aceGenolator enables protein function interpretation using a multimodal large language model fusing genomic and structural interpretation with natural language interaction
Abstract: Background Decoding the genetic code to unveil its genome functionality is a monumental task which would greatly advance the understanding of disease mechanisms and development of targeted treatments. Although large language models (LLMs) have transformed natural language processing across diverse domains, translating the complex language of DNA into human-readable form remains challenging due to genomic data complexity and unexplored regions of the human genome. Current language models either are capable of pro…
Breakage-Reunion-Domain-of-Streptococcus-pneumoniae-Topoisomerase-IV-Crystal-Structure-of-a-Gram-pone.0000301.s003 by Laponogov I, Veselkov D, Sohi M, Pan X, Achari A, Yang C, Ferrara J, Fisher L, Sanderson M. CC BY 2.5 · https://creativecommons.org/licenses/by/2.5
Researchers present Genolator, a multimodal large language model that combines DNA sequence, amino acid sequence, and protein structure embeddings with natural language interaction. Fine-tuned on over 365,000 question-answer pairs generated from abstracted Gene-Ontology terms, the model is evaluated on answering queries about protein subcellular localization, molecular function, and biological processes.
The work matters because it attempts to bridge genomic code and human language, making protein data queryable in natural language to support biological discovery and clinical research. What remains uncertain from the supplied text is how the reported accuracy generalizes beyond GO-derived QA pairs, to unexplored regions of the human genome noted as challenging, and to real-world laboratory or clinical workflows.
- Genolator is described as a multimodal large language model that integrates embeddings from DNA sequences, amino acid sequences, and protein structures with natural language queries.
- Model was fine-tuned on over 365,000 question-answer pairs generated using abstracted Gene-Ontology (GO) terms.
- Evaluation reported high accuracy in confirming or denying protein function associations versus GPT 4.1 and smaller domain-specific protein structure transformer models.
- Authors report analysis of hidden states, attention heads, and an ablation study as evidence for benefit of the multi-modal approach.
Genolator integrates DNA, amino acid and protein structure embeddings with natural language queries and answers questions about protein localization and function with high accuracy, outperforming general LLMs.
The rundown
The system was built by fusing three biological modalities with language: DNA sequence embeddings, amino acid sequence embeddings, and protein structure embeddings, then coupling them to natural language queries. Training used more than 365,000 question-answer pairs derived from abstracted Gene-Ontology terms covering subcellular localization, molecular function, and biological processes.
Reported evaluations focus on confirming or denying protein function associations, where Genolator outperformed openly available allrounder LLMs like GPT 4.1 and smaller domain-specific models integrating knowledge from a protein structure transformer. The authors also describe explorations of hidden states showing biologically and linguistically plausible organization and attention-head analysis supporting the multimodal design.
Sources
- Peer-reviewedGenome Biology2026-09-16
How should this claim be treated?
The debate