A result can be reproducible without being a clinical test

When a cancer study identifies a metabolic gene-expression signature, it is tempting to tell a simple story: a pathway is switched on or off, a vulnerability has been found, and a treatment may follow. Each step, however, requires different evidence. A transcriptional pattern can be reproducible and biologically informative without measuring a metabolic process or predicting a patient’s outcome.

Our recent Frontiers in Oncology paper, “A predefined sulphur-handling transcriptional programme is associated with colorectal cancer molecular subtypes,” provides a concrete example. We examined a predefined sulphur-handling gene panel and found that its combined expression score tracked molecular subtype across independent cohorts. The result supports a question about tumour biology, not a shortcut to a clinical assay. The distinction matters both when designing follow-up experiments and when explaining computational research beyond its immediate specialist audience [1] .

What the score actually measures

Sulphur handling spans several processes, including cysteine handling, transsulfuration and mitochondrial sulphide oxidation. Genes assigned to these processes do not perform interchangeable tasks. Our panel contained 14 genes selected before the expression analysis; its score was an unweighted average of gene-wise, standardised expression values within each cohort.

That design gives researchers a specified transcriptional feature to examine. It does not turn the score into an assay of hydrogen sulphide production, enzyme activity or sulphur flux. Nor does the predefined panel capture every component of sulphur metabolism. A lower score therefore needs an explicitly transcriptional interpretation, rather than the broader claim that a tumour has less sulphur-handling activity[1] .

Replication, with the denominators kept visible

Guinney and colleagues’ Nature Medicine paper, “The consensus molecular subtypes of colorectal cancer,” established a framework that groups these cancers by distinguishing biological features. CMS4 is associated with mesenchymal and stromal biology, whereas CMS3 has prominent metabolic features. This framework offers context for interpreting expression programmes, but an association with subtype does not establish what caused the association [2].

We assigned CMS classifications using CMScaller [3]. CMS4 tumours had lower programme scores in TCGA-COAD and in each of six independent validation cohorts. Those validation datasets contained 1,337 input tumours. Of these, 1,187 received qualifying CMS assignments and entered the subtype meta-analysis; 150 were unclassified. The pooled CMS4-minus-non-CMS4 difference was −0.302 programme-score units. Effect sizes varied across cohorts, with heterogeneity of I²=63%. Consistent direction should not be confused with identical effects, and these score units are not absolute expression or metabolic-rate units [1] .

Why tissue composition changes the question

A bulk tumour-expression profile combines signals from epithelial, stromal and immune cells. A difference between tumour groups might therefore reflect altered expression within cells, a changed mixture of cells, or both. Aran, Sirota and Butte’s Nature Communications study, “Systematic pan-cancer analysis of tumour purity,” demonstrated why purity can confound transcriptomic comparisons [4].

This issue is especially relevant to a stromal-rich subtype such as CMS4. In our TCGA-COAD sensitivity analyses, the association remained statistically detectable after adjustment using ESTIMATE [5], consensus purity estimates [4] and ABSOLUTE [6]. That strengthens the observation but does not establish complete independence from tissue composition. Purity estimators remain estimates, and the models used different complete-case samples.

Single-cell-informed analyses provided supporting compartmental context, not direct validation of biochemical function in the same tumours. Cross-resource agreement in aggregated expression patterns also does not prove that inferred cell fractions are accurate in individual samples. The defensible interpretation is that measured global purity differences do not fully explain the association not that cellular composition has been ruled out [1].

One composite, several expression directions

The gene-level results make a simple pathway-on or pathway-off narrative particularly unsuitable. Several oxidation and clearance transcripts, including SQOR, ETHE1 and TST, were reduced in tumour-versus-adjacent-normal comparisons. SLC7A11 and AHCY increased, while CBS was not significantly altered. These tumour–normal findings should remain distinct from the composite subtype comparison.

The programme therefore describes a coordinated but nonuniform transcriptional architecture. Averaging genes into a score can reveal an association while concealing important biological differences among its components. For follow-up research, both the composite and the individual branches deserve attention [1].

What the findings do not establish

The score showed no statistically detectable independent survival association. It should therefore be interpreted as subtype-associated biology, not a validated prognostic classifier. Likewise, SLC7A11/xCT a transporter at the centre of ferroptosis biology, a dependency landscape we have previously reviewed in another cancer type remains a preclinical hypothesis, not evidence of CMS4-specific drug sensitivity or treatment benefit [1,7] .

The next experiments should bridge the actual gaps

A useful next stage would combine compartment-resolved expression analysis with direct measurements of enzyme activity, metabolites or flux in appropriately characterised colorectal cancer models. Related glioblastoma work that I coauthored integrated bulk, spatial and single-cell transcriptomics to identify niche-specific expression programmes. That provides methodological context, not validation of the CRC findings. Genetic perturbation and orthogonal experimental approaches would be needed to examine proposed dependencies. These are proposed investigations, not results of the present study.

The broader lesson is not that computational signatures lack value. Their value becomes clearer when each claim matches its evidence. Reproducible transcriptional patterns can map tumour diversity and focus experimental questions. Preserving the distance between association, mechanism and clinical utility makes that contribution more useful not less compelling.

Author perspective

We fixed the gene panel before examining expression results. A central interpretive challenge was separating the subtype association from what it might mean biologically. Checking tumour purity helped address one alternative explanation but could not turn a transcriptional pattern into a direct measurement of metabolism. That distinction shapes the experiments needed next.

 References:

1. Shafi, S., Hamid, A., Yang, J., Cheng, X. & Sun, M. A predefined sulphur-handling transcriptional programme is associated with colorectal cancer molecular subtypes. Front. Oncol. 16, (2026).
2. Guinney, J. et al. The consensus molecular subtypes of colorectal cancer. Nat. Med. 21, 1350–1356 (2015).
3. Eide, P. W., Bruun, J., Lothe, R. A. & Sveen, A. CMScaller: an R package for consensus molecular subtyping of colorectal cancer pre-clinical models. Sci. Rep. 7, 16618 (2017).
4. Aran, D., Sirota, M. & Butte, A. J. Systematic pan-cancer analysis of tumour purity. Nat. Commun. 6, 8971 (2015).
5. Yoshihara, K. et al. Inferring tumour purity and stromal and immune cell admixture from expression data. Nat. Commun. 4, 2612 (2013).
6. Carter, S. L. et al. Absolute quantification of somatic DNA alterations in human cancer. Nat. Biotechnol. 30, 413–421 (2012).
7. Lu, M. et al. The molecular mechanisms of ferroptosis and its role in glioma progression and treatment. Front. Oncol. 12, (2022).