CellQuery-ST: Cell-Aware Query Grounding for Single-Cell and Neighborhood Retrieval from Histology

CellQuery-ST addresses a key gap in computational pathology: while current models can predict gene expression or answer coarse slide-level questions, they do not support cell-aware querying of histology images. In practice, researchers and clinicians want to ask questions such as: Where are B-cell follicles? Which regions show inflammatory myeloid activity? Which neighbourhoods resemble a vascular niche? This project aims to make such biologically grounded querying possible on new histology slides. To solve this, I will build a cell-aware query grounding framework that uses spatial omics data for supervision during training but supports image-only inference at test time. Each slide will be preprocessed into a spatial index of cells, patches, and neighbourhoods, and natural-language queries will be matched against this index to retrieve or score relevant spatial evidence. The system will combine spatial pathology data with CellNet, an existing paired single-cell and language resource, to connect text queries with cell identities, cell states, and higher-level biological concepts. The main deliverables will be: (1) a benchmark covering four task families—cell type grounding, cell state/programme grounding, spatial niche grounding, and communication hotspot grounding; (2) a reusable slide indexing and retrieval pipeline for histology images; (3) reproducible baseline models and evaluation utilities for seen/unseen query generalisation; and (4) documentation and tutorial notebooks showing how to preprocess a new slide, issue queries, and evaluate results.