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

Multimodal large language model fusing embeddings of DNA sequences, amino acid sequences and protein structures with natural language queries

Language models today handle either natural language or genomic code, rarely both. Genolator is a multimodal large language model that integrates embeddings of DNA sequences, amino acid sequences and protein structures with natural language queries. Fine-tuned on over 365,000 question-answer pairs generated from abstracted Gene Ontology terms, it answers queries about protein subcellular localization, molecular function and biological processes. The authors report high accuracy in confirming or denying protein function associations, outperforming openly available generalist models such as GPT-4.1 and smaller domain-specific models incorporating a protein structure transformer. Exploration of hidden states, attention head analysis and an ablation study support the benefit of the multimodal approach.

Synthesis written by Geno'X. For the full original abstract, please refer to the source publication.

Fusing sequence, structure and language is architecturally interesting, but the evaluation never leaves Gene Ontology territory: no patient variant, no diagnostic task. Beating GPT-4.1 on questions derived from functional annotations says nothing about settling a gene-disease hypothesis, which is the question actually asked in front of an exome. Worth following as an annotation component, not as clinical interpretation support.

Analysis by Dr Thibaut Benquey