Microbiome signals in rectal cancer iTNT need a stricter clinical frame
A new commentary in rectal cancer places a sharp methodological lens on longitudinal stool profiling from the TORCH trial, where immunotherapy-based total neoadjuvant therapy was paired with serial microbiome and metabolome sampling. The underlying study followed 102 stool samples across three treatment timepoints in patients with proficient mismatch repair or microsatellite stable locally advanced disease, then linked microbial and metabolic patterns to treatment response and toxicity. The commentary does not dispute the biological signal. It asks a harder question: which parts of that signal are ready for bedside use, and which remain tied to a small, incomplete, and internally fitted cohort.
One of the most striking observations is the claim that short-course radiotherapy-based iTNT reorganizes the gut ecosystem rather than simply erasing it. In cancer immunotherapy, that distinction matters because broad depletion and selective reshaping can point to very different mechanisms. Depletion would imply a loss of resilience and microbial diversity; reorganization suggests a selective pressure that preserves some communities while favoring others with potential relevance to immune tone, metabolite output, and treatment tolerance. The commentary treats that framework as promising, but insists that the clinical meaning depends on how response is defined and whether the microbiome signature predicts durable organ preservation rather than only an early restaging snapshot.
Response definitions may shift the biomarker list
The most immediate statistical issue sits in the definition of “responder.” In the Methods, near-complete response is counted alongside pathological complete response and clinical complete response, yet the Results define responders only as those with pathological or clinical complete response. Near-complete response is the least stable category in rectal cancer response assessment, and folding it into one analysis while excluding it from another can alter the apparent 29 versus 15 split between groups. If the response label changes, the taxa and metabolites that appear to discriminate responders from nonresponders may change as well.
A one-line biomarker definition can redraw the entire analysis set.
That concern is not academic. Restaging response is an imperfect proxy for organ preservation. Clinical complete response detects ypT0 with limited sensitivity, and in suspected regrowth, endoscopy outperforms magnetic resonance imaging for predicting complete pathologic response. Even when a patient reaches clinical complete response, later regrowth is common enough to weaken confidence in any marker that only tracks the initial restaging event. A microbiome feature that separates early responders from nonresponders may still fail to predict whether a patient truly avoids total mesorectal excision.
Sampling depth and attrition shape what the cohort can show
The commentary also focuses on who was actually sampled. Baseline stool collection captured 44 participants, roughly one third of the 121 evaluable patients in TORCH. Responders made up 29 of those 44 samples, a proportion that is numerically higher than the response rate reported in the parent trial arms, but the sample is too small to establish genuine enrichment. In other words, the sampled subgroup may be suggestive, yet it is not a clean miniature of the whole trial population.
Attrition complicates interpretation further. The sample count fell from 44 to 37 to 21 across the longitudinal series, and recent antibiotic exposure prevented sampling. That creates a realistic risk that patients with infection or febrile neutropenia were missing more often at later timepoints, exactly the kind of patients most relevant to toxicity analyses. Selective missingness would tend to weaken toxicity associations rather than inflate them, but it still distorts inference. For a microbiome study trying to predict adverse events, missing the sickest participants can quietly flatten the most important signal.
The commentary also raises confounding from diet and concomitant medication. Proton-pump inhibitors and metformin are among the strongest non-antibiotic modifiers of gut composition, and neither appears to have been modeled. Age, supportive care, and treatment context can all alter microbial composition and metabolite production. Without prospectively captured covariates, a tumor-associated signal can look more precise than it really is.
In-sample discrimination is not the same as prediction
Several of the strongest statistical concerns center on model development. Random-forest feature selection and receiver operating characteristic analysis were performed inside the same 44-patient dataset, which creates the usual risk of optimism. The combined Klebsiella pneumoniae and GABA model reached an area under the curve of 72.6%, but K. pneumoniae alone performed at 73.8%. That pattern points less to a biologically additive multivariable model than to instability in a small training set.
A model that fits the same patients used to build it can look stronger than it is.
The same problem appears in the toxicity score. Metabolite-derived toxicity indices were built from coefficients fitted to hematologic measures, then correlated back to those same measures. That establishes internal fit, not out-of-sample predictive validity. For a clinically useful test, the next step has to be nested cross-validation with feature selection confined to inner folds, followed by independent testing in another short-course iTNT cohort. Decision curve analysis would also be useful because a predictor only matters if it changes management in a way that improves net benefit, not merely if it separates curves on paper.
GABA biology is provocative, but the causal chain is not closed
The mechanistic arm of the study is the most distinctive part of the work. Oral gamma-aminobutyric acid, or GABA, blunted the antitumor effect of radiotherapy plus PD-1 blockade in vivo. That finding makes the GABA axis biologically plausible as more than a bystander, yet the experimental design does not prove that microbial GABA reaches the same concentrations or compartments in patients. The mouse model used systemic GABA at 6 mg/mL and a subcutaneous MC38 tumor, which is not the same biological setting as pelvic irradiation in rectal cancer.
Context matters because GABA signaling is not uniformly suppressive across all systems. GABA receptor activity can behave differently depending on receptor subtype, tumor type, and whether the signal is local or systemic. A microbial metabolite that appears inhibitory in one setting could have a different effect in another, especially when immune activation, tumor biology, and the irradiated tissue compartment all shift at once. The commentary therefore calls for gnotobiotic colonization experiments comparing wild-type and GABA-pathway-deficient K. pneumoniae, stable-isotope tracing, receptor antagonism, and targeted quantification of GABA in serum and tumor tissue.
What should come next in rectal cancer biomarker work
The broader clinical message is not that microbiome research in iTNT is overreaching. It is that the field is now far enough along to demand validation with stricter endpoints and cleaner design. Response features should be re-tested against regrowth-free survival and total mesorectal excision–free survival, not only against the first restaging event. Sampling should be harmonized across future cohorts, with explicit capture of antibiotic exposure, diet, and medication use. Feature selection should move into an independent validation structure rather than remaining inside a single exploratory set.
This commentary also draws a boundary around disease subtype. The discussion applies to pMMR and MSS rectal cancer, not to dMMR or MSI-H tumors, where PD-1 blockade has a different therapeutic role. It also notes that taxa linked to favorable response may overlap with organisms associated with immune-related adverse events, which makes any intervention on the microbiome a clinical question that deserves randomized testing before it reaches routine practice.
The core achievement of the TORCH microbiome analysis remains intact: it provides one of the more ambitious longitudinal fecal multi-omics datasets in rectal cancer immunotherapy. The commentary’s message is narrower and more demanding. Before the field treats taxa such as K. pneumoniae or metabolites such as GABA as actionable biomarkers, it needs proof that the signal survives clearer response definitions, more complete sampling, and external validation in a separate trial population.