AI Model Reads Proteins in Parallel to Predict How Mutations Reshape Stability

A single swapped amino acid can be the difference between a protein that folds into a precise, functioning molecular machine and one that unravels into useless, sometimes dangerous, tangles. Predicting which way that coin will flip has occupied computational biologists for decades, because the energetics of protein folding are governed by an intricate web of interactions spanning just a few neighboring residues to chains of hundreds of positions. Now, a team at Northeast Forestry University in Harbin, China, reports a new deep learning framework that tackles the problem by reading proteins at two scales at once — and by borrowing a trick from thermodynamics to keep its predictions physically sensible. The work, published in BMC Bioinformatics, offers a fresh architectural answer to one of protein engineering’s most persistent computational challenges.

The quantity at the heart of the study is ΔΔG, the change in a protein’s folding free energy caused by a mutation. When ΔΔG is favorable, the mutated protein holds together at least as well as the original; when it is unfavorable, the mutation destabilizes the structure. Getting this number right matters far beyond academic curiosity. Enzyme engineers hunting for catalysts that survive industrial temperatures rely on stability predictions to prioritize which of thousands of possible mutations to synthesize. Clinicians interpreting genomic variants need to know whether a single amino acid substitution in a human protein is likely to sabotage its folding, a frequent mechanism underlying inherited disease. Experimental measurement of ΔΔG is slow and expensive, so accurate computational surrogates have enormous leverage.

Why has prediction remained hard? Because the information that determines stability lives at multiple distances simultaneously. A mutation’s immediate neighborhood — the few residues packed around it in the folded structure — sets the local physicochemical environment: hydrogen bonds broken or formed, hydrophobic pockets filled or emptied, electrostatic contacts gained or lost. Yet the global context matters too. Distant segments of the chain can pack against the mutation site in three-dimensional space, allosteric communication can propagate perturbations far from their origin, and the overall sequence encodes the evolutionary constraints that keep the fold viable. A model that sees only the local scene misses long-range consequences; one that sees only the global picture blurs the fine chemical detail where many decisive interactions live.

The new framework, developed by Yuming Zhao, Ziqi Liu, Benzhi Dong, and Dali Xu, confronts this two-scale problem with an explicitly parallel design built on top of ESM-2, a large protein language model trained on billions of evolutionary sequences. ESM-2 converts each amino acid sequence into a rich numerical embedding in which positions carry contextual information learned from the statistics of natural proteins. From these embeddings, the researchers compute a difference sequence — a position-by-position numerical portrait of what the mutation changes. That difference representation then feeds two independent processing branches running side by side.

The first branch is a convolutional neural network, a family of architectures descended from image-recognition research. Convolutional filters slide across the sequence and respond only to a finite window of positions, making them naturally suited to capturing short-range motifs: the local pattern of chemical-property changes surrounding the mutation. The second branch is a Transformer, the architecture behind modern large language models, whose self-attention mechanism lets every position weigh information from every other position in the sequence. This gives the model unrestricted reach for sequence-wide dependencies, the long-range couplings that convolutions cannot span. Crucially, the two branches do not process the data one after the other; they operate simultaneously, and a gated fusion mechanism — a learned weighting layer — decides adaptively how much of each representation to blend into the final prediction for any given input.

The authors frame the architectural question precisely: is there complementary value in modeling short-range patterns and long-range dependencies independently and in parallel, rather than successively within a single pathway? Their ablation experiments address this directly. Introducing the serial version of the architecture provides the largest gain over raw ESM-2 embeddings alone, confirming that structured local-global processing is the core of the method’s power. Switching from serial to parallel processing then yields additional improvements across most evaluated metrics. In other words, allowing the local and global views to be formed without contaminating each other — and then fused on learned terms — extracts information that a strictly sequential pipeline leaves on the table.

Two further ingredients push the model toward physical realism. First, explicit physicochemical difference features encode how the wild-type and mutant residues differ in fundamental properties such as size, charge, and hydrophobicity, giving the network direct descriptors of mutation-type heterogeneity rather than forcing it to rediscover basic chemistry from embeddings. Second, the researchers confronted a subtle logical property of ΔΔG itself: antisymmetry. The stability change of mutating residue A to B is, in principle, the mirror image of mutating B back to A. Deep networks do not respect such symmetries unless they are built in or encouraged. The team used reverse-mutation data augmentation, training on both directions of each mutation, and embedded a transitivity constraint in the joint loss function — a thermodynamic regularization that ties predictions together so the model’s outputs obey the consistency rules that real free-energy cycles demand.

The performance evaluation rested on two widely used benchmarks. S669 is the standard single-mutation dataset for ΔΔG prediction, while PTmul-NR tests generalization to multiple mutations and other, harder regimes. On both, the parallel framework achieved results comparable to state-of-the-art sequence-based methods. On S669 specifically, the complete model delivered numerical improvements over the serial baseline in the three canonical metrics: the Pearson correlation coefficient, which measures how well predictions track experimental values; the root-mean-square error; and the mean absolute error. Notably, the reduction in mean absolute error remained statistically significant even after correction for multiple comparisons — a rigorous standard that many benchmark studies skip — and the parallel design incurred only limited additional computational cost, an important practical consideration for labs scanning large mutational libraries.

The significance of the work extends beyond its leaderboard position. It contributes an architectural lesson to the fast-growing field of protein language model applications: as pretrained embeddings become commodities, the differentiating factor is increasingly the downstream network that interprets them. Demonstrating that parallel local-global extraction with gated fusion beats serial processing — and that thermodynamically motivated constraints measurably improve consistency — gives other developers concrete design principles. It also aligns with a broader trend of injecting physical prior knowledge into machine learning models, a strategy that tends to improve generalization on exactly the sparse, noisy datasets that characterize experimental biophysics.

There remain, of course, honest limits. Sequence-based models infer stability without explicit three-dimensional structures, so their mechanistic interpretability is indirect, and benchmark datasets like S669 carry well-documented biases, including errors and overrepresentation of certain protein families, that cap achievable accuracy for everyone in the field. The authors are careful to describe their method as comparable to, rather than decisively surpassing, the strongest competitors — with the statistically robust gain in mean absolute error as the clearest architectural dividend. Still, for protein engineers hoping to design thermostable enzymes and for geneticists triaging which variants in a patient’s genome are most likely to destabilize a protein, the message is encouraging: reading a mutation’s local neighborhood and its global context simultaneously, disciplined by the laws of thermodynamics, squeezes genuinely better predictions out of the same underlying biological information. As protein language models continue to improve, frameworks like this one — architected to respect both the chemistry of the immediate environment and the physics of the whole fold — are positioned to ride that rising tide directly into the lab.

Subject of Research: Prediction of mutation-induced protein stability changes using parallel local and global feature modeling on protein language model embeddings

Article Title: Parallel local–global feature modeling with thermodynamic regularization for predicting protein stability changes

Article References: Zhao, Y., Liu, Z., Dong, B., & Xu, D. (2026). Parallel local–global feature modeling with thermodynamic regularization for predicting protein stability changes. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06669-9

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06669-9

Keywords: protein stability, ΔΔG prediction, ESM-2, deep learning, Transformer, convolutional neural network, gated fusion, thermodynamic regularization, protein engineering, bioinformatics, S669 benchmark, machine learning