🧬 biology

What is shared across 33 cancer types is the tissue, not the tumour: a non-malignant axis of the microenvironment

Analysis of over 10,000 TCGA tumor samples reveals that reproducible causal network structures across 33 cancer types are driven by shared non-malignant tissue microenvironment components rather than tumor-specific drivers, indicating that pan-cancer causal inference from bulk data primarily reflects sample composition rather than intrinsic tumor regulation.

Original authors: Shuaidong Gao

Published 2026-11-03

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Shuaidong Gao

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Cancer is not a single disease but a vast collection of hundreds of different conditions, each arising from a specific type of tissue in the body. While the cells that form a tumor are the most obvious part of the problem, a tumor does not exist in a vacuum. It is surrounded by a complex neighborhood of healthy cells, including immune defenders, structural support cells, and blood vessels. Scientists have long tried to map the internal wiring of these tumors, looking for the specific instructions that tell cancer cells how to grow and spread. The hope has been that by studying many different types of cancer at once, researchers could find a universal set of rules—a common regulatory program—that drives all of them. If such a program existed, it would offer a powerful target for new treatments that could work across the board, rather than requiring a unique solution for every single type of cancer.

A new study challenges this long-held hope by looking at the data in a way that had not been done before. The researchers analyzed gene activity from more than 10,000 tumor samples representing 33 different types of cancer. They used a method designed to find cause-and-effect relationships between genes, asking which genes turn on or off other genes. However, they discovered a fundamental flaw in how these comparisons were usually made. Because different cancers naturally use different sets of genes to function, comparing them directly is like trying to compare two maps of different cities by looking at the same coordinate on the page; the location might be the same, but one map shows a park while the other shows a highway. To fix this, the researchers forced all 33 cancer types to be analyzed using the exact same list of 100 genes. Only when they aligned the data this way could they see what was truly shared across the different diseases.

What they found was surprising. Instead of revealing a common set of instructions that drive the cancer cells themselves, the shared structure pointed entirely to the healthy tissue surrounding the tumor. The most consistent pattern they found was a group of three genes that act as signals to recruit immune cells, specifically a type of white blood cell that fights infection. These genes appeared together in nearly every cancer type, but they were not part of the cancer's internal machinery. Instead, they were markers of the tumor's environment, indicating how much immune activity and structural tissue was present in the sample. The study showed that the genes most commonly used to build these cancer maps are the ones that vary the most in their activity. These highly variable genes happen to be the ones that describe the mix of healthy cells in the sample, not the genes that drive the cancer. The actual genes known to cause cancer, the "drivers," were so stable in their activity that they did not even appear in the top 100 genes selected for the analysis.

The researchers also uncovered a practical rule about how much data is needed to trust these maps. They found that when a study has too few samples, the computer program used to find the connections starts inventing patterns out of random noise. By shuffling the data to see what happened when there was no real biology left, they identified a clear threshold: a study needs at least 172 samples to reliably find real connections when looking at 100 genes. Below this number, the results are likely just statistical accidents. This finding suggests that many previous studies with smaller sample sizes may have been reporting false connections. Furthermore, the study confirmed that the few connections that did appear consistently across all 33 cancers were not about the cancer cells talking to each other, but about the healthy tissue responding to the presence of the tumor.

The implications of this work are significant for how scientists interpret cancer data. It suggests that when researchers see the same gene connections appearing in many different cancers, they should not assume they have found a universal cancer mechanism. Instead, they are likely seeing a reflection of the common healthy tissue that surrounds every tumor. The study does not say that cancer cells lack their own unique wiring, but rather that the standard tools used to find that wiring are currently tuned to hear the noise of the surrounding environment instead. To find the true drivers of cancer, scientists will need to look at a much wider range of genes and ensure they have enough samples to separate real signals from random chance. Until then, the most reproducible feature shared across all cancers is not the tumor itself, but the tissue it lives in.

Technical Summary: "What is shared across 33 cancer types is the tissue, not the tumour: a non-malignant axis of the microenvironment"

Problem Statement
Pan-cancer causal studies routinely assume that regulatory edges inferred in one tumor type transfer to others, interpreting shared edges as evidence of a common regulatory program. However, a fundamental methodological obstacle prevents rigorous testing of this assumption across The Cancer Genome Atlas (TCGA) cohorts: standard analyses select tissue-specific gene panels (e.g., the top 100 most variable genes per cohort). Because these panels differ across cancer types, the adjacency matrices index different gene pairs at the same positions, rendering element-wise comparison undefined. Existing approaches either analyze cohorts in isolation (precluding comparison), pool samples (assuming homogeneity that contradicts biological reality), or restrict comparisons to shared regulators without examining the edges themselves. This paper addresses the question: if a common gene panel is held fixed across 33 TCGA cohorts, what cross-cohort causal structure is recoverable, and does it describe tumor-cell regulation or the composition of the samples?

