Research map / Vision-language and multimodal / Human and embodied modeling
MutaPLM: Protein Language Modeling for Mutation Explanation and Engineering
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 MutaPLM, which adds a protein delta network and a CoT-based transfer learning pipeline to ESM-2 and a biomedical LLM, plus the MutaDescribe dataset, for explaining and engineering protein mutations via text.
Problem
Protein language models model mutations only implicitly via evolutionary plausibility, offering neither human-readable explanations of mutational effects nor text-guided mutation proposals.
Approach
Model mutations explicitly as delta features (difference of ESM-2 wild-type and mutant representations) compressed by cross-attention into a shared space aligned with LLM text, trained with protein-literature pre-training and a two-round chain-of-thought fine-tuning on mutation annotations.
Key results
- MutaPLM reaches 21.34 average ROUGE-L and 13.61 BLEU-2 on MutaDescribe mutation explanation, versus 14.88 ROUGE-L for fine-tuned ESM-2.
- For mutation engineering, MutaPLM achieves 53.51% accuracy and 40.94 Rec@50 on average, about 1.6-fold the ESM-2 Rec@50 of 25.15.
- Ablation shows removing the predicted protein function drops explanation ROUGE-L from 21.34 to 18.54, and removing delta features drops it to 17.36.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MutaPLM instead of Fine-tuned ESM-2 | Mutation explanation · MutaDescribe test average | ROUGE-L: 14.88 -> 21.34 | Helped |
| Use MutaPLM instead of GPT-4 + ESM-2 | Mutation explanation · MutaDescribe test average | BLEU-2: 10.37 -> 13.61 | Helped |
| Use MutaPLM instead of Fine-tuned ESM-2 | Mutation engineering · MutaDescribe test average | accuracy: 51.43 -> 53.51 | Helped |
| Use MutaPLM instead of Fine-tuned ESM-2 | Mutation engineering · MutaDescribe test average | Rec@50: 33.77 -> 40.94 | Helped |
| add protein function information to the CoT dialogue (ablation) | MutaPLM mutation explanation · MutaDescribe test average | ROUGE-L: 18.54 -> 21.34 | Helped |
| add protein function information to the CoT dialogue (ablation) | MutaPLM mutation engineering · MutaDescribe test average | Rec@50: 39.81 -> 40.94 | Helped |
| replace predicted protein function with ground-truth function (ablation) | MutaPLM mutation explanation · MutaDescribe test average | ROUGE-L: 21.34 -> 23.80 | Helped |
| replace predicted protein function with ground-truth function (ablation) | MutaPLM mutation engineering · MutaDescribe test average | Rec@50: 40.94 -> 41.26 | Helped |
| add delta features to mutation explanation (ablation) | MutaPLM · MutaDescribe test average | ROUGE-L: 17.36 -> 21.34 | Helped |
| add mutational effects to mutation engineering input (ablation) | MutaPLM · MutaDescribe test average | Rec@50: 35.10 -> 40.94 | Helped |
Surprising findings
- Supervised baselines underperform few-shot GPT-4 on Medium and Hard sets and on BLEU-2, which the authors attribute to overfitting and random recombination of training text segments.
- Fine-tuning on mutation data improves Easy-set performance but degrades Medium and Hard sets, indicating overfitting on out-of-domain samples.
Limitations
- Restricted to single-site substitutions; multi-point mutations and indels are future work.
- Performance drops substantially on Medium and Hard (out-of-homology) splits, and fine-tuning overfits out-of-domain samples.
- Safety concern: the model could be misused to generate pathogenic mutations, so release is restricted to research purposes.
Experimental setup
Datasets: UniProtKB/SwissProt PubMed abstracts ProteinGym
Benchmarks: MutaDescribe
Metrics: ROUGE-L BLEU-2 accuracy Recall@50 human-AI accurate/relevant rate
Baselines: ProLLaMA Mol-Instructions Galactica-6.7B GPT-4-0613 ESM-2 OntoProtein AugmentedESM Fine-tuned ESM-2 ProtST (ESM-2) Fine-tuned BioMedGPT EvoProtGrad
Builds on: ESM-2 BioMedGPT-LM LLaMA2 BLIP-2 LM-DESIGN ProtST
Closest papers by meaning
- DPLM-2: A Multimodal Diffusion Protein Language Model (ICLR 2025 · similarity 0.88)
- DePLM: Denoising Protein Language Models for Property Optimization (NeurIPS 2024 · similarity 0.88)
- Understanding protein function with a multimodal retrieval-augmented foundation model (NeurIPS 2025 · similarity 0.88)
- Knowledge-aware Reinforced Language Models for Protein Directed Evolution (ICML 2024 · similarity 0.88)
- LEMON-ZEST: Evolution-Informed Tokenization for Efficient Protein Language Modeling (NeurIPS 2026 · similarity 0.88)
- Steering Protein Language Models (ICML 2025 · similarity 0.87)
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.