Full text
1,981 characters
· extracted from
oa-doi-fallback
· click to expand
Abstract
Foundation models enable knowledge transfer across data modalities and tasks, yet foundation models for spatial biology remain in their early stages, largely centered on encoding single-cell representations in spatial context without fully integrating transcriptomic and morphological information to delineate functional niches. Here we introduce SpatialFusion, a lightweight multimodal foundation model that identifies biologically coherent microenvironments defined by distinct pathway activation patterns rather than spatial proximity alone. SpatialFusion integrates paired histopathology, gene expression, and inferred pathway activity into a unified representation. Compared with two specialist niche-detection methods and four spatial foundation models, SpatialFusion performs competitively and consistently resolves fine-grained spatial niches with unique pathway-level signatures. Applying the model to two Visium HD cohorts uncovered a pre-malignant niche in morphologically normal mucosa adjacent to colorectal tumors and revealed distinct malignant microenvironments in non-small cell lung cancer that were predictive of tumor stage. Overall, SpatialFusion offers a versatile framework for multimodal spatial analysis, enabling the discovery of new morpho-molecular niches with significant biological and clinical relevance.
Competing Interest Statement
E.M.V.A., advisory/consulting: Enara Bio, Manifold Bio, Monte Rosa, Novartis Institute for Biomedical Research, Serinus Bio, and TracerDx. Research support: Novartis, BMS, Sanofi, and NextPoint. Equity: Tango Therapeutics, Genome Medical, Genomic Life, Enara Bio, Manifold Bio, Microsoft, Monte Rosa, Riva Therapeutics, Serinus Bio, Syapse, and TracerDx. Travel reimbursement: none. Patents: institutional patents filed on chromatin mutations and immunotherapy response, and methods for clinical interpretation; intermittent legal consulting on patents for Foaley & Hoag. Editorial boards: Science Advances.
Text is read by the "Ask this paper" AI Q&A widget below.
Extraction quality varies by source — PMC NXML preserves structure
cleanly, OA-HTML may include some navigation residue, and OA-PDF can
have broken hyphenation. The publisher copy
(via DOI)
is the canonical version.