Research map / Large language models / Alignment and preferences
Controllable Protein Sequence Generation with LLM Preference Optimization
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 CtrlProt, a prefix-tuned protein LLM (ProtGPT2-based) further refined with a multi-listwise DPO variant that uses Rosetta stability and structure-based functionality scores, enabling single- and multi-attribute controllable protein sequence generation.
Problem
Controllable protein sequence generation for GO-defined functional attributes often yields sequences with poor functionality and structural stability, and finetuning data is scarce, especially for multi-attribute combinations.
Approach
Builds preference pairs from candidate sequences scored by Rosetta-energy-based stability and ProteinMPNN-embedding-based functionality, then applies a DPO variant whose regularization term uses a beta-CDF-weighted quality-score difference to emphasize pairs with large quality gaps; multi-attribute generation concatenates single-attribute prefixes and reuses the same optimization without multi-att
Key results
- On single-attribute generation, CtrlProt reports the best TM-score and pLDDT on the six GO attributes, e.g., 94.1 TM-score and 83.8 pLDDT on BPO: Phosphorylation versus 75.0 and 76.6 for the best baselines.
- Ablation: full CtrlProt averages CLS 79.5, TM 77.7, RMSD 4.30, pLDDT 73.9, beating DPO (CLS 76.4, TM 72.0), ORPO and Lipo-lambda variants.
- Multi-attribute generation with concatenated prefixes plus the proposed optimization gives higher CLS and TM and lower RMSD than Average and Concat baselines across six attribute combinations.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the structural stability metric γ to preference optimization (ablation) | CtrlProt · Six single-attribute datasets (average) | pLDDT: 70.2 -> 73.9 | Helped |
| add the functionality metric τ to preference optimization (ablation) | CtrlProt · Six single-attribute datasets (average) | CLS: 71.9 -> 79.5 | Helped |
| add the functionality metric τ to preference optimization (ablation) | CtrlProt · Six single-attribute datasets (average) | TM: 66.2 -> 77.7 | Helped |
| add the functionality metric τ to preference optimization (ablation) | CtrlProt · Six single-attribute datasets (average) | RMSD: 4.90 -> 4.30 | Helped |
| replace multi-listwise preference optimization with DPO (ablation) | CtrlProt · Six single-attribute datasets (average) | CLS: 79.5 -> 76.4 | Hurt |
| replace multi-listwise preference optimization with ORPO (ablation) | CtrlProt · Six single-attribute datasets (average) | CLS: 79.5 -> 74.0 | Hurt |
| replace multi-listwise preference optimization with Lipo-λ (ablation) | CtrlProt · Six single-attribute datasets (average) | TM: 77.7 -> 70.1 | Hurt |
| apply preference optimization after concatenating single-attribute prefixes for multi-attribute generation | CtrlProt · six attribute combinations | CLS-score | Helped |
| apply preference optimization after concatenating single-attribute prefixes for multi-attribute generation | CtrlProt · six attribute combinations | pLDDT | Helped |
Surprising findings
- Lipo-lambda, a listwise method, underperforms plain DPO on protein generation, which the authors attribute to overemphasis on top-ranked sequences.
Limitations
- Achieving precise, programmable generation for certain attribute combinations remains challenging.
- Multi-attribute performance is evaluated only on six attribute combinations.
- ProLLaMA underperforms likely due to its 256-length instruction-tuning restriction, a confound in comparison.
Experimental setup
Datasets: UniProtKB protein sequences with GO annotations AlphaFold structures
Benchmarks: UniProtKB GO-attribute evaluation sets (100k per attribute)
Metrics: CLS-score TM-score RMSD pLDDT 3-gram inter-output similarity training-set similarity
Baselines: ESM-1b ESM-2 EvoDiff PrefixProt ProGen2 ProLLaMA DPO ORPO Lipo-lambda
Builds on: DPO prefix-tuning ProtGPT2 ProLLaMA Lipo-lambda
Closest papers by meaning
- Steering Protein Language Models (ICML 2025 · similarity 0.88)
- Diffusion Language Models Are Versatile Protein Learners (ICML 2024 · similarity 0.87)
- Diffusion on Language Model Encodings for Protein Sequence Generation (ICML 2025 · similarity 0.87)
- ProteinOPD: Towards Effective and Efficient Preference Alignment for Protein Design (NeurIPS 2026 · similarity 0.87)
- Steering Generative Models with Experimental Data for Protein Fitness Optimization (NeurIPS 2025 · similarity 0.87)
- Generalists vs. Specialists: Evaluating LLMs on Highly-Constrained Biophysical Sequence Optimization Tasks (ICML 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.