Light-DDG

Predicting the change in binding free energy (ΔΔG) when a protein complex is mutated is the standard prior used to steer antibody optimization toward variants that bind better. The accurate predictors are all expensive: they lean on invariant-point-attention backbones inherited from structure prediction, which makes a single ΔΔG estimate slow enough that exhaustively screening a mutation space is impractical. Since the interesting mutation spaces are combinatorially large, throughput matters as much as accuracy.

Light-DDG, from Westlake University and DP Technology, takes the opposite design decision: shrink the model radically, then recover the lost capacity with data. The backbone is a four-layer structure-aware transformer with a hidden dimension of 128 — small enough that its predictions come back in milliseconds — and it is trained in two stages. First, supervised pretraining on a large augmented mutation corpus whose labels were generated by Prompt-DDG, a heavyweight state-of-the-art predictor. Then fine-tuning on experimental SKEMPI v2.0 data under joint supervision from ground-truth labels and distillation from the same teacher. The result is both faster and more accurate than its teacher.

The paper's second argument is that once ΔΔG scoring is cheap enough, a predictor is all you need for antibody optimization: no generative model, no per-target retraining. Around Light-DDG the authors build Uni-Anti, which explains which mutations matter and searches the space they define. It was published at ICLR 2025.

#Key Features

• Distilled lightweight backbone: Four transformer layers at hidden dimension 128 with four attention heads, distilled from Prompt-DDG, delivering 89.7× faster inference than the teacher it was trained from.
• Cross-augmented pretraining data: A K-fold cross-augmentation scheme trains a fresh teacher on K-1 folds and uses it to annotate random mutations from the held-out fold, so the data a teacher labels is never data it saw, avoiding leakage.
• Mutation Explainer: An iterative Shapley value estimation algorithm assigns each residue substitution its marginal contribution to ΔΔG through coarse-to-fine iteration, capturing synergy between mutations without enumerating the full combinatorial space.
• Preference-guided optimization: Learned site-wise mutation preferences steer sampling toward promising regions, and candidates are scored immediately by the same model, so multiple CDRs can be co-optimized in one pass.

#Technical Details

Wild-type and mutant complexes are encoded separately by a parameter-shared transformer over residue graphs, with backbone-atom geometry in local residue frames as edge features, and an MLP head predicts ΔΔG. Pretraining uses SKEMPI-Aug, an augmented and released mutation dataset that the paper's dataset table lists at 640k annotated mutations — the experimental section describes it as 670k, and the abstract as "millions." Fine-tuning and evaluation use SKEMPI v2.0: 348 complexes and 7,085 mutation combinations, split into three folds by complex for cross-validation.

Averaged over three folds, Light-DDG reaches 0.5440 per-structure Pearson and 0.5004 per-structure Spearman, against 0.4712 and 0.4257 for Prompt-DDG — gains of 15.45% and 17.55%. Overall Pearson is 0.7429, RMSE 1.3837, MAE 0.9697, and AUROC 0.7935, ahead of ProMIM, DiffAffinity, RDE-Network, MIF-Network, FoldX, and Rosetta. In the SARS-CoV-2 case study it screens 494 candidate human antibodies with mutations across 26 sites in three heavy-chain CDRs, achieving the best average ranking of the five known-favorable mutations among nine methods, and it evaluates 10,000 sampled variants in under five minutes.

#Applications

Light-DDG suits antibody engineering campaigns that need to triage large mutation libraries before committing to synthesis: affinity maturation of a lead, joint optimization across multiple CDRs, or re-ranking designs produced by a generative model such as RefineGNN, MEAN, DiffAb, or dyMEAN. Its speed also makes it usable as the fitness function inside an outer search loop, and the Mutation Explainer gives a per-residue rationale that engineers can inspect rather than a single opaque score.

#Impact

Light-DDG's main result is a useful inversion of the usual tradeoff: a much smaller model, given enough teacher-annotated data, beat the teacher outright, showing that ΔΔG prediction had been architecture-limited less than data-limited. Its second claim — that a sufficiently fast supervised predictor is itself a competitive unsupervised antibody optimizer — undercuts the assumption that optimization requires a generative model. Code and the trained checkpoint are released under the MIT license, alongside the augmented dataset. The evaluation is entirely in silico, with no wet-lab binding measurements, and the pretraining labels come from a model rather than experiment, so systematic errors in Prompt-DDG propagate into the student.