SPaRTA: Spatial Reasoning in Text-to-SQL for Omics Analysis
- 1Vanderbilt University
Abstract
Spatial-omics technologies provide clinicians, pathologists, and biomedical researchers with high-resolution maps of transcriptomic, proteomic, and metabolomic activity within tissues. However, querying these complex datasets often requires specialized computational expertise, limiting their accessibility to domain experts without programming experience. In the current manuscript, we develop SPaRTA (Spatial Reasoning in Text-to-SQL for Omics Analysis), a natural-language interface that translates free-text questions into executable spatial SQL queries over such data. We release SPaRTA-Bench, comprising dataset-specific database schemas, explicit query conventions, and 7,500 question-SQL pairs validated through LLM-assisted verification and targeted human review, spanning breast cancer, pancreatic cancer, and melanoma tissue sections, and identify 15 failure modes in which a query runs but returns a wrong answer. Open-weight large language models (LLMs) of the Llama, Mistral, Qwen, Gemma and Phi families reach 10-40% execution accuracy zero-shot, and at most 61% with reasoning and execution feedback. Fine-tuning raises them to 63-77%, and they lose at most 6 percentage points on a tissue unseen in training. Execution feedback repairs queries that fail to run, but rarely queries that run and return a wrong answer.
Related
- OmicsLM: A Multimodal Large Language Model for Multi-Sample Omics Reasoning
- Spatial multi-omics reveals novel vascular adipocyte and B cell biology in human abdominal aortic aneurysm
- Integrating multi-omics analysis reveals shared genetic architecture and spatial correlations between heart failure and osteoporosis
- Scale-resolved correlation functions unify spatial omics across platforms and resolutions
- REZCUE: Rescuing 3D spatial omics from archived FFPE tissue
BibTeX
@misc{dutta2026sparta,
title = {{SPaRTA: Spatial Reasoning in Text-to-SQL for Omics Analysis}},
author = {A. Dutta and S. Ghosh and K. Prado and A. KhudaBukhsh and H. Sarkar},
howpublished = {bioRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.31.763119},
url = {https://www.biorxiv.org/content/10.64898/2026.10.31.763119v1},
note = {CC BY-NC},
}
RIS
TY - UNPB
TI - SPaRTA: Spatial Reasoning in Text-to-SQL for Omics Analysis
AU - A. Dutta
AU - S. Ghosh
AU - K. Prado
AU - A. KhudaBukhsh
AU - H. Sarkar
PY - 2026
DA - 2026/10/31
PB - bioRxiv
DO - 10.64898/2026.10.31.763119
UR - https://www.biorxiv.org/content/10.64898/2026.10.31.763119v1
AB - Spatial-omics technologies provide clinicians, pathologists, and biomedical researchers with high-resolution maps of transcriptomic, proteomic, and metabolomic activity within tissues. However, querying these complex datasets often requires specialized computational expertise, limiting their accessibility to domain experts without programming experience. In the current manuscript, we develop SPaRTA (Spatial Reasoning in Text-to-SQL for Omics Analysis), a natural-language interface that translates free-text questions into executable spatial SQL queries over such data. We release SPaRTA-Bench, comprising dataset-specific database schemas, explicit query conventions, and 7,500 question-SQL pairs validated through LLM-assisted verification and targeted human review, spanning breast cancer, pancreatic cancer, and melanoma tissue sections, and identify 15 failure modes in which a query runs but returns a wrong answer. Open-weight large language models (LLMs) of the Llama, Mistral, Qwen, Gemma and Phi families reach 10-40% execution accuracy zero-shot, and at most 61% with reasoning and execution feedback. Fine-tuning raises them to 63-77%, and they lose at most 6 percentage points on a tissue unseen in training. Execution feedback repairs queries that fail to run, but rarely queries that run and return a wrong answer.
ER -