The integration of artificial intelligence systems into scientific research processes has become a subject of growing interest in the life sciences in recent years. Research workflows in fields such as biology, drug discovery, and translational medicine involve multi-step processes including extensive literature reviews, access to specialized databases, interpretation of experimental data, and hypothesis generation. It has been reported that the process from target discovery to regulatory approval in the United States takes an average of 10 to 15 years. This duration is due not only to fundamental scientific challenges but also to the fragmented and hard-to-scale nature of research workflows. The performance of AI models in tasks requiring multi-step reasoning on molecules, proteins, genes, and biological pathways remains a prominent research question in the field.
Developed by the OpenAI research team, the study introduces a reasoning model called GPT-Rosalind, which is tailored specifically for life sciences research. The model was named after Rosalind Franklin, who significantly contributed to elucidating the structure of DNA and played a crucial role in establishing the foundations of modern molecular biology. The initiative was driven by the observation that general-purpose language models often fall short in domain-specific tasks such as chemical reaction mechanisms, protein structure, mutation effects, and the phylogenetic interpretation of DNA sequences. In this context, the goal was to evaluate a domain-specific model capable of providing higher performance in scientific tool usage and multi-step research workflows, including literature reviews, sequence-to-function interpretation, experimental planning, and data analysis.
The model was evaluated across a range of competencies deemed essential for scientific discovery. The evaluation covered sub-domains such as organic chemistry, biochemistry and protein understanding, genomics, experimental design and analysis, and tool usage. The model’s performance was compared with earlier models like GPT-5, GPT-5.2, and GPT-5.4. Furthermore, researchers utilized publicly available benchmark sets: BixBench, which includes bioinformatics and data analysis tasks, and LABBench2, which encompasses research tasks such as literature access, database utilization, sequence manipulation, and protocol design. In a collaboration with Dyno Therapeutics, a company working on AI-based gene therapies, the model was tested on unpublished RNA sequences without data contamination. Its performance in sequence-to-function prediction and sequence generation tasks was compared against the historical scores of 57 human experts.
The findings indicate that the GPT-Rosalind model achieved higher scores in chemistry, biochemistry and protein understanding, phylogenetics, and experimental design and analysis compared to previous models. In the BixBench evaluation, it achieved leading performance among models with published scores. On LABBench2, the model outperformed GPT-5.4 in 6 out of 11 tasks, with the most notable improvement observed in the CloningQA task, which requires the end-to-end design of DNA and enzyme reagents for molecular cloning protocols. In the evaluation with Dyno Therapeutics, based on the top ten submissions, the model performed above the 95th percentile of human experts in the RNA sequence-to-function prediction task and at approximately the 84th percentile in the sequence generation task. These results demonstrate the potential of domain-specific models to support researchers in multi-step, tool-intensive scientific workflows. However, it is emphasized that additional independent evaluations and studies in real research environments are necessary before these results can be directly applied to clinical practice.
Translated by: Zeynep Naz Darbaroğlu
Editor: Elinsu Ak
Reference: Gor, K., Geissen, E. M., & Duss, O. (2026). Functional characterization of dynamic nascent RNA folding ensembles in real time. Science Advances, 12(12), eaec4037. https://doi.org/10.1126/sciadv.aec4037
-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)