Tumor-Budding Signature in Colorectal Cancer Stands Out Only When the Cells Are Separated
A colorectal cancer preprint finds that a tumor-budding-associated gene signature behaves very differently depending on transcriptomic resolution. In bulk RNA-seq from TCGA-COADREAD, the co-occurrence signal across the TB-linked genes was weak, with fold-enrichment ranging from 1.16 to 2.72. The weakest pair, AIF1 and LIF, showed no significant association at all, with a fold-enrichment of 1.16 and FDR of 0.31. In single-cell RNA-seq, the same gene pairs rose sharply to 31- to 184.5-fold enrichment, and every pair passed a stringent false discovery threshold with FDR below 8 × 10⁻³⁸.
The contrast is not subtle. It points to a signature that is biologically real but diluted when stromal and epithelial compartments are averaged into one bulk measurement.
Tumor budding has long been linked to aggressive behavior in colorectal cancer, including invasion, metastatic spread, and epithelial plasticity. The four genes centered in this analysis — AIF1, LIF, MRC1, and TNC — were previously enriched in tumor-budding regions by spatial transcriptomics, and this study asked a practical question: can that same module be recovered in routine bulk data, or does the signal vanish unless the tumor microenvironment is parsed at single-cell scale? The answer here favors the latter. When transcripts from many cell types are blended together, rare but coordinated cellular neighborhoods are easy to miss.
Bulk Averaging Blunts a Microenvironmental Program
Bulk RNA-seq remains indispensable for large cohorts, but it imposes a biological compromise. A signature that is driven by a localized population of budding cells, macrophage-like cells, or surrounding stromal elements may be real in tissue and still appear faint in the aggregate. In this dataset, pairwise enrichment in bulk stayed modest, and the absence of a significant AIF1-to-LIF association suggests that the co-detected pattern may depend on cellular adjacency rather than uniform expression across the tumor mass.
By contrast, single-cell data from GSE178341 captured the same pairs with striking frequency. That pattern is exactly what one would expect if the TB program is concentrated within specific cells or within a tightly organized microenvironmental niche. Co-expression at this scale can reflect shared cell states, local signaling loops, or transcriptional programs induced by the same invasive milieu. It can also reveal how much bulk data compresses biology into a single average that hides compartment-specific behavior.
Doublets Did Not Explain the Signal
A useful concern in single-cell work is whether apparent co-expression is merely an artifact of doublets, where two cells are captured together and falsely appear to express the same genes. Total UMI counts provided no evidence that doublet contamination was driving these co-detection patterns. That leaves the high enrichment signal more plausibly tied to real cellular architecture rather than technical noise.
The implication is straightforward. If tumor budding is a localized biological state, then its molecular readout will often be strongest where the cells are least mixed. Single-cell sequencing does not just add resolution; in this context, it restores the architecture of the signal itself.
Different Regulation May Be Operating Across the Gene Module
Exploratory enhancer-promoter interaction analysis in HCT116 cells also suggested that these genes are controlled differently. While the preprint does not overstate that finding, it adds another layer to the story: genes enriched in the same pathologic niche may not be wired identically at the regulatory level. Distinct enhancer-promoter relationships can produce coordinated phenotypes through different upstream controls, especially in tumors where signaling, inflammation, and extracellular matrix remodeling overlap.
That matters for interpretation. A shared spatial signature does not require a shared regulatory mechanism, and a shared phenotype does not guarantee the same transcriptional switch. In a tumor-budding context, one gene may reflect immune-macrophage activity, another stromal remodeling, and another invasive epithelial behavior, even while all three cluster in the same microscopic region.
Late-Stage Disease and Metabolic Gene Suppression
The analysis also extended beyond the budding module itself. GOT1 and MDH1, two genes involved in the malate-aspartate shuttle, showed lower expression in late-stage colorectal cancer, both with FDR less than or equal to 0.010. This is biologically plausible in a malignancy that reshapes redox balance and amino-acid handling during progression. The malate-aspartate shuttle helps coordinate mitochondrial and cytosolic metabolism, so reduced expression may reflect broader metabolic reprogramming as tumors advance.
This part of the work sits adjacent to the budding analysis but still fits the same disease logic. Invasive colorectal cancer does not alter one pathway at a time; it shifts cell state, metabolism, and intercellular signaling together. Gene modules tied to invasion and metabolic adaptation often move in parallel, even if the mechanistic links are not identical.
Modest Prognostic Signal, Limited by the Bulk Framework
A bulk-derived risk score showed a modest association with survival, with Cox HR=1.21 and p=0.021. That finding is directionally consistent with a clinically relevant signature, but it remains restrained in magnitude. A hazard ratio in this range suggests signal, not a stand-alone prognostic tool. It may be useful as part of a larger molecular model, though this preprint does not establish clinical readiness or calibration against established colorectal cancer markers.
The key limitation is built into the method itself. If the biology is strongly cell-state dependent, then a score derived from bulk data will inevitably compress the very architecture that makes the signature informative. A modest hazard ratio in the bulk setting may therefore underestimate the true relevance of the underlying tumor-budding program.
Resolution, Not Just Discovery, Drives Interpretation
This preprint makes a clear technical and biological point: transcriptomic resolution can determine whether a tumor-microenvironment signature is visible at all. The TB-linked genes were detectable in both bulk and single-cell datasets, but only single-cell analysis revealed the strength of their co-expression. For colorectal cancer research, that is a reminder that some aggressive programs are not absent in bulk tissue; they are simply blurred by it.
The work also stops short of proving clinical utility. It does not validate the signature prospectively, compare it against current risk stratification tools, or establish whether it can guide treatment decisions. What it does show is narrower and more defensible: in colorectal cancer, a tumor-budding gene module looks far more coherent when the tissue is dissected cell by cell than when it is averaged into one bulk transcriptomic profile.