Tumor heterogeneity -- the genetic and molecular diversity that exists within a single cancer -- is a fundamental obstacle to effective treatment. Different regions of the same tumor can have different genetic profiles, enabling cancer cells in some areas to survive therapies that kill others. Understanding this spatial complexity is essential for designing targeted treatments.
Copy number variations (CNVs) -- gains or losses of sections of chromosomal DNA -- are among the most important genomic alterations driving cancer progression. They influence cell proliferation rates, response to therapy, and mechanisms of drug resistance. Mapping CNVs across the physical space of a tumor would provide a comprehensive picture of its evolution.
Traditional methods for detecting CNVs (whole-genome and whole-exome sequencing) analyze bulk tissue -- averaging signals from millions of mixed cells -- and therefore cannot reveal the spatial arrangement of different genetic clones within a tumor. Single-cell sequencing captures individual cell profiles but loses information about where those cells sit in the tissue architecture.
Spatial transcriptomics technologies (such as 10x Genomics Visium) represent a breakthrough: they measure gene expression at hundreds to thousands of defined locations (called spots) across a tissue section, preserving both molecular information and spatial context. However, existing tools for inferring CNVs from spatial data were originally designed for single-cell data and fail to exploit the spatial relationships between neighboring tissue spots.
The researchers developed SCOIGET (Spatial COpy number Inference by Graph on Evolution of Tumor), a novel computational framework that treats the spatial transcriptomics data as a graph -- a network where each tissue spot is a node and edges connect neighboring spots based on their physical proximity and gene expression similarity.
The core of SCOIGET is a Graph Neural Network (GNN) with Graph Attention Network (GAT) layers. Unlike standard neural networks, GNNs pass information between connected nodes, allowing each spot's representation to be influenced by its spatial neighbors. The attention mechanism means the model learns to weight the influence of each neighbor differently -- more similar or informative neighbors contribute more.
The architecture uses an encoder-decoder design. The encoder uses three GAT layers to learn compact representations capturing both transcriptomic profiles and spatial context. A specialized component called the CNEncoder then applies a Hidden Markov Model (HMM) to these representations to detect chromosomal CNV segments -- regions of consistent gain or loss across the genome. HMMs are well-suited to this task because CNVs typically span contiguous genomic regions.
Training proceeds in two phases. First, the model learns from raw gene expression data organized into 25-gene bins. Then, the initial CNV estimates are fed back as additional features to refine the spatial graph, and the model is retrained on this enriched graph. This iterative approach progressively improves the spatial accuracy of CNV localization. Total training time on an NVIDIA A100 GPU is approximately 35 minutes per dataset.
SCOIGET was evaluated on eight colorectal cancer samples from the HTAN-WUSTL dataset, with whole-exome sequencing (WES) data providing an independent ground truth for large-scale CNV patterns. Across all samples and metrics, SCOIGET consistently outperformed four existing methods (InferCNV, CopyVAE, CopyKAT, SCEVAN).
In one representative sample (HT260C1), SCOIGET achieved a Mean Squared Error (MSE) of 1.076 and cosine similarity of 0.903, compared to InferCNV's MSE of 1.639 and cosine similarity of 0.853. Across all tested samples, SCOIGET reduced error metrics by 30-80% relative to the best competing method while achieving superior similarity to the WES ground truth.
For clustering quality -- the ability to identify distinct tumor clones based on CNV profiles -- SCOIGET-derived features produced Silhouette Scores ranging from 0.35 to 0.46 across samples. In contrast, InferCNV-based clustering yielded scores from -0.03 to 0.10, near or below zero, indicating poor cluster separation. CopyVAE showed highly variable performance. These results demonstrate that SCOIGET produces not just more accurate CNV estimates, but ones that translate into more meaningful biological groupings.
Validation on simulated data with known ground truth confirmed spot-level accuracy: the model achieved a cosine similarity of 0.632 and MSE of 0.032, and correctly recovered the spatial architecture of three predefined clonal domains (normal, gain-heavy, and loss-heavy regions) from noisy synthetic data.
The researchers applied SCOIGET to four colorectal cancer tissue sections from the same patient, covering both precancerous (G1 stage) and cancerous (G2 stage) tissue. The model identified spatially distinct clonal domains corresponding to different stages of tumor development, with high-CNV-score regions localized centrally in cancerous areas.
Four tumor subclones were identified: two (cloneA, cloneB) exclusive to the cancerous stage, one (cloneC) spanning both precancerous and cancerous tissue, and one (cloneD) marking a transitional boundary. This pattern suggests that cloneC represents an evolutionary intermediate -- a genetic lineage present before malignant transformation that persists into cancer. These inferences were validated by pathologist annotations of hematoxylin-eosin stained tissue sections.
Survival analysis of genes specifically upregulated in cloneB revealed prognostic significance. TUBA1C (encoding a tubulin essential for cell division) and H2AFZ (a histone variant involved in gene regulation) showed significantly different survival curves in colon cancer patients from TCGA data, suggesting these cloneB-specific molecular features have real clinical relevance.
To test generalizability, SCOIGET was applied to four spatial transcriptomics samples from a single prostate cancer patient. The model successfully inferred spatial CNV features and identified distinct tissue domains corresponding to Gleason Grades (GG1, GG2, GG4), border regions, and benign tissue -- aligning with the established histological classification system for prostate cancer aggressiveness.
The CNV-based spatial clustering revealed finer substructure than traditional Gleason grading alone, identifying molecular distinctions between tumor grades. Chromosome-level CNV heatmaps showed unique and shared alterations across the three tumor grades, providing a genomic basis for the morphological differences that pathologists use to grade prostate cancer.
This demonstration in prostate cancer, combined with the colorectal cancer results, establishes SCOIGET's versatility across cancer types and spatial transcriptomics platforms (both standard Visium and the higher-resolution Visium HD). The ability to automatically map Gleason-grade spatial heterogeneity using genomic features rather than subjective morphological assessment opens new avenues for objective prostate cancer characterization.
SCOIGET represents a meaningful advance in computational oncology by being the first framework to combine graph neural networks with HMM-based CNV inference using spatial transcriptomics data. Its graph-based design allows spatial neighborhood information to propagate across tissue sections, capturing local heterogeneity that spot-level methods miss.
The practical value lies in SCOIGET's three integrated modules: inferring CNV features, identifying spatial tumor clones, and reconstructing tumor evolutionary trajectories. Together, these outputs provide a comprehensive spatial map of tumor heterogeneity that can guide decisions about which tumor regions are most aggressive, which have already evolved drug resistance, and which might harbor the key evolutionary transitions driving metastasis.
With a computation time of approximately 35 minutes per sample on modern hardware -- compared to 2-8 hours for competing methods -- SCOIGET is practical for research use today and positions itself for eventual integration into clinical tumor profiling workflows. The authors envision this technology supporting personalized cancer treatment strategies by identifying patient-specific evolutionary vulnerabilities that could be targeted therapeutically.