The problem Genolator addresses is one of the most stubborn bottlenecks in modern biology. Sequencing a genome has become fast and cheap, but interpreting it has not. Vast stretches of the human genome remain functionally unexplored, and even for the genes we do know, assigning a protein to a role—where it operates inside the cell, which molecular tasks it performs, and which biological processes it participates in—still relies heavily on laborious experimental work and expert curation. The Gene Ontology framework, the standard vocabulary researchers use to describe protein functions, has grown enormous precisely because this annotation task is so demanding. Every uncharacterized protein is a potential blind spot in our understanding of disease mechanisms and, by extension, in the development of targeted treatments.

Large language models have transformed how machines handle human text, but genomic data has stubbornly resisted the same revolution. The reason, the Aachen team argues in the study, is that existing models are divided into two camps that rarely speak to each other. Some language models can process natural language; others can process genomic code; models that genuinely fuse both modalities have been largely lacking. Genolator was built explicitly to close that gap. It integrates embedding representations derived from three very different kinds of biological data—DNA sequences, amino acid sequences, and protein structures—with the ability to receive and answer natural language queries. In other words, the model holds the molecular alphabet and the human dictionary in the same computational mind.

The engineering behind that fusion is substantial. The researchers fine-tuned Genolator on more than 365,000 question-answer pairs, generated using abstracted Gene Ontology terms. This training corpus teaches the model to respond to queries spanning three core categories of protein function: subcellular localization, meaning where in the cell a protein is found; molecular function, meaning the specific biochemical activity it carries out; and biological processes, meaning the larger cellular programs it contributes to. By abstracting the Gene Ontology terms, the team could turn the structured, hierarchical vocabulary of functional genomics into the kind of question-answer dialogue that language models learn best from.

How well does it work? The evaluation results reported in the study show that Genolator achieves high accuracy in confirming or denying associations between proteins and functions—the fundamental yes-or-no question at the heart of functional annotation. Crucially, the model was benchmarked against serious competition. It outperformed openly available general-purpose large language models such as GPT 4.1, demonstrating that general AI fluency in English does not automatically translate into biological fluency. It also outperformed smaller domain-specific models that incorporate knowledge from a protein structure transformer, indicating that the multimodal integration itself, rather than sheer scale, is what delivers the advantage.

Perhaps the most intriguing part of the study is what the researchers found when they looked inside the model. By examining the hidden states—the internal representational layers that language models build as they process input—the team uncovered an organization that is both biologically and linguistically plausible. That finding matters because a common criticism of large language models in science is that they act as inscrutable black boxes, producing answers without any interpretable internal logic. If Genolator’s learned representations mirror meaningful biological and linguistic structure, it suggests the model is not merely memorizing training pairs but is organizing its knowledge in ways that reflect real relationships among proteins, functions, and the words used to describe them.

The Aachen group went further, dissecting the attention heads—the components of the underlying language model that decide which parts of the input to focus on—and conducting an ablation study, in which parts of the system are systematically removed to test their contribution. Both lines of analysis provided evidence for a genuine benefit of the multimodal approach. Removing or ignoring the genomic and structural inputs diminished performance, supporting the argument that fusing sequence data, structural data, and language is not a gimmick but the source of the model’s power. For a field still debating whether multimodal biology models justify their complexity, this is a meaningful data point.

The practical implications reach into the clinic. The study’s authors are affiliated with the Center for Human Genetics and Genomic Medicine at Uniklinik RWTH Aachen, and the work was funded in part by the START Program and the Clinician Scientist program of the university’s Faculty of Medicine. That institutional context hints at the intended use case: helping clinicians and researchers interpret genetic variants and uncharacterized proteins encountered in patient genomes. When a diagnostician encounters a gene of unknown function in a patient with an unexplained condition, a tool that can rapidly and accurately answer natural-language questions about that gene’s protein product could accelerate the path from raw sequence to mechanistic insight—and, ultimately, to targeted treatment.

It is worth being precise about what Genolator does and does not claim. The model answers queries about protein subcellular localization, molecular function, and biological processes, and its evaluated strength lies in confirming or denying protein-function associations. It is a step toward bridging genomic code and human language, as the authors conclude, rather than a finished oracle of the genome. The training data was generated from Gene Ontology terms, so the model’s competence is bounded by the scope of that vocabulary and the proteins represented within it. And like all language models, its judgments would need to inform, not replace, experimental validation. Still, the direction of travel is clear: a tool that lets biologists converse with molecular data in their own language, rather than forcing every question through specialized computational pipelines.

The open-access publication in Genome Biology, with a peer-review process led by Casey Greene and Tim Sands, means the full study, including supplementary analyses such as t-SNE visualizations of the model’s internal representations, cosine distance heatmaps, and a small case study utilizing GPT 4.1, is freely available for scrutiny and reuse. The research was made possible in part by a cloud environment established in collaboration with the IT Department at Uniklinik RWTH Aachen, and the published version carries a Creative Commons Attribution 4.0 license. As artificial intelligence and genomics continue to converge, Genolator stands as a template for what that convergence might look like: not a machine that replaces biologists, but one that finally speaks their language and the genome’s at the same time. If the multimodal approach holds up as larger models and richer datasets arrive, the monumental task of decoding genetic functionality may become less monumental—and far more conversational—than it has ever been.

Subject of Research: A multimodal large language model for protein function prediction and genomic interpretation

Article Title: Genolator enables protein function interpretation using a multimodal large language model fusing genomic and structural interpretation with natural language interaction

Article References: Danner, M., Islam, T., Begemann, M., Kraft, F., Elbracht, M., Kurth, I., & Krause, J. (2026). Genolator enables protein function interpretation using a multimodal large language model fusing genomic and structural interpretation with natural language interaction. Genome Biology, 27(1), Article 290. https://doi.org/10.1186/s13059-026-04274-w

Image Credits: AI Generated

DOI: 10.1186/s13059-026-04274-w

Keywords: Genolator, large language models, protein function prediction, Gene Ontology, genomics, multimodal AI, protein structure, functional annotation, genome informatics, natural language processing, computational biology, Genome Biology