ProteinReasoner: A Multi-Modal Protein Language Model with Chain-of-Thought Reasoning for Efficient Protein Design

Abstract

Reasoning has emerged as a central capability of large language models, yet how it should be formulated for scientific foundation models remains unclear because scientific knowledge is distributed across interdependent, domain-specific representations. Here we introduce modality-chain reasoning, which organizes representations into ordered computational chains, each conditioning prediction or generation of the next. Based on this principle, we develop ProteinReasoner, a multimodal generative protein foundation model that sequentially connects amino acid sequence, evolutionary constraints and three-dimensional structure within a shared autoregressive architecture. Across zero-shot structure prediction, inverse folding and fitness prediction, ProteinReasoner outperformed two multimodal protein foundation models, while controlled comparisons supported the functional contribution of the modality chain. We further extended this principle beyond pretraining: reorganizing the chain across successive structural states enabled multiple-conformation prediction, while introducing experimental feedback as an additional modality enabled an in-context learning paradigm for protein optimization without target-specific parameter updates. In particular, across thermostability and affinity-maturation evaluations, this paradigm improved over matched fine-tuned models and showed stronger mean performance than target-specific active-learning baselines. These results establish modality-chain reasoning as a unified and effective foundation-modelling strategy in protein science. More broadly, they suggest a general route towards reasoning across interdependent representations in other scientific domains.