Proteins, as the fundamental functional units of cellular life, govern a wide range of biochemical processes, from metabolism to intracellular signaling. Advances in high-throughput sequencing technologies have led to the accumulation of over 250 million protein sequences in databases such as UniProt. However, the functional interpretation of these data lags significantly behind; less than one percent of known proteins possess experimentally validated functional annotations. This situation creates a serious knowledge gap in the fields of molecular biology and drug discovery. Current computational methods either rely on superficial sequence similarity or treat function prediction tasks as independent classification problems. These approaches fail to adequately reflect the integrated reasoning performed by expert biologists, who evaluate multiple data sources—such as sequences, structures, protein domains, and interaction networks—in concert.
In a study conducted by Fallahpour and colleagues, the goal was to develop and evaluate BioReason-Pro, the first multimodal large language model (LLM) for protein function prediction. The central question of the research is whether a language model can generate expert-level structured reasoning traces by integrating biological context with protein representation embeddings. The fact that current methods treat Gene Ontology (GO; a standard dictionary that defines protein functions in a hierarchical order) terms as independent classification targets, and that a fixed ontological vocabulary limits the expressive power for proteins with unique or combinatorial functions, has been identified as the key gaps this study aims to address.
BioReason-Pro is designed as a multimodal large language model built on the Qwen3-4B base model. The model integrates ESM3 protein embeddings, a GO graph encoder, and additional biological context information such as target organism, domain annotations, and protein-protein interactions. GO-GPT, a key input component, is a transformer model that automatically generates GO terms in a hierarchical manner by taking ESM2 protein embeddings and previously generated terms as conditions. The training dataset was constructed from a collection containing 133,492 proteins from 3,135 organisms, featuring experimental GO annotations sourced from UniProt. Over 130,000 synthetic reasoning traces were generated via GPT-5, followed by supervised fine-tuning (SFT) on these traces. The model’s GO term prediction accuracy was evaluated using temporal decomposition according to the CAFA evaluation framework. Additionally, LLM-based predictions were evaluated through a blind human assessment conducted by 27 molecular biologists.
As a result, GO-GPT outperformed the most accessible methods in the CAFA 5 competition, achieving performance in the range of 0.65–0.70 on the weighted F-max metric. It was observed that GO-GPT’s learned organism embeddings reproduced known phylogenetic relationships, and that attention patterns in DNA-binding proteins were significantly concentrated in binding site residues (average AUROC = 0.81 ± 0.06). BioReason-Pro RL achieved an average score of 8.03 out of 10 in the LLM-based evaluation and significantly outperformed the best comparison method, Prot2Text-v2 (4.15 points). In the human expert evaluation, BioReason-Pro SFT annotations were found to be equal to or superior to curated UniProt database entries in 79% of cases. The model also correctly predicted SBP2 as the interaction partner for the selenocysteine-specific elongation factor eEFSec based solely on sequence information; this prediction was supported by attention maps that aligned with the contact interface validated by cryo-electron microscopy structures. The results indicate that structured reasoning can go beyond approaches based solely on sequence similarity to generate insights into protein function, and this framework provides a reference point for future studies aimed at the functional characterization of millions of uncharacterized proteins.
Translated by : Yağmur Aktepe
Editor: Elinsu Ak
Reference: Fallahpour, A., Seyed-Ahmadi, A., Idehpour, P., Ibrahim, O., Gupta, P., Naimer, J., Zhu, K., Shah, A., Ma, S., Adduri, A., Güloğlu, T., Liu, N., Cui, H., Jain, A., de Castro, M., Fallahpour, A., Cembellin-Prieto, A., Stiles, J. S., Nemčko, F., Nevue, A. A., Moon, H. C., Sosnick, L., Markham, O., Duan, H., Lee, M. Y. Y., Salvador, A. F. M., Maddison, C. J., Thaiss, C. A., Ricci-Tam, C., Plosky, B. S., Burke, D. P., Hsu, P. D., Goodarzi, H., & Wang, B. (2026). BioReason-Pro: Advancing protein function prediction with multimodal biological reasoning. bioRxiv. https://doi.org/10.64898/2026.03.19.712954
-Bioinfocodes Scientific News Service-
News articles prepared by our team members, reviewing and compiling scientific research
published in journals with and impact factor greater than 20 (click here for the list)