Research map / Probabilistic and scientific modeling / Molecules and proteins

Learning to Predict Mutational Effects of Protein-Protein Interactions by Microenvironment-aware Hierarchical Prompt Learning

ICML 2024 · Probabilistic and scientific modeling › Molecules and proteins · approach: Protein fitness and variant effect prediction · 4 citations (OpenAlex)

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.

Prompt-DDG pre-trains a hierarchical prompt codebook (residue type, angular, local conformation) via masked microenvironment modeling and uses lightweight prompts to predict mutational effects on protein-protein binding (ΔΔG), outperforming pre-training baselines on SKEMPI v2 without external pre-training data.

Problem

Predicting how amino acid mutations change protein-protein binding free energy (ΔΔG) is limited by scarce annotated mutation data and missing mutated complex structures; existing pre-training methods are costly and ignore per-mutation local microenvironment conformational changes and higher-order mu

Approach

Build a hierarchical discrete prompt codebook with three sub-codebooks (residue type, angular statistics, local conformation) pre-trained jointly by masked microenvironment modeling, then encode each mutation's microenvironment into prompts that are added to residues in the local microenvironment, avoiding heavy large-scale pre-training.

Key results

• On SKEMPI v2 3-fold CV, Prompt-DDG reaches per-structure Pearson 0.4712 and Spearman 0.4257, beating RDE-Network (0.4448/0.4010) and DiffAffinity (0.4220/0.3970).
• On multi-point mutations, Prompt-DDG achieves overall Pearson 0.6780 and AUROC 0.8187, outperforming the best pre-training baselines by a large margin.
• In the SARS-CoV-2 antibody case study, Prompt-DDG gives an average ranking of 10.69% for five favorable mutations versus 18.26% for RDE-Network and 24.49% for DiffAffinity.

Measured effects reported in the paper

Surprising findings

• Prompt-DDG, trained only on SKEMPI v2 (7k mutations) without any additional pre-training data, outperforms methods pre-trained on 143k PDB-REDO structures.
• The local conformation (3D coordinate) prompt contributes more than residue-type or angular prompts in the ablation.

Limitations

• Focuses on two-protein interactions; not applicable to multimeric proteins or substrate-related enzymes (stated).
• Biological validation limited to a single antibody case study without wet-lab experiments (stated in Impact Statement).

Experimental setup

Datasets: SKEMPI v2.0 PDB-REDO

Benchmarks: SKEMPI v2.0

Metrics: per-structure Pearson per-structure Spearman Pearson Spearman RMSE MAE AUROC

Baselines: ESM-1v PSSM MSA Transformer Tranception Rosetta FoldX DDGPred End-to-End ESM-1F MIF-Network RDE-Network DiffAffinity

Builds on: MAPE-PPI RDE DiffAffinity MEAN PromptProtein

Closest papers by meaning

• MAPE-PPI: Towards Effective and Efficient Protein-Protein Interaction Prediction via Microenvironment-Aware Protein Embedding (ICLR 2024 · similarity 0.87)
• Enhancing Protein Mutation Effect Prediction through a Retrieval-Augmented Framework (NeurIPS 2024 · similarity 0.86)
• MSAGPT: Neural Prompting Protein Structure Prediction via MSA Generative Pre-Training (NeurIPS 2024 · similarity 0.86)
• MutaPLM: Protein Language Modeling for Mutation Explanation and Engineering (NeurIPS 2024 · similarity 0.85)
• Learning to design protein-protein interactions with enhanced generalization (ICLR 2024 · similarity 0.84)
• Predicting mutational effects on protein binding from folding energy (ICML 2025 · similarity 0.84)