ProStab

Predicting the change in folding free energy (ΔΔG) caused by a point mutation is the workhorse calculation of protein engineering: it decides which of thousands of possible substitutions are worth expressing and assaying. Physics-based methods such as FoldX and Rosetta are constrained by their force-field approximations, while data-driven predictors have historically been trained on small, alanine-scanning-biased datasets like S2648 (2,648 measurements) and generalize poorly outside them.

ProStab, released in August 2025 by a group at Shanghai Jiao Tong University, attacks the generalization problem by combining two complementary pretrained representations rather than scaling a single one. It reads the wild-type sequence, the mutant sequence, and the wild-type backbone structure, drawing sequence features from ESM-2 and geometric features from the inverse-folding model ProteinMPNN. The model is trained once on the Megascale cDNA-display proteolysis dataset and then applied without retraining, including through a public web server.

What separates ProStab from most ΔΔG papers is that its predictions were carried into the lab. The authors selected the top 20 predicted stabilizing mutations for the amine transaminase from Exophiala xenobiotica, expressed them in E. coli, and measured thermostability — an end-to-end test of whether the ranking is actually usable for enzyme engineering.

#Key Features

• Two pretrained representations, jointly fused: Mutation-specific signal comes from the difference between wild-type and mutant language-model embeddings at the substitution site; site-specific priors come from inverse-folding features on the wild-type backbone.
• Trained on the largest available stability dataset: The Megascale corpus supplies roughly 900–1,000 mutants per protein, versus as few as 7 per protein in conventional deep mutational scanning sets.
• Tested under deliberate distribution shift: All ten evaluation sets share under 25% sequence identity with the training split, and several span far wider ΔΔG ranges than Megascale itself.
• Experimentally validated ranking: Four of 16 solubly expressed transaminase variants showed markedly improved thermostability, and the mutation A245I retained 99.6% of catalytic activity after thermal challenge against 27.3% for wild type.
• Works from predicted structures: Substituting Chai-1 models for crystal structures across the 94 S669 proteins changes performance only marginally, so the method does not require experimental coordinates.

#Technical Details

ProStab is an end-to-end network with three inputs. ESM2-650M, held frozen, encodes both wild-type and mutant sequences; the token representations at the mutation site are subtracted to give a mutation-difference feature. A structure encoder adapted from ProteinMPNN derives residue-level geometric features from wild-type backbone coordinates. Sequence and structure features are concatenated, passed through a transformer encoder to produce a fused representation, and decoded by an MLP to a scalar ΔΔG. Training follows the SPURS protocol for splitting Megascale, uses AdamW (learning rate 1 × 10⁻⁴, β = [0.9, 0.98], weight decay 0.01) with MSE loss for up to 50 epochs, one protein per batch, on a single RTX 4090; model selection is by validation Spearman correlation with early stopping.

Against physics-based baselines (FoldX, Rosetta) and deep-learning baselines (SPURS, ThermoMPNN, StabilityOracle, Pythia) across ten independent test sets, ProStab attains the best Spearman correlation on 7 of 10 and is competitive on the rest, with a substantial margin on the widely used S669 set. It matches or beats ThermoMPNN and SPURS on area under the precision–recall curve across all ten benchmarks, reaching 0.635 on Megascale, 0.629 on S2648, and 0.586 on S8754. On the ΔTm sets S571 and S4346 it preserves ranking (Spearman 0.463 and 0.629) even though absolute predictions compress badly: Megascale spans roughly −3 to 5 kcal/mol, so extrapolation beyond that dynamic range is where the model is weakest.

#Applications

The primary use is prioritizing stabilizing substitutions for enzyme and biologic engineering, where thermostability determines process viability. The transaminase campaign is the template: score all single mutants, take the top-ranked surface-exposed positions distal from the active site, and express a couple of dozen variants instead of screening thousands. A public web server at the authors' institution accepts a structure and mutation list, putting the model within reach of wet-lab groups without GPU infrastructure, and molecular dynamics of the best variant traced its gain to reduced backbone fluctuation around the mutation site and a coupled distal region.

#Impact

ProStab is a clean demonstration that the bottleneck in ΔΔG prediction is representation and data breadth rather than model size, and that inverse-folding geometry contributes signal a sequence model alone does not capture. Its wet-lab follow-through is unusual for the genre and makes the ranking claim concrete rather than benchmark-bound. The work builds openly on the SPURS framework, reusing its training pipeline, dataset splits, and metrics. Code and datasets are public on GitHub, but the repository carries no license file, which restricts reuse; pretrained weights are distributed separately via a cloud drive link rather than bundled. The paper is a preprint and has not yet been peer reviewed.