PhenoGraph: A Multi-Agent Framework for Phenotype-driven Discovery in Spatial Transcriptomics Data Augmented with Knowledge Graphs

Spatial transcriptomics (ST) provides powerful insights into gene expression patterns within tissue structures, enabling the discovery of molecular mechanisms in complex tumor microenvironments (TMEs). Phenotype-based discovery in ST data holds transformative potential for linking spatial molecular expression patterns to clinical outcomes; however, appropriate ST data analysis remains fundamentally fragmented and highly labor-intensive. Due to its limited scalability considering the size of typical ST data in large cohorts, researchers must rely on other phenotype-annotated omics data modalities (e.g. bulk RNA-sequencing) and align them with ST data to extract clinically meaningful spatial patterns. Yet, this process requires manually identifying relevant cohorts, aligning multi-modal data, selecting and tuning analysis pipelines, and interpreting results—typically without any built-in support for biologically context-aware reasoning. In this paper, we present \textit{PhenoGraph}, a large language model (LLM) based multi-agent system, that automates the full pipeline for phenotype-driven ST data analysis, augmented by biological knowledge graphs for enhanced interpretability. Built on a modular agent architecture, PhenoGraph dynamically selects, executes, and corrects phenotype analysis pipelines based on user-defined queries. We showcase the flexibility and effectiveness of PhenoGraph across a variety of TME ST datasets and phenotype classes, highlighting its potential to enhance biological discovery efficacy.