Bridging large language model-based agents and workflow languages for automated multi-perspective analysis of single-cell RNA sequencing data
Abstract
Motivation: In single-cell RNA sequencing data analysis, the same dataset can be examined from multiple complementary perspectives, each requiring different analytical strategies and potentially leading to distinct biological interpretations. Systematically exploring such diverse strategies in a reproducible and scalable manner remains challenging.
Results: We developed a framework combining large language model-based agents with workflow languages. Given an objective, the framework generates analytical strategies, converts each into a workflow plan, verifies that each plan reflects its intended strategy, executes accepted workflows in parallel, and integrates the outputs into a final report. Evaluation across three datasets and three large language models showed that plan validation improved the executable analysis intent fulfillment rate from 0.061 to 0.378. Diversity of analytical themes was established at as few as 10 generated strategies, while the number of distinct executable analysis paths grew continuously, reaching 59 unique paths at 100 generated strategies. Parallelization of both plan generation and workflow execution achieved up to a 29.9-fold speedup in workflow execution over sequential execution, with plan generation completing in near-constant time regardless of portfolio size.
Availability and implementation: Source code is available at https://github.com/bioinfo-tsukuba/mpdas_paper under the MIT license.
Related
- From conservation scores to language models: a feature engineering perspective on machine learning for genetic variant interpretation
- From Language-Cognition Coupling to AI Foundations: An Evolutionary Perspective on 13,415 SNVs in 33 Language/Cognition Genes
- Learning Latent Protein Languages for Autoregressive Generation
- Bridging the antiviral gap: A BAC-based non-infectious SARS-CoV-2 replicon reveals potent antisense gapmers
- Targeted protein degradation: bridging chemical biology and clinical translation
BibTeX
@article{hayakawa2026bridging,
title = {{Bridging large language model-based agents and workflow languages for automated multi-perspective analysis of single-cell RNA sequencing data}},
author = {Yoshinori Hayakawa and Haruka Ozaki},
journal = {Bioinformatics Advances},
year = {2026},
month = oct,
doi = {10.1093/bioadv/vbag301},
url = {https://doi.org/10.1093/bioadv/vbag301},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Bridging large language model-based agents and workflow languages for automated multi-perspective analysis of single-cell RNA sequencing data
AU - Yoshinori Hayakawa
AU - Haruka Ozaki
PY - 2026
DA - 2026/10/08
JO - Bioinformatics Advances
DO - 10.1093/bioadv/vbag301
UR - https://doi.org/10.1093/bioadv/vbag301
AB - Motivation: In single-cell RNA sequencing data analysis, the same dataset can be examined from multiple complementary perspectives, each requiring different analytical strategies and potentially leading to distinct biological interpretations. Systematically exploring such diverse strategies in a reproducible and scalable manner remains challenging. Results: We developed a framework combining large language model-based agents with workflow languages. Given an objective, the framework generates analytical strategies, converts each into a workflow plan, verifies that each plan reflects its intended strategy, executes accepted workflows in parallel, and integrates the outputs into a final report. Evaluation across three datasets and three large language models showed that plan validation improved the executable analysis intent fulfillment rate from 0.061 to 0.378. Diversity of analytical themes was established at as few as 10 generated strategies, while the number of distinct executable analysis paths grew continuously, reaching 59 unique paths at 100 generated strategies. Parallelization of both plan generation and workflow execution achieved up to a 29.9-fold speedup in workflow execution over sequential execution, with plan generation completing in near-constant time regardless of portfolio size. Availability and implementation: Source code is available at https://github.com/bioinfo-tsukuba/mpdas_paper under the MIT license.
SN - 2635-0041
ER -