EBM-DDG

Engineering a protein-protein interaction — tightening an antibody against its antigen, or weakening an unwanted interface — depends on predicting the change in binding free energy (ΔΔG) caused by a mutation. Deep learning methods have exploited the observation that ΔΔG correlates strongly with conformational entropy, estimating it from proxies such as side-chain torsion angles or residue identities. The underlying quantity, the full conformational distribution of a protein complex, is generally intractable to compute, so these methods substitute a surrogate and inherit its blind spots.

EBM-DDG, developed at the University of Virginia and presented at KDD '25, takes a different route. It decomposes ΔΔG into two terms: a sequence-based component estimated by an inverse folding model, and a structure-based component ΔΔE estimated by an energy-based model. Assuming equilibrium between the bound and unbound states lets the degeneracy factors associated with each state largely cancel, which is what makes the energy term tractable. The result connects the widely used sequence log-odds-ratio approach to a term grounded in statistical mechanics.

The practical consequence is that EBM-DDG does not hold the backbone fixed. Because the energy model samples mutant backbone coordinates through Langevin dynamics rather than scoring a frozen structure, it handles mutations that reorganize the local backbone — precisely the cases where fixed-backbone predictors degrade.

#Key Features

• Two-term free energy decomposition: ΔΔG is split into a sequence term computed from inverse folding log-odds and a structure term ΔΔE from a learned energy function, joined by a physically motivated equilibrium assumption rather than a fitted head.
• Pretrained inverse folding backbone: the sequence term reuses a pretrained ProteinMPNN checkpoint, taking half the log-probabilities of mutated residues rather than scoring the entire sequence.
• Denoising energy model: the ΔΔE term is a Gaussian denoising score-matching energy model whose encoder is initialized from frozen ESM-2 3B residue embeddings, trained so that sampled conformations track the learned energy landscape.
• Backbone-flexible mutations: on mutations to glycine and proline — residues that reshape backbone geometry — EBM-DDG reaches the lowest error of the compared structure-aware predictors.
• Transfers without retraining: the SKEMPI-trained checkpoint is applied directly to antibody CDR mutation ranking, a task and structural regime it never saw in training.

#Technical Details

The energy model is a one-layer Invariant Point Attention encoder without a pair representation, replacing the SRU++ frame-averaging encoder used in DSMBind, followed by an all-atom interaction block over pairwise atom embeddings. Residue features come from ESM-2 3B as 2560-dimensional embeddings, projected down to a hidden dimension of 256 by a linear layer; the inverse folding model is a ProteinMPNN variant pretrained on the CATH 4.2 40% non-redundant set with an embedding dimension of 256. The energy model is trained from scratch on SKEMPI v2.0 crystal structures for 30K steps with Adam on a single NVIDIA A100 80GB GPU, with a KL term against the original inverse folding distribution to prevent catastrophic forgetting. Sampling uses ten Langevin denoising steps, which the authors find sufficient for optimal accuracy.

Under three-fold cross-validation on SKEMPI v2.0 (7,085 mutations across 348 complexes), EBM-DDG reaches per-structure Pearson 0.5681 and Spearman 0.5184, overall Pearson 0.7385 and Spearman 0.6516, RMSE 1.3901, MAE 0.9871, and AUROC 0.7941 — ahead of Prompt-DDG (per-structure Pearson 0.4712), RDE-Network (0.4448), FoldX (0.3789), and Rosetta (0.3284). Because the BA-DDG baseline's code was unavailable at submission, the authors reimplemented it, and their reported numbers for that baseline differ from the original publication.

#Applications

The direct use is antibody affinity maturation and protein-protein interface engineering, where a candidate list of point mutations must be ranked before any of them are made. The paper's case study takes the SKEMPI-trained model unchanged and scores all 494 single-point mutations across 26 CDR positions of a SARS-CoV-2 neutralizing antibody heavy chain. Five mutations known experimentally to improve neutralization land at an average rank of 11.90%, the best of the compared methods (BA-DDG 12.87%, DDGPred 20.00%, FoldX 21.78%).

#Impact

EBM-DDG's contribution is methodological: it reinstates a statistical-mechanics inductive bias in a problem that deep learning had largely reduced to supervised regression, and shows that relaxing the fixed-backbone assumption pays off on exactly the mutation classes where that assumption fails. Its limits are clear. No code or trained weights have been released, so reproduction requires reimplementation. Evaluation rests on one cross-validation split of SKEMPI v2.0 plus a single antibody case study, a narrowness the authors themselves flag as a problem for the whole ΔΔG literature. And the method assumes a representative static structure for each complex, which excludes intrinsically disordered and membrane proteins.