DPStab

Predicting how a single amino acid substitution shifts a protein's folding free energy (ddG) or melting temperature (dTm) is routine in protein engineering, and the accurate methods have historically needed a structure. Force-field tools such as FoldX and learned predictors such as ThermoMPNN — which transfers features from ProteinMPNN — read the three-dimensional environment of the mutated residue directly. Sequence-only methods scale far better across large mutant libraries, but have lagged behind because a protein language model trained on single sequences has no explicit view of the contacts a substitution disrupts.

DPStab, from the Yang Zhang lab at the National University of Singapore with collaborators at Central South University, closes that gap without requiring a structure. It initializes its sequence encoder from ESM-2 and, crucially, reads out the model's attention maps as predicted residue contacts. A neighboring encoder then applies cross-attention over the residues in contact with the mutated position, so the network sees the local packing environment inferred from sequence alone rather than ignoring it.

Two training strategies address the well-known pathologies of stability datasets. Experimental ddG measurements are heavily skewed toward destabilizing mutations, and the true quantity is antisymmetric — the forward and reverse substitutions must sum to zero. DPStab trains one model on direct mutations and a second on symmetrized data under an explicit antisymmetric constraint, then uses their averaged predictions as pseudo-labels to self-distill a single final model.

#Key Features

• No structure required: Contact information is inferred from ESM-2 attention maps rather than read from a PDB entry or a predicted model, so the method applies to proteins with no experimental or reliable predicted structure.
• Antisymmetric constraint: A second teacher trained on symmetrized data enforces the physical requirement that ddG for wild-type to mutant equals the negative of the reverse, correcting a bias that inflates apparent accuracy in most benchmarks.
• Self-distillation at inference: Averaged predictions from the two teachers become pseudo-labels for a final student model, trained on the combination of labeled data and pseudo-labeled test data.
• Both stability endpoints: Separate heads predict ddG (conditioned on pH and temperature) and dTm (conditioned on pH), covering the two measurements protein engineers most often need.

#Technical Details

The sequential encoder reproduces the ESM-2 architecture — one token embedding layer, 33 stacked transformer layers, and a linear head over concatenated attention maps for contact prediction — initialized from pretrained ESM-2 weights with a hidden dimension of 1,280. Transformer weights are frozen for the first two epochs and then fine-tuned with AdamW under a decaying learning rate. Training uses S8754 (8,754 single-point mutations across 301 proteins with pH and temperature) for ddG and S4346 (4,346 mutations across 358 proteins with pH) for dTm, both drawn from ProThermDB and ThermoMutDB; symmetrization expands these to 17,235 and 8,692 samples.

On S461, an independent test set of 461 mutations across 48 proteins held below 25% sequence identity to training, DPStab reaches Spearman 0.86 and Pearson 0.84 (versus 0.83 and 0.82 for the best prior method), with RMSE under 0.93 kcal/mol and MAE under 0.68 kcal/mol. Classifying mutations as stabilizing, neutral, or destabilizing, it scores MCC and Cohen's kappa above 0.61 against 0.51 for the next-best method, GeoDDG-3D. For dTm on S571 (571 mutations across 39 proteins) it reaches 0.56 accuracy against 0.52 and 0.53 for the two GeoDTm variants. Per-protein evaluation across 15 proteins with at least 6 mutations each gives median gains of 9.1% in Spearman and 19.4% in accuracy over GeoDDG-3D, and on deep mutational scanning data DPStab improves Spearman by 13% and Kendall tau by 15% over the same baseline.

#Applications

The natural use is triaging mutant libraries in protein engineering and enzyme stabilization, where a ranked list of candidate substitutions determines what gets cloned and assayed. Because DPStab needs only a sequence, it extends to targets without structures and to proteome-scale screens that structure-based tools cannot cover. Validation on cDNA display proteolysis data — 389,068 inferred ddG values across 412 proteins — and on deep mutational scanning cohorts shows the model ranks mutations within a single protein reliably, the regime that matters for directed evolution. Analysis of the learned contacts also identifies which residue interactions a mutation disrupts, adding a mechanistic handle to the numeric prediction.

#Impact

DPStab is a preprint and has not yet been peer reviewed. Code is released under an MIT license, with trained weights distributed through a Google Drive link from the repository. Its main contribution is showing that sequence-only stability prediction can match or exceed structure-based predictors when contact information is recovered from language-model attention and when the antisymmetry of mutation effects is enforced during training rather than assumed. The scope is narrow by design: it predicts stability change for single substitutions only, and does not address multi-point mutants, binding, or activity.