarser than explicit geometry, which shows up as weaker performance on distant double mutations and buried core–core pairs. Code and the processed benchmark splits are on GitHub, with weights distributed through a cloud drive link. The README declares an MIT license in prose, but the repository carries no license file, leaving reuse terms less settled than they appear.