Research map / Probabilistic and scientific modeling / Molecules and proteins
Protein Multimer Structure Prediction via Prompt Learning
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 PromptMSP, a pre-training plus prompt-tuning framework (with MAML-based meta-initialization) that reformulates conditional protein-protein interaction link prediction as graph-level regression to assemble protein multimers step by step.
Problem
Protein multimer structure prediction via step-wise assembly suffers poor generalization across multimers of different chain numbers, because condition PPI knowledge varies with scale and large-scale experimental multimer data is scarce.
Approach
Pre-train a GIN on random small-scale (3-5 chain) assembly graphs to regress assembly correctness (TM-Score), then learn a cross-attention prompt model that turns any arbitrary-scale query chain pair into a fixed 4-node virtual path, so the linking probability is read from the pre-trained correctness score; a MAML-style meta-initialization helps adapt to scarce large multimers.
Key results
- On 3<=N<=10 with GT dimers, PromptMSP reaches RMSD 13.57 (avg) and TM-Score 0.67 (avg), versus MoLPC 18.53 / 0.52.
- With AFM dimers on 3<=N<=10, PromptMSP achieves TM-Score 0.55 vs MoLPC 0.43; with ESMFold dimers it reaches 0.45 vs 0.31.
- Inference for an N-chain multimer with GT dimers takes about 0.04 min for 11-30 chains, versus 365.87 min for MoLPC total.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PromptMSP instead of AFM† with GT dimers | Multimer structure prediction · PDB-M test, 3≤N≤10, GT Dimer | RMSD (average): 16.79 -> 13.57 | Helped |
| Use PromptMSP instead of AFM† with GT dimers | Multimer structure prediction · PDB-M test, 3≤N≤10, GT Dimer | TM-Score (average): 0.59 -> 0.67 | Helped |
| Use PromptMSP instead of AFM† with AFM dimers | Multimer structure prediction · PDB-M test, 3≤N≤10, AFM Dimer | RMSD (average): 18.98 -> 17.36 | Helped |
| Use PromptMSP instead of AFM† with AFM dimers | Multimer structure prediction · PDB-M test, 3≤N≤10, AFM Dimer | TM-Score (average): 0.50 -> 0.55 | Helped |
| Use PromptMSP instead of AFM with ESMFold dimers | Multimer structure prediction · PDB-M test, 3≤N≤10, ESMFold Dimer | RMSD (average): 20.99 -> 22.55 | Hurt |
| Use PromptMSP instead of AFM with ESMFold dimers | Multimer structure prediction · PDB-M test, 3≤N≤10, ESMFold Dimer | TM-Score (average): 0.47 -> 0.45 | Hurt |
| Add the prompt model to PromptMSP (ablation) | PromptMSP · GT Dimer, 11≤N≤30 | TM-Score (mean): 0.29 -> 0.37 | Helped |
| Use C-PPI modeling instead of I-PPI modeling (ablation) | PromptMSP · GT Dimer, 11≤N≤30 | TM-Score (mean): 0.33 -> 0.37 | Helped |
| Use PromptMSP-GT instead of MoLPC | Multimer structure prediction · PDB-M test, 3≤N≤10 | total inference time (min): 177.37 -> 0.01 | Helped |
| Use PromptMSP-ESMFold instead of MoLPC | Multimer structure prediction · PDB-M test, 11≤N≤30 | total inference time (min): 365.87 -> 1.10 | Helped |
Surprising findings
- Directly applying the link prediction task for both pre-training and fine-tuning (without the prompt model) hurts large-scale performance; the paper reports an approximately 21.6% drop.
- Ablation indicates that independent PPI (ignoring third-party chains) brings negative transfer to MSP.
Limitations
- Test set is small (80 to 180 samples) because baselines are slow, limiting statistical power.
- Performance drops with dimers from ESMFold, and the method still requires pre-computed dimer structures (N-1 of them).
- Baselines failing on large multimers are padded with upper-bound performance, so comparisons on large scales are partly constructed.
Experimental setup
Datasets: PDB-M
Benchmarks: PDB-M
Metrics: RMSD TM-Score
Baselines: Multi-LZerD RL-MLZerD AlphaFold-Multimer MoLPC
Builds on: MoLPC AlphaFold-Multimer Graph Isomorphism Network MAML prompt learning
Closest papers by meaning
- Deep Reinforcement Learning for Modelling Protein Complexes (ICLR 2024 · similarity 0.88)
- MSAGPT: Neural Prompting Protein Structure Prediction via MSA Generative Pre-Training (NeurIPS 2024 · similarity 0.86)
- PPDiff: Diffusing in Hybrid Sequence-Structure Space for Protein-Protein Complex Design (ICML 2025 · similarity 0.86)
- MIPT: Multilevel Informed Prompt Tuning for Robust Molecular Property Prediction (ICML 2025 · similarity 0.86)
- PSC-CPI: Multi-Scale Protein Sequence-Structure Contrasting for Efficient and Generalizable Compound-Protein Interaction Prediction (AAAI 2024 · similarity 0.86)
- Evaluating Representation Learning on the Protein Structure Universe (ICLR 2024 · 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.