TACTIC: An Explainable Multi-Agent Architecture for Classification & Interpretable Reasoning in Spatial Transcriptomics

Abdel Rahman Alsabbagh, Mahmoud Zahran, Ali Balubaid, Sumeer Ahmad Khan, Robert Lehmann, Xabier Martinez de Morentin, Vincenzo Lagani, Narsis A. Kiani, David Gomez-Cabrero, Jesper Tegnér

Keywords: Multi-Agent Systems, Spatial Transcriptomics, Cell Type Classification, Explainable AI, Chain-of-Thought Reasoning, Graph Autoencoders

TL;DR: A multi-agent system using structured dialogue to accurately and interpretably classify cell types in spatial transcriptomics data without needing task-specific fine-tuning.

Abstract: While spatial transcriptomics captures the spatial localization of transcriptional signatures, its analysis demands significant bioinformatics expertise. We investigate whether a collaborative multi-agent system can automate this complex workflow. We present Transcriptomic Agents for Classification via Thoughtful Inference and Coordination (TACTIC), for interpretable cell type annotation in spatial transcriptomics. TACTIC integrates graph autoencoders and large language models in a chain-of-thought architecture featuring specialized agents, a junior and a senior bioinformatician, engaging in structured dialogue to deliver accurate, human-interpretable annotations. Evaluated on MERFISH, MIBI-TOF, and Drosophila Stereo-seq datasets, TACTIC achieves F1-scores of 0.80, 0.94, and 0.46, respectively, without task-specific fine-tuning across diverse platforms. Ablation studies show that agent collaboration enhances interpretability, reinforcing the value of structured reasoning. These results position TACTIC as a generalizable and explainable AI framework for spatial omics, requiring no task-specific fine-tuning.

Submission Number: 96