Abstract A030: Mutation2Text: A unified protein and text language model for explaining mutation effects

Abstract Background: Understanding the functional consequences of protein mutations is crucial for diagnosing, treating, and preventing diseases such as cancer. Even a single amino acid change can disrupt cellular pathways and influence therapeutic outcomes. However, existing protein language models (PLMs) remain limited by their black-box nature, their inability to process full-length proteins without truncation, and their primary focus on single-nucleotide variants. More complex alterations such as insertions, deletions, and frameshifts remain underexplored. Methods: We developed Mutation2Text, a multimodal generative PLM designed to generate human-understandable, rationale-based explanations for diverse mutation types. The architecture integrates a gated cross-attention mechanism to explicitly contrast wild-type and mutated sequences, as well as a Perceiver Resampler for length-invariant encoding of entire proteins. To support training and evaluation, we constructed Mutation2TextQA, the largest mutation interpretation dataset to date, comprising question–answer pairs mined from biomedical literature and curated for lexical and semantic diversity. Results: Mutation2Text sets a new state-of-the-art across multiple benchmarks for mutation explanation, pathogenicity prediction, and disease association. Importantly, our evaluation revealed a significant discrepancy between the model’s highly accurate internal representations (0.96 AUC on pathogenicity prediction) and its final text-based output (0.85 AUC). This gap underscores a fundamental “lost-in-translation” challenge for generative AI in science, where reasoning captured internally by the model may not be fully expressed in natural language output. Conclusions: Mutation2Text represents a step toward interpretable and rationale-driven generative models for protein mutations, addressing a central need in precision medicine and cancer research. All code, data, and pretrained models are released at https://anonymous.4open.science/r/mutation2text-94A6/ to accelerate community-driven research and applications. Citation Format: Oladimeji Macaulay, Ala Jararweh, Yue Hu, David Arredondo, Shrey Poshiya, Luis Tafoya, Robert Mccourt, Kushal Virupakshappa, Avinash Sahu. Mutation2Text: A unified protein and text language model for explaining mutation effects [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Cancer Evolution: The Dynamics of Progression and Persistence; 2025 Dec 4-6; Albuquerque, NM. Philadelphia (PA): AACR; Cancer Res 2025;85(23_Suppl):Abstract nr A030.