CNA detection from spatial transcriptomics using SpaCNA

Zihui Zhang, Xiaochen Wang, Hong Xuan, Zijie Jin, Ruibin XibioRxiv 2025 · DOI

TL;DR

This paper introduces SpaCNA, a novel computational framework that leverages spatial transcriptomics data and H&E images to accurately detect copy number alterations (CNAs) within tumors. By integrating spatial context and multi-modal information, SpaCNA significantly improves CNA detection, enabling a deeper understanding of tumor heterogeneity and spatial evolution.

What Problem Does This Paper Address?

Detecting copy number alterations (CNAs) in cancer is crucial for understanding tumor evolution and identifying malignant cell populations. While single-cell sequencing methods have well-established tools for CNA detection, spatial transcriptomics (ST) data presents unique challenges. ST data often suffers from high noise and sparsity, and the "spot" nature of many ST platforms means each spot can contain multiple cells, diluting CNA signals and increasing the chance of false negatives. Furthermore, ST data often includes multiple modalities like spatial coordinates and histology images, which are difficult to integrate effectively for improved CNA inference. Existing methods either ignore spatial context or rely on assumptions that don't hold true at tumor boundaries, leading to less accurate results. This paper aims to bridge this gap by developing a specialized computational framework for accurate CNA detection in ST data.

Methods

SpaCNA employs a multi-module approach:

  1. Gene Expression Enhancement: It starts by smoothing and denoising gene expression data. This involves constructing an adjacency graph of spots based on both spatial proximity and similarity in H&E stained images. This graph is then used to average neighboring spots, amplifying weaker CNA signals.
  2. Chromosome Segmentation: The genome is divided into non-overlapping bins (1 million base pairs), and gene expression is aggregated to the bin level. A baseline expression from non-tumor cells is used to estimate copy number ratios.
  3. CNA State Inference: A Hidden Markov Random Field (HMRF) model is used to infer discrete copy number states (loss, neutral, gain) for each chromosomal segment. This model incorporates spatial similarity, encouraging neighboring spots to have similar CNA states.

SpaCNA also includes downstream modules for:

  • CNA Visualization: Heatmaps for identifying recurrent CNA events.
  • Malignancy Level Estimation: A regression model to estimate the malignancy of each spot.
  • Tumor Boundary Detection: A boundary entropy score to quantify CNA changes between adjacent spots.
  • 3D-ST Application: Adapting the framework for 3D spatial transcriptomics data.

Datasets

The authors evaluated SpaCNA on:

  • Simulated Datasets: Generated under various conditions (high-purity vs. mixture tumor tissue, different dropout rates, dispersion levels, and tumor proportions) to mimic real ST data.
  • Real Tumor Datasets:
    • Liver cancer patient samples (hepatocellular carcinoma, intrahepatic cholangiocarcinoma) with both spatial transcriptomics and whole-exome sequencing (WES) data. WES data processed by FACETS was used as the gold standard.
  • Breast cancer and colorectal cancer datasets.
  • A 3D spatial transcriptomics dataset of human head and neck squamous cell carcinoma (HNSCC) generated by Open-ST.

Key Insights and Ideas

  • Spatial Context is Crucial for CNA Detection: Integrating spatial proximity and histological features significantly improves the accuracy of CNA detection in ST data, especially in complex tumor microenvironments.
  • Multi-modal Integration Enhances Robustness: Combining gene expression with H&E image similarity in the graph construction helps to refine spot relationships and improve CNA inference, particularly at tumor boundaries.
  • SpaCNA Uncovers Spatially Distinct Subclones: The framework can identify spatially segregated tumor subclones with distinct CNA profiles, providing insights into their evolutionary trajectories.
  • CNAs Define Tumor Architecture: SpaCNA's ability to map CNAs spatially allows for precise delineation of tumor core, boundary, and non-tumor regions, offering a more robust way to define tumor structure than marker genes alone.
  • 3D Spatial CNA Landscape: The method can be extended to 3D ST data, revealing spatially coherent CNA patterns across multiple slices and uncovering clone-specific alterations in 3D.

Key Results

  • Superior Performance on Simulated Data: SpaCNA consistently achieved high precision and recall (above 0.95) in high-purity tumor scenarios and outperformed other methods in mixture tumor scenarios, demonstrating robustness to signal dilution.
  • Accurate CNA Detection in Real Tumors: SpaCNA achieved the highest F1 scores in liver cancer datasets compared to WES gold standards and demonstrated superior accuracy in identifying tumor regions across breast and colorectal cancer datasets (highest ROC AUC).
  • Identification of Therapeutic Biomarkers: In HNSCC, SpaCNA identified clone-specific CNAs, including gains on chromosome 3q harboring genes like PODXL2, which is associated with resistance to EGFR-targeted therapy (cetuximab).
  • Elucidation of Tumor Microenvironment: SpaCNA's CNA profiles were linked to distinct tumor microenvironments, such as immune cell infiltration patterns and the presence of cancer-associated fibroblasts.

Limitations and Future Work

  • H&E Image Requirement: While the authors note that the impact of excluding H&E features is not significant in certain samples, the method's full potential is realized when H&E images are available for graph construction.
  • Computational Efficiency: For very high-resolution ST data (millions of cells), merging cells into pseudo-spots is necessary to manage computational resources, which can sacrifice some detection resolution.
  • Future Directions: The authors suggest that SpaCNA's model architecture could be extended to integrate other modalities (e.g., epigenome) by treating CNA as a common hidden state and using a HMRF model to balance different data types. They also envision its broader application in understanding intra-tumor heterogeneity and progression.

Summary updated

Abstract

Abstract Spatial transcriptomics (ST) enables genome-wide profiling of gene expression while preserving spatial context, yet accurate detection of copy number alterations (CNAs) in tumor ST data remain challenging. Here we present SpaCNA, a spatial-aware computational framework that integrates multi-modal spatial information to improve CNA detection. SpaCNA constructs adjacency graphs using spatial proximity and H&E stained image similarity, refines the raw gene expressions, and implements a Hidden Markov Random Field model for robust CNA state inference. Through extensive benchmarking on simulated data and multi-cancer cohorts, SpaCNA demonstrates superior accuracy and outperforms existing methods in CNA detection and tumor region identification. In real-world applications to breast cancer and colorectal cancer, SpaCNA reveals spatially distinct subclones with context-dependent interactions within the microenvironment. Additionally, applied to a 3D ST dataset of head and neck squamous cell carcinoma, SpaCNA uncovers clone-specific CNAs associated with therapeutic resistance biomarkers across multiple slices. By facilitating precise spatial mapping of CNAs and tumor architecture, SpaCNA significantly enhances our understanding of intratumoral heterogeneity and spatial evolutionary patterns.
BibTeX
@misc{zhang2025cna,
  author = {Zihui Zhang and Xiaochen Wang and Hong Xuan and Zijie Jin and Ruibin Xi},
  title = {{CNA detection from spatial transcriptomics using SpaCNA}},
  year = {2025},
  doi = {10.1101/2025.07.08.663688},
  howpublished = {bioRxiv},
}
Figure