InstructPLM-mu

InstructPLM-mu addresses a practical question in protein modeling: how much of the benefit of a natively multimodal protein language model can be recovered by cheaply grafting structural information onto an existing sequence-only model? Structure-aware models such as ESM3 are trained from scratch to jointly reason over sequence and structure, an expensive undertaking. InstructPLM-mu instead takes the widely used, sequence-only ESM2 and fine-tunes it with structural inputs, showing that a lightweight adaptation can match a purpose-built multimodal model on mutation-effect prediction.

The method fuses structure-derived features — obtained from inverse-folding-style representations in the ProteinMPNN and ESM-IF lineage — into ESM2 and fine-tunes the combined model. The authors systematically compare three fusion designs and several tuning recipes, finding that both the fusion mechanism and the tuning strategy strongly shape final accuracy. The headline result is efficiency: with roughly one hour of fine-tuning, the adapted ESM2 reaches performance comparable to ESM3 on protein mutation predictions. The work comes from researchers at the Chinese University of Hong Kong and the Hangzhou Institute of Medicine, Chinese Academy of Sciences, and was posted as a preprint in October 2025.

By demonstrating that structural conditioning can be retrofitted onto a frozen-then-tuned sequence backbone at low cost, InstructPLM-mu offers a resource-conscious alternative to training large multimodal protein models from scratch.

#Key Features

• Structure-conditioned ESM2: Injects structural features into the sequence-only ESM2 backbone rather than training a joint sequence–structure model from scratch.
• One-hour adaptation: Reaches ESM3-comparable mutation-prediction accuracy after roughly one hour of fine-tuning, a fraction of the cost of native multimodal pretraining.
• Systematic fusion study: Evaluates three distinct fusion designs and multiple fine-tuning recipes, quantifying how each choice affects downstream accuracy.
• Builds on established components: Reuses ESM2 and inverse-folding representations from the ProteinMPNN / ESM-IF family, keeping the recipe reproducible with existing open models.

#Technical Details

InstructPLM-mu is a multimodal fine-tuning framework applied to the ESM2 transformer family (evaluated across the 35M, 150M, and 650M-parameter variants). Structural context enters through features drawn from inverse-folding-style encoders, and the study contrasts alternative ways of fusing these features with ESM2's sequence representations alongside different parameter-update strategies. On protein mutation-effect benchmarks, the best configurations match ESM3-level accuracy while requiring only about one hour of fine-tuning. The model is released under a CC BY-NC-SA 4.0 license, and the work is a preprint awaiting peer review.

#Applications

InstructPLM-mu is useful for researchers scoring the effects of point mutations — for stability, fitness, or function — who want structure-aware accuracy without the compute budget of training a native multimodal model. Because it starts from the widely deployed ESM2 and reuses off-the-shelf structural encoders, groups can adapt existing protein-language-model infrastructure to incorporate structure quickly, making it attractive for variant interpretation and protein-engineering triage.

#Impact

The result reframes structure-aware protein modeling as an efficient fine-tuning problem rather than an expensive from-scratch training problem, showing that a short adaptation of ESM2 can rival ESM3 on mutation prediction. Its systematic comparison of fusion and tuning choices also provides practical guidance for others retrofitting structural conditioning onto sequence models. As a recent, non-commercially licensed preprint, its broader adoption and independent validation remain to be established.