CancerSTFormer

CancerSTFormer is a pair of spatially aware transcriptomic foundation models from Baylor College of Medicine, released as a bioRxiv preprint in late 2025 and updated in March 2026. It is designed to study the niche-level behavior of tumors — how cells coordinate gene expression within their spatial neighborhoods — from spot-resolution spatial transcriptomics data. Because biologically meaningful niches exist at different length scales, the authors train two complementary models: a 50um-Local model that captures fine-grained neighborhood structure and a 250um-Extended model that captures broader tissue context.

Spot-resolution platforms such as Visium, DBiT-seq, and Slide-seq measure expression over small tissue regions rather than single cells, and modeling the spatial organization of these spots within tumors poses distinct challenges from single-cell foundation models. CancerSTFormer is purpose-built for this setting and for cancer specifically, learning representations that recover the regulatory logic of tumor microenvironments.

The model demonstrates that a transcriptomic foundation model can recover ligand- target relationships, identify niche-specific and organ-specific expression programs, and — in a zero-shot setting — recapitulate spatial perturbation effects, including the regulatory consequences of immune-checkpoint blockade and other targeted therapies.

#Key Features

• Multi-scale niche modeling: Two paired models (50um-Local and 250um-Extended) capture spatial neighborhoods at complementary length scales.
• Spot-resolution and cancer-focused: Trained on sequencing-based spot-resolution spatial transcriptomics from human cancer studies rather than single-cell data.
• Ligand-target recovery: Recovers ligand-target gene relationships and niche-specific differentially expressed genes from spatial context.
• Zero-shot perturbation prediction: Recapitulates spatial Perturb-map-style perturbation effects and the regulatory responses of targeted therapies without task-specific training.
• Multi-platform training: Spans Visium, DBiT-seq, and Slide-seq, improving generalization across spatial assays.

#Technical Details

CancerSTFormer is a transformer-based, sequencing-based spatial transcriptomic foundation model trained on over 1,000,000 spots drawn from 50 human cancer studies and 511 samples, spanning multiple spot-resolution platforms including Visium, DBiT-seq, and Slide-seq. The two model variants differ in the spatial radius of neighborhood context they encode — 50um (Local) versus 250um (Extended) — enabling analysis of niches at different scales. Reported capabilities include recovery of ligand-target relationships, niche and organ-specific differentially expressed genes, metastasis-associated genes, and zero-shot recapitulation of spatial perturbation effects for immune-checkpoint blockade and other targeted therapies.

#Applications

CancerSTFormer supports cancer researchers studying the tumor microenvironment, spatial organization of immune and malignant cells, and how niches respond to targeted and immunotherapies. Its multi-scale design lets users probe both fine-grained cell-cell signaling and broader tissue architecture, and its zero-shot perturbation predictions can help prioritize hypotheses about therapeutic response before costly spatial perturbation experiments.

#Impact

CancerSTFormer extends the foundation-model paradigm from single-cell transcriptomics to spot-resolution spatial transcriptomics in cancer, where modeling spatial niches at multiple scales is essential. Its ability to recover ligand-target relationships and recapitulate perturbation effects zero-shot points toward in silico exploration of how tumor niches respond to therapy. Adoption is constrained by access: the preprint is released under a restrictive CC-BY-NC-ND license, and no public code or model weights have been confirmed, limiting independent validation and reuse.