StaB-ddG

Predicting how a point mutation changes the strength of a protein-protein interaction — the binding free energy change, ΔΔG — is a core problem in antibody engineering, therapeutic design, and interface biology. For years the awkward situation in this field was that deep learning lost to physics: empirical force-field methods such as FoldX and Rosetta's Flex ddG remained more accurate than neural predictors, because experimental binding measurements are scarce (a few thousand mutations across a few hundred complexes) and models trained directly on them overfit.

StaB-ddG, developed at Stanford University and presented at ICML 2025, closes that gap by changing what the network is asked to learn. Instead of regressing binding energy directly, it exploits the thermodynamic identity ΔG_bind(A:B) = ΔG_fold(A:B) − ΔG_fold(A) − ΔG_fold(B): binding energy is the folding energy of the complex minus the folding energies of its isolated partners. That reframing lets the model borrow from two much larger data sources — the sequence likelihoods of a pretrained inverse-folding network, and hundreds of thousands of folding-stability measurements — before it ever sees binding data.

The folding-energy proxy is the log-likelihood of an inverse-folding model, ProteinMPNN, evaluated on the complex and on each partner separately. Applied zero-shot this already gives usable signal; two stages of fine-tuning turn it into a predictor that matches the accuracy of FoldX while running roughly 1,000 times faster.

#Key Features

• Thermodynamic decomposition: Binding ΔΔG is computed as a difference of folding energies rather than predicted end-to-end, so abundant stability data can be used to train a binding model.
• Two-stage transfer learning: The inverse-folding backbone is fine-tuned first on 776,298 folding-stability measurements from the Megascale dataset, then on 7,085 binding ΔΔG measurements from SKEMPI v2.
• Force-field-level accuracy at neural speed: 0.2 seconds per mutation on a GPU versus roughly 210 seconds per mutation for FoldX.
• Multi-chain and multi-point mutations: Because it inherits ProteinMPNN's multi-chain conditioning, it handles complexes of several chains and mutations at several positions at once, which many stability predictors cannot.
• Variance-reduced estimator: Predictions average 20 forward passes that share the same decoding permutation and backbone noise between wild type and mutant, cancelling much of the sampling noise in the likelihood estimate.

#Technical Details

The backbone is the ProteinMPNN message-passing graph neural network, which encodes backbone geometry and autoregressively models sequence likelihood under a randomized decoding order with Gaussian noise added to backbone coordinates. Stage one minimizes squared error between predicted and measured folding ΔΔG on the Megascale dataset of 776,298 measurements across 412 domains (Adam, learning rate 3e-5, 70 epochs). Stage two applies the binding decomposition and fine-tunes on SKEMPI v2, which contains 7,085 binding ΔΔG measurements across 345 complexes; after filtering, 121 complexes are used for fine-tuning and 80 held out for testing (Adam, learning rate 1e-6, 200 epochs).

On the held-out SKEMPI v2 split, StaB-ddG reaches a per-interface Spearman correlation of 0.45, statistically indistinguishable from FoldX and Flex ddG and above previously published deep learning predictors. Ensembling StaB-ddG with FoldX raises per-interface Spearman to 0.53. Accuracy is stratified by interface rigidity: RMSE is 1.12 kcal/mol for rigid interfaces with under 30% loop content and degrades to 1.50 kcal/mol for the most flexible ones — a direct consequence of the model holding the backbone fixed rather than relaxing it after mutation. A case study on six TCR-mimic antibody structures measured by surface plasmon resonance gave a correlation of 0.13 ± 0.39, and the authors are explicit that loop-dominated antibody interfaces remain hard.

#Applications

Binding ΔΔG prediction underpins affinity maturation of antibodies and other binders, interpretation of interface variants in disease, and design campaigns where thousands of candidate substitutions must be triaged before anything is expressed. The speed difference is what changes the workflow: a force-field scan that takes CPU-days can be run in minutes, making exhaustive site-saturation scans over an entire interface practical, and the released checkpoint applies to new PDB complexes without per-target retraining. Predictions are most reliable on rigid, compact interfaces.

#Impact

StaB-ddG is the first deep learning method to reach parity with empirical force fields on protein-protein binding ΔΔG, and its central argument — that scarce binding data should be attacked through a thermodynamic identity that lets plentiful stability data do the work — generalizes beyond this particular backbone. Code, training scripts, and all three checkpoints (the ProteinMPNN initialization, the stability-fine-tuned intermediate, and the final inference model) are released under an MIT license, and the method is also hosted on third-party inference platforms. It remains a specialized single-task regressor rather than a general-purpose representation model, and its fixed-backbone assumption caps performance on flexible interfaces, the antibody loops included.