AI-Discovered Prognosis from the Glass Slide: The DoMore Deep-Learning Biomarker in Colorectal Cancer

The clinical question

For patients with colorectal cancer—most often adenocarcinoma of the colon—one of the central decisions after surgery is how aggressively to pursue adjuvant therapy. That decision hinges on prognosis: how likely is the cancer to recur or to cause death? Traditional prognostication relies on tumour stage, nodal status, lymphovascular invasion, and a handful of molecular markers. Yet patients within the same stage category can have strikingly different outcomes, and clinicians have long recognised that staging alone leaves considerable uncertainty, particularly in stage II and III disease where the benefit of chemotherapy is debated.

This residual uncertainty raises a practical question: does the ordinary haematoxylin-and-eosin (H&E) stained slide—the same slide a pathologist examines every day—contain prognostic information that human observers systematically underuse? The DoMore project set out to test whether a computer could learn to read that information directly from tumour morphology [1].

What the method is and how it works

The DoMore biomarker is a tissue-based prognostic classifier built with deep learning, a form of artificial intelligence in which multilayer neural networks learn predictive patterns directly from image data rather than from features a human specifies in advance [1]. Crucially, the input is nothing exotic: it is the conventional H&E-stained section, digitised as a whole-slide image. No special stain, no immunohistochemistry, no sequencing—just the morphology already captured in routine practice [1].

The conceptual novelty is worth emphasising for a mixed audience. In classical pathology, a human expert defines what to measure (gland formation, mitotic count, tumour budding) and then grades it. In the DoMore approach, the network is instead trained against a hard clinical endpoint—patient outcome—and left to discover, on its own, which microscopic patterns correlate with that endpoint [1]. The result is an AI-discovered morphologic marker: a prognostic signal that emerges from the pixels of the slide rather than from a pre-specified histologic rule. The output is a simple categorical classification of a patient's tumour as having a good or poor prognosis (with an intermediate/uncertain category), which can then be tested as an independent predictor of survival [1].

Several design choices in the study deserve attention because they bear on how much we should trust the result. The classifiers were developed on large training material and then locked before being applied to separate patient cohorts, so that the performance figures reflect genuine validation on previously unseen cases rather than performance on the data used to build the model [1]. This train-then-validate discipline is the single most important safeguard against the over-optimism that plagues machine-learning studies, and it is what distinguishes a credible prognostic claim from a preliminary one.

The evidence

The pivotal evidence comes from the 2020 report in The Lancet by Skrede and colleagues [1]. After developing the deep-learning classifier, the investigators evaluated it in a validation cohort of 1,122 patients with colorectal cancer—a substantial size for a biomarker study and one that lends statistical weight to the findings [1].

The headline result is the association between the AI classification and patient survival. Patients classified by the model as poor-prognosis had a markedly higher risk of a cancer-specific death event than those classified as good-prognosis, with a hazard ratio of 3.84 in the univariable comparison [1]. A hazard ratio of this magnitude means the instantaneous risk of the adverse outcome was nearly fourfold higher in the poor-prognosis group across follow-up—a large effect for a single marker.

The more demanding test, however, is whether the marker adds information beyond what pathologists and clinicians already have. To address this, the investigators adjusted for established prognostic variables. After adjustment, the hazard ratio was 3.04—attenuated, as expected, but still substantial and statistically meaningful [1]. The persistence of a strong, independent signal after accounting for conventional factors is the crux of the paper: it indicates the network is capturing prognostic morphology that is not simply a restatement of stage or standard histopathology [1]. The evidence suggests, in other words, that the ordinary H&E slide carries exploitable prognostic information that routine reporting does not fully extract.

For readers parsing the numbers: the univariable HR tells you how well the marker separates outcomes on its own; the multivariable (adjusted) HR tells you how much of that separation is genuinely new. That the adjusted figure remained high—3.04—is what makes the result interesting rather than merely confirmatory [1].

Where it stands

It is important to be precise and sober about status. The DoMore biomarker as reported is a research-validated prognostic marker, demonstrated in a large external validation cohort, but it is not something a patient or clinician should treat as an established, deployable test in the way one treats, say, nodal status [1]. Within the scope of the provided evidence, this is a prognostic tool—it stratifies risk—and it should not be read as directing any specific therapy for any individual. Any downstream implication for treatment intensity belongs to the clinical team working within the usual decision-making frameworks, not to the marker itself.

Several limitations and open questions frame how this work should be interpreted.

First, prognostic is not predictive. Demonstrating that the model identifies higher- and lower-risk groups is distinct from demonstrating that those groups benefit differently from a particular class of therapy. The cited study establishes association with outcome; it does not, in itself, establish that acting on the classification changes outcomes [1]. That distinction matters enormously for clinical use and would require prospective, interventional evaluation.

Second, deep-learning models are sensitive to the characteristics of the slides they were built on—staining protocols, scanners, and tissue-handling practices can all vary between laboratories. Robust generalisability across diverse real-world settings is a general concern for this class of tool and warrants continued, independent validation before broad adoption.

Third, interpretability remains an active area of discussion. Because the network discovers its own features, the precise morphologic substrate of the prognostic signal is not transparent in the way a mitotic count is. This "black box" quality is part of what makes AI-discovered markers both powerful and, for some clinicians, uncomfortable.

Finally, this is a fast-moving field, and the appropriate posture is epistemic humility. The DoMore result is a strong, well-validated proof of concept that H&E morphology, read by deep learning, yields independent prognostic value in colorectal cancer [1]. Translating that into a regulated, reproducible, clinically actionable assay—and demonstrating that it improves patient outcomes rather than merely predicting them—is the work that lies ahead.

The enduring lesson for trainees and clinicians is conceptual as much as technical: the humble H&E slide, scanned and interrogated by modern computation, may hold more actionable information than a century of human pattern-recognition has managed to codify. DoMore is one of the clearest demonstrations to date that this information exists and can be extracted [1].

---

References

1. Skrede OJ, et al. Deep learning for prediction of colorectal cancer outcome: a discovery and validation study. Lancet. 2020. PMID 32007170. (Deep-learning H&E-based prognostic biomarker in colorectal cancer; validation cohort n = 1,122; poor- vs good-prognosis hazard ratio 3.84, adjusted 3.04.)