Assessment of LLMs on basic bioinformatics tasks requiring multi-step reasoning and code writing
5 affiliations
- 1University of Michigan College of Engineering, , Ann Arbor, MI, 48109-2054,
- 2University of Michigan Department of Molecular, Cellular and Developmental Biology, , Ann Arbor, MI, 48109-2054,
- 3University of Michigan Medical School Department of Learning Health Sciences, , Ann Arbor, MI, 48109-2054,
- 4University of Michigan Medical School Department of Pharmacology, , Ann Arbor, MI, 48109-2054,
- 5University of Michigan Center for Global Health Equity, , Ann Arbor, MI, 48109-2054,
Abstract
Motivation: Large Language Models (LLMs), AI agents and co-scientists promise to accelerate scientific discovery across fields ranging from chemistry to biology. Bioinformatics—the analysis of DNA, RNA and protein sequences plays a crucial role in biological research and is especially amenable to AI-driven automation given its computational nature. Here, we assess the bioinformatics capabilities of popular general-purpose LLMs on a set of tasks covering basic analytical questions that include code writing and multi-step reasoning in the domain.
Results: Utilizing questions from Rosalind, a bioinformatics educational platform, we compare the performance of the LLMs on 104 questions relative to expert ground truth and 68 760 humans with varying levels of expertise. Among the frontier LLMs evaluated (GPT-5.2, GPT-4o, GPT-3.5, Llama-3-70B, Llama-4 and Claude Sonnet 4.6), GPT-5.2 exhibited top performance, providing correct answers for 80/104 (77%) questions followed by Claude Sonnet 4.6. LLM performance varied widely depending on the biological categories of questions. While the latest generation of frontier LLM models showed overall improvements in performance, their performance in specific categories of biological questions was lower than first generation models. Overall, LLMs performance mirrored that of humans with lower performance in tasks in which humans had low performance and vice versa. However, LLMs also failed in some instances where most humans were correct and, in a few cases, LLMs excelled where most humans failed. To the best of our knowledge, this presents the first assessment of general purpose LLMs on basic bioinformatics tasks in distinct areas relative to both expert and non-expert ground truths.
Availability and Implementation: The complete analysis was implemented in Python; all code is available at https://github.com/SiwoResearch/lllm-bioinformatics.
Related
- The Reasoning Paradox in Scientific Sandboxes: Why Extended Reasoning Fails in Autonomous Bioinformatics and How to Fix It.
- Valid and Reliable? Basic and Expanded Recommendations for Psychometric Reporting and Quality Assessment
- Correlation assessment of SARS-CoV-2 variants and their subvariants present in clinical and wastewater samples in Oregon, USA (February 7, 2021 - February 26, 2022) using the Freyja bioinformatics approach
- Assessment of diagnostic reasoning in acute vertigo using vignette-based tools: A cross-sectional comparison between general practitioners and final-year medical students
- ‘(m6)A’ stands for ‘autoimmunity’: reading, writing, and erasing RNA modifications during inflammation
BibTeX
@article{rajesh2026assessment,
title = {{Assessment of LLMs on basic bioinformatics tasks requiring multi-step reasoning and code writing}},
author = {Varsha Rajesh and Maceo Blumberg and Shreya Jain and Geoffrey H Siwo},
journal = {Bioinformatics},
year = {2026},
month = oct,
doi = {10.1093/bioinformatics/btag749},
url = {https://academic.oup.com/bioinformatics/article/doi/10.1093/bioinformatics/btag749/9005024},
note = {CC BY},
}
RIS
TY - JOUR
TI - Assessment of LLMs on basic bioinformatics tasks requiring multi-step reasoning and code writing
AU - Varsha Rajesh
AU - Maceo Blumberg
AU - Shreya Jain
AU - Geoffrey H Siwo
PY - 2026
DA - 2026/10/30
JO - Bioinformatics
DO - 10.1093/bioinformatics/btag749
UR - https://academic.oup.com/bioinformatics/article/doi/10.1093/bioinformatics/btag749/9005024
AB - Motivation: Large Language Models (LLMs), AI agents and co-scientists promise to accelerate scientific discovery across fields ranging from chemistry to biology. Bioinformatics—the analysis of DNA, RNA and protein sequences plays a crucial role in biological research and is especially amenable to AI-driven automation given its computational nature. Here, we assess the bioinformatics capabilities of popular general-purpose LLMs on a set of tasks covering basic analytical questions that include code writing and multi-step reasoning in the domain. Results: Utilizing questions from Rosalind, a bioinformatics educational platform, we compare the performance of the LLMs on 104 questions relative to expert ground truth and 68 760 humans with varying levels of expertise. Among the frontier LLMs evaluated (GPT-5.2, GPT-4o, GPT-3.5, Llama-3-70B, Llama-4 and Claude Sonnet 4.6), GPT-5.2 exhibited top performance, providing correct answers for 80/104 (77%) questions followed by Claude Sonnet 4.6. LLM performance varied widely depending on the biological categories of questions. While the latest generation of frontier LLM models showed overall improvements in performance, their performance in specific categories of biological questions was lower than first generation models. Overall, LLMs performance mirrored that of humans with lower performance in tasks in which humans had low performance and vice versa. However, LLMs also failed in some instances where most humans were correct and, in a few cases, LLMs excelled where most humans failed. To the best of our knowledge, this presents the first assessment of general purpose LLMs on basic bioinformatics tasks in distinct areas relative to both expert and non-expert ground truths. Availability and Implementation: The complete analysis was implemented in Python; all code is available at https://github.com/SiwoResearch/lllm-bioinformatics.
SN - 1367-4803
SN - 1367-4811
ER -