Articol susținut de S.N.O.M.R.

Marius Florentin Popa¹, Paul Siancu^2,*, Călin-Ilie Mohor^2,*, Lilioara-Alexandra Oprinca-Muja², George-Călin Oprinca², Cosmin-Ioan Mohor², Ciprian Tănăsescu², Denisa Tănăsescu², Alina Cristian², Vicentiu-Vasile Veres² , Alina Bereanu²

1. Faculty of Medicine, Ovidius University of Constant,a, 900527 Constant,a, Romania; marius_popa2005@yahoo.com
2. Faculty of Medicine, Lucian Blaga University of Sibiu, 550169 Sibiu, Romania; lilioaraalexandra.muja@ulbsibiu.ro (L.-A.O.-M.); georgecalin.oprinca@ulbsibiu.ro (G.-C.O.); cosmin.mohor@ulbsibiu.ro (C.-I.M.); tanasescuciprian@yahoo.fr (C.T.); denisa.tanasescu@ulbsibiu.ro (D.T.); alina.cristian@ulbsibiu.ro (A.C.); vicentiu.veres@ulbsibiu.ro (V.-V.V.); alina.bereanu@ulbsibiu.ro (A.B.) * Correspondence: pauls.siancu@ulbsibiu.ro (P.S, .); calin.mohor@ulbsibiu.ro (C.-I.M.)

Int. J. Mol. Sci. 2026, 27, 8025

DOI: https://doi.org/10.3390/ijms27188025

Abstract:
From a molecular perspective, colorectal cancer is a heterogeneous disease in which microsatellite instability (MSI) phenotypes, mismatch repair (MMR) status, chromosomal instability, or mutations in BRAF, KRAS, NRAS, and TP53 can influence prognosis, hereditary cancer risk assessment, and therapeutic approaches. Conventional biomarker testing via immunohistochemistry, polymerase chain reaction, and next-generation sequencing remains the diagnostic standard, but it can be limited by cost, turnaround time, tissue consumption, and uneven access. This narrative review synthesizes 30 peer-reviewed studies, identified across major scientific databases, that evaluate artificial intelligence (AI) platforms designed to detect molecular alterations from routine hematoxylin and eosin-stained colorectal cancer tissue slides. The strongest and most reproducible evidence exists for MSI phenotypes and MMR status, for which weakly supervised, attention-based, transformer-based, foundation-model, and clinically oriented multiple-instance learning systems achieve high discriminatory performance and particularly high negative predictive values at screening thresholds. BRAF mutation status is moderately predictable, although often through morphology associated with microsatellite instability or the CpG island methylator phenotype (CIMP). In contrast, despite promising single-center results, the predictability of KRAS, NRAS, PIK3CA, and other point mutations remains less consistently generalizable. Interpretability analyses indicate that these algorithms rely on features such as tumor-infiltrating lymphocytes, plasma cells, mucinous and medullary differentiation, necrosis, stromal architecture, tumor heterogeneity, nuclear morphometry, and tumor purity. Current evidence supports the use of AI as a triage and enrichment tool rather than a replacement for validated molecular diagnostics. Ultimately, prospective validation, pre-analytical standardization, calibrated thresholds, regulatory oversight, and pathologist-centered workflow integration are essential for successful clinical translation.
Keywords: colorectal cancer; digital pathology; MSI phenotype; MMR status; BRAF; KRAS; artificial intelligence; hematoxylin and eosin