Use StaB-ddG Online
Commercially Available StaB-ddG No-Code Web Server
StaB-ddG: Deep Learning Prediction of Mutational Effects on Protein Binding
StaB-ddG is a deep learning predictor that estimates how amino acid mutations change protein-protein binding free energy (binding ΔΔG) from a wild-type complex structure. It was developed by Arthur Deng, Karsten Householder, Fang Wu, Sebastian Thrun, K. Christopher Garcia and Brian Trippe at Stanford University, and was presented at ICML 2025. Training and inference code is openly available on GitHub under the MIT license.
Predicting how a mutation alters binding energy matters for structural biology and therapeutic design, for example when avoiding off-target binding. According to the authors, deep learning predictors have lagged behind physics-based force-field methods, partly because binding data is scarce: the largest curated dataset covers fewer than 350 interfaces. Force-field methods, in turn, are computationally expensive.
StaB-ddG addresses this by transferring knowledge from folding stability. It treats binding energy as the folding energy of the complex minus the folding energies of its two binding partners, so abundant stability data can support a task with little binding data.
Attribute | Detail |
|---|---|
Task | Mutational effect on binding energy (ΔΔG, kcal/mol) |
Base model | ProteinMPNN inverse-folding model, fine-tuned |
Inputs | Wild-type complex structure (PDB), mutation(s), the two binding partner chain groups |
Fine-tuning data | Megascale stability data, then SKEMPIv2.0 binding data |
Reported speed | About 0.2 s per mutation vs. about 210 s for FoldX |
Venue | ICML 2025 (PMLR 267) |
Why StaB-ddG Matters for Protein Binding Affinity Prediction
The authors describe StaB-ddG as the first deep learning predictor to match the accuracy of the empirical force-field method FoldX, while running over 1,000 times faster on a single device. They report about 13 GPU-minutes for 4,451 mutants, compared with roughly 260 CPU-hours for FoldX on the same set. Flex ddG is reported to need roughly 15 CPU-hours per mutation.
How StaB-ddG Works
Folding energy as a proxy for binding energy
A pre-trained inverse-folding model scores a sequence by its log-likelihood given a backbone. The binding score is the score of the complex minus the scores of each partner. The predicted ΔΔG is the binding score of the mutant minus that of the wild type. The authors note the method is antisymmetric and independent of the mutational path, and that it supports multi-chain complexes and multiple mutations.
Training and sampling pipeline
Stability fine-tuning: the model is fine-tuned on the Megascale dataset (776,298 measurements across 412 structures).
Binding fine-tuning: the stability-tuned model is then fine-tuned on SKEMPIv2.0 binding measurements.
Antithetic variates: wild type and mutant share the same decoding order and backbone noise, which reduces variance. Removing this lowered per-interface Spearman to 0.30 in an ablation.
Monte Carlo ensembling: predictions are averaged over 20 samples in the main experiments.
StaB-ddG Benchmark Performance
SKEMPIv2.0 test split
Method | Per-interface Spearman | Overall Spearman | Overall RMSE (kcal/mol) | Overall AUROC |
|---|---|---|---|---|
StaB-ddG | 0.45 ± 0.04 | 0.53 ± 0.06 | 1.72 ± 0.11 | 0.73 ± 0.05 |
StaB-ddG + FoldX ensemble | 0.53 ± 0.03 | 0.61 ± 0.05 | 1.62 ± 0.11 | 0.78 ± 0.04 |
FoldX | 0.48 ± 0.03 | 0.56 ± 0.05 | 1.92 ± 0.12 | 0.77 ± 0.04 |
Flex ddG | 0.42 ± 0.04 | 0.54 ± 0.05 | 3.98 ± 1.65 | 0.74 ± 0.03 |
PPIformer | 0.20 ± 0.04 | 0.42 ± 0.06 | 1.77 ± 0.09 | 0.71 ± 0.05 |
Surface-VQMAE | 0.33 ± 0.05 | 0.44 ± 0.05 | 1.76 ± 0.09 | 0.65 ± 0.03 |
The authors found no statistically significant difference among StaB-ddG, its zero-shot variant, FoldX and Flex ddG. Combining StaB-ddG with FoldX gave the best scores in the table.
Other evaluations
Yeast surface display: on 28,293 single mutants across 33 mini-binder interfaces, StaB-ddG reached a per-interface Spearman of 0.296 ± 0.026.
Folding stability: on the Megascale test set (single mutations), the stability-tuned model reached Spearman 0.69 versus 0.73 for ThermoMPNN.
Complex size: RMSE was 0.94 kcal/mol for complexes under 150 residues and 1.50 for those under 1000 residues.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with StaB-ddG on Tamarind Bio
Therapeutic design: estimate how mutations shift binding energy when engineering binders.
Off-target avoidance: the authors highlight avoiding off-target binding as a motivation for accurate ΔΔG prediction.
Large mutation scans: at about 0.2 s per mutation in the authors' batched runs, many variants can be scored quickly.
Consensus scoring: pairing StaB-ddG with FoldX improved every reported metric over either alone.
How to Use StaB-ddG on Tamarind Bio
Open the tool: log in to tamarind.bio and select StaB-ddG.
Provide the structure: supply the wild-type complex structure as a PDB file.
Specify mutations: give one or more mutations, each defined by wild-type residue, chain, position and new residue (for example YH103H); multiple mutations are separated by commas.
Define binding partners: indicate which chains form each side of the interface (for example ABC and DE).
Set sampling: the reference code uses 20 Monte Carlo samples by default, if this parameter is exposed.
Run the job: submit and wait for completion.
Download results: the reference implementation outputs a CSV with the predicted ΔΔG per mutation.
Interpret and use: a negative ΔΔG indicates a stabilizing (more favorable binding) mutation; shortlist variants for design or follow-up experiments.
Things to Keep in Mind
The model does not account for backbone changes upon mutation and uses a single static structure, which the authors link to weaker performance on flexible interfaces.
Performance is better on rigid, smaller complexes.
Accuracy for multiple mutations is weaker, likely due to limited data.
On a small TCR mimic case study (30 measurements), results were inconclusive (Spearman 0.13 ± 0.39).
Yeast display readouts confound binding with expression, which limited the binding predictor's advantage there.
Source: Deng A, Householder K, Wu F, Thrun S, Garcia KC, Trippe B. "Predicting mutational effects on protein binding from folding energy." Proceedings of the 42nd International Conference on Machine Learning (ICML), PMLR 267, 2025. arXiv:2507.05502. Code: https://github.com/LDeng0205/StaB-ddG.