Methodology
The study employs a two-stage multi-batch causal decomposition framework using the NOTEARS algorithm (Zheng et al., 2018) for continuous optimization of Directed Acyclic Graphs (DAGs).

Data and Alignment: The author utilized HiSeqV2 RSEM-normalized expression data from 33 TCGA cancer types ( $N=10,459$ samples). To enable comparison, they constructed a fixed "aligned panel" of 100 genes selected by median absolute deviation (MAD) across all cohorts, rather than using the standard per-cohort top-100 lists.
Causal Estimation: A NOTEARS graph was fitted for each of the 33 cohorts against this common panel. The model minimizes the reconstruction error with an $L_1$ penalty and a differentiable acyclicity constraint.
Decomposition: The author modeled the causal adjacency matrix for each cancer $k$ as $W_k = W_0 + \Delta_k$ , where $W_0$ is a shared causal backbone and $\Delta_k$ represents cancer-specific deviations. $W_0$ was estimated via the element-wise median of the per-cohort adjacency matrices, projected onto the DAG manifold.
Diagnostics and Validation:
- Null Controls: Sample labels were shuffled within cohorts to test if recurrence was an artifact of sample size or estimator properties.
- Composition Control: A composite score of non-malignant markers (stromal, immune, endothelial) was regressed out of the panel genes to test if the recurring structure was mediated by bulk tissue composition.
- External Validation: Recurring edges were compared against STRING (physical interactions) and TRRUST (transcription factor-target) databases.
- Driver Analysis: The dispersion percentile of canonical driver genes was compared against the selection threshold of the 100-gene panel.

Key Results

Alignment is Prerequisite for Comparison: In the original misaligned design (per-cohort top-100 genes), the element-wise median adjacency matrix was identically zero (largest median coefficient 0.048). When aligned to a common panel, a non-zero structure emerged (largest median coefficient 0.731), demonstrating that cross-cohort comparison is only defined after alignment.
Recurring Structure is Microenvironmental, Not Malignant: Only 14 directed gene pairs recurred in $\ge10$ of the 33 cohorts. These pairs did not involve canonical driver genes. Instead, they formed a coherent module centered on a co-induced, interferon- $\gamma$ -inducible CXCR3-ligand triplet (CXCL9, CXCL10, CXCL11). The remaining recurring edges mapped to six non-malignant axes: neutrophil/CXCR2 chemokines, blood content (hemoglobin), plasma-cell/immunoglobulin, stroma/complement, and myeloid/cytokine signaling.
Dominant Axis Reflects Tissue Composition: The first principal component (PC1) of the pooled aligned matrix correlated strongly ( $r=0.803$ ) with external stromal and immune markers across all 33 cohorts. Canonical driver genes, which sit at the 50.8th percentile of transcriptome dispersion, were structurally excluded from the top-100 panel (cut at the 0.487th percentile). Drivers only entered the panel systematically when the panel width expanded to $k \gtrsim 1,000$ .
Sample-Size Dependence and Identifiability Floor: The number of inferred edges per cohort was strongly negatively correlated with sample size ( $r = -0.652$ ) across the full set, but this correlation vanished for cohorts with $n \ge 200$ . Shuffling sample labels abolished the recurring set (14 $\to$ 0 edges) but preserved the sample-size dependence. This identified an empirical "identifiability floor" at $n^* \approx 1.7d$ (where $d=100$ ); cohorts below this threshold return edges driven by noise rather than biology.
Orientation and Recurrence: While the recurring edges showed statistical significance against gene-wise permutation nulls, their orientations were not identified by the Gaussian observational data (e.g., the CXCL9-CXCL10-CXCL11 module is a co-induced set, not a sequential cascade). Three of the 14 pairs were recovered in both orientations.

Significance and Claims
The paper claims three primary contributions that redefine the interpretation of pan-cancer causal networks:

The Alignment Requirement: Cross-cohort causal comparison is mathematically undefined without a fixed feature panel. Without alignment, the "shared structure" is zero by construction.
Composition over Regulation: In standard variance-selected panels (top 100 genes), the reproducible causal structure across cancer types describes the composition of the tissue (immune infiltration, stromal content, blood content) rather than the regulatory architecture of the tumor cells. The absence of driver genes is a structural consequence of the gene-selection criterion, not a biological coincidence.
Operational Sample-Size Floor: The number of edges inferred by sparse causal estimators is a finite-sample property of the estimator, not a biological signal. The study establishes an empirical rule ( $n \gtrsim 1.7d$ ) for determining when a cohort is large enough to distinguish signal from noise in structure learning.

The author concludes that pan-cancer studies reporting conserved edges should interpret them as microenvironment-conditional dependencies. To recover driver-level causal structures, studies must either widen the gene panel significantly (accounting for sample size constraints) or explicitly condition on tumor purity. The recurring structure is the tissue, not the tumor.

More like this