Research map / Large language models / Alignment and preferences
ProteinOPD: Towards Effective and Efficient Preference Alignment for Protein Design
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 ProteinOPD, which adapts protein language models into preference-specific SFT teachers and distills them into a student via on-policy distillation, with a normalized product-of-experts consensus target for multi-objective alignment.
Problem
Aligning protein language models to multiple competing design preferences (foldability, solubility, thermostability) while preserving designability (plausibility and novelty), without the cost and forgetting of RL or the limits of test-time steering.
Approach
Extends on-policy distillation to multi-teacher settings: the student matches a normalized product-of-experts (geometric mean) of preference-specific SFT teachers, giving a consensus target that keeps only shared support across teachers and provides bounded signals under conflict.
Key results
- ProteinOPD achieves HV 0.62 in unconditional multi-objective alignment, versus 0.46 for MoMPNN and 0.33 for ProtGPT2.
- Relative to ProtGPT2, ProteinOPD reduces PPL by 83.7% and improves pLDDT, solubility, and thermostability by 14.8%, 16.9%, and 54.2%.
- OPD reaches thermostability 0.25 in 3.86 minutes versus 31.2 minutes for ProtRL, an 8x training speed-up.
Surprising findings
- Teacher SFT on ProtGPT2 cuts novelty by 42.4%, whereas the OPD student keeps novelty loss to 1.7%, showing OPD mitigates SFT forgetting.
- Preliminary study: RL preserves designability but thermostability improvement saturates quickly.
Limitations
- Evaluation relies on computational oracles (ESMFold, TemBERTure, Protein-Sol) without wet-lab validation.
- MoMPNN comparison is cross-paradigm (inverse folding with backbones), not task-identical.
- Multi-objective weights for ASPO and PoE teachers are manually predefined.
Experimental setup
Datasets: UniRef50 UniProtKB SwissProt
Benchmarks: PDFBench-based evaluation of unconditional and lysozyme-like conditional protein generation
Metrics: PPL Novelty-U Novelty-T pLDDT pAE Solubility Thermostability ProTrek score hypervolume
Baselines: ESM2 ESM3 ProGen2 ProLLaMA ProtGPT2 Pinal ProteinDT ProDVa ASPO MoMPNN ProtRL
Builds on: on-policy distillation GRPO ASPO MoMPNN ProLLaMA
Closest papers by meaning
- Aligning Transformers with Continuous Feedback via Energy Rank Alignment (NeurIPS 2025 · similarity 0.87)
- Protein Design with Agent Rosetta: A Case Study for Specialized Scientific Agents (ICML 2026 · similarity 0.87)
- Controllable Protein Sequence Generation with LLM Preference Optimization (AAAI 2025 · similarity 0.87)
- Optimistic Games for Combinatorial Bayesian Optimization with Application to Protein Design (ICLR 2025 · similarity 0.86)
- Protein Design with Dynamic Protein Vocabulary (NeurIPS 2025 · similarity 0.86)
- Protein Inverse Folding From Structure Feedback (NeurIPS 2025 · similarity 0.86)
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.