Research map / Probabilistic and scientific modeling / Molecules and proteins
A Simple yet Effective $\Delta\Delta G$ Predictor is An Unsupervised Antibody Optimizer and Explainer
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Light-DDG, a lightweight structure-aware Transformer ΔΔG predictor trained via supervised pretraining on a newly augmented SKEMPI-Aug dataset and knowledge distillation from Prompt-DDG, and uses it with a Mutation Explainer and preference-guided search for unsupervised antibody optimization (Uni-Anti).
Problem
Predicting binding free energy change (ΔΔG) upon mutations of protein complexes is slow and data-scarce for existing methods, and mutation preferences/synergistic effects are hard to explain and search efficiently in the huge combinatorial mutation space of antibodies.
Approach
Replace heavy IPA-style backbones with a lightweight Transformer, compensate with a large cross-augmented dataset labeled by Prompt-DDG and knowledge distillation from Prompt-DDG, then use an iterative Shapley-value Mutation Explainer to learn site/mutation preferences that guide sampling-based antibody search using the fast predictor as the fitness function.
Key results
- Light-DDG reaches per-structure Pearson 0.5440 vs 0.4712 for Prompt-DDG (+15.45%) on SKEMPI v2.0 3-fold CV.
- Light-DDG overall Pearson 0.7429 and RMSE 1.3837, outperforming Prompt-DDG (0.6772, 1.5207).
- Reported 89.7x inference acceleration over Prompt-DDG; directed mutation search on SARS-CoV-2 achieves -2.872 kcal/mol best ΔΔG versus -1.865 for random mutation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Light-DDG instead of Prompt-DDG | ΔΔG prediction · SKEMPI v2.0 | Per-structure Pearson: 0.4712 -> 0.5440 | Helped |
| add knowledge distillation (ablation) | Light-DDG without KD or augmentation · SKEMPI v2.0 | Per-structure Pearson: 0.3888 -> 0.4809 | Helped |
| add supervised data augmentation (ablation) | Light-DDG without KD or augmentation · SKEMPI v2.0 | Per-structure Pearson: 0.3888 -> 0.4516 | Helped |
| combine knowledge distillation and augmentation (ablation) | Light-DDG · SKEMPI v2.0 | Per-structure Pearson: 0.4809 -> 0.5440; 0.4516 -> 0.5440 | Helped |
| add wild-type structure input (ablation) | sequence-only Light-DDG · SKEMPI v2.0 | Per-structure Pearson: 0.4154 -> 0.5440 | Helped |
| add mutant structure input (ablation) | Light-DDG with wild-type structure · SKEMPI v2.0 | Per-structure Pearson: 0.5440 -> 0.5496 | Helped |
| increase augmentation data size (ablation) | Light-DDG · SKEMPI v2.0 | Performance | Helped |
| add Gaussian noise to wild-type structures (ablation) | Light-DDG · SKEMPI v2.0 | Performance | Hurt |
| use preference-guided directed mutations instead of random mutations | SARS-CoV-2 antibody optimization · SARS-CoV-2 antibody optimization, CDR-H1/2/3 | Minimum ΔΔG: -1.325 -> -2.872 | Helped |
| use Uni-Anti instead of dyMEAN | SAbDab antibody optimization · SAbDab, CDR-H3 optimization | Average optimal ΔΔG: -3.05 -> -3.87 | Helped |
| use Light-DDG instead of Prompt-DDG | ΔΔG prediction · SKEMPI v2.0 | Inference speed: 89.7× inference acceleration | Helped |
Surprising findings
- Sequence-only Light-DDG outperforms energy-based and supervised baselines, but structure-based input is better.
- Mutated structures from ESMFold only slightly improve performance because the teacher is already implicitly mutated-structure-aware.
Limitations
- ΔΔG is only one prior constraining protein evolution; other priors are not combined.
- Only antibodies studied; no wet-lab experimental validation of optimized antibodies.
- Performance degrades with 3D structure noise above around 1 Å, so structure-based input is sensitive to predicted structure errors.
Experimental setup
Datasets: SKEMPI v2.0 SKEMPI-Aug PDB-REDO AFDB SAbDab
Benchmarks: SKEMPI v2.0 SARS-CoV-2 antibody case study SAbDab CDR-H3 optimization
Metrics: Pearson Spearman RMSE MAE AUROC per-structure Pearson per-structure Spearman minimal ΔΔG of optimized antibodies average ranking
Baselines: ESM-1v Rosetta FoldX DDGPred ESM-1F MIF-Network RDE-Network DiffAffinity ProMIM Prompt-DDG RefineGNN MEAN
Builds on: Prompt-DDG ProMIM RDE Iterative Target Augmentation
Closest papers by meaning
- Bayesian Optimization of Antibodies Informed by a Generative Model of Evolving Sequences (ICLR 2025 · similarity 0.87)
- Improving protein optimization with smoothed fitness landscapes (ICLR 2024 · similarity 0.86)
- Pareto-Optimal Energy Alignment for Designing Nature-Like Antibodies (NeurIPS 2025 · similarity 0.85)
- AffinityFlow: Guided Flows for Antibody Affinity Maturation (ICML 2025 · similarity 0.84)
- DePLM: Denoising Protein Language Models for Property Optimization (NeurIPS 2024 · similarity 0.84)
- Retrieval Augmented Diffusion Model for Structure-informed Antibody Design and Optimization (ICLR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.