Clinical bottom line
Artificial intelligence (AI) platforms are emerging as promising tools to predict the tumor cell-of-origin from molecular and pathological data, potentially aiding diagnosis and guiding therapy in challenging cases. Current evidence supports AI models trained on somatic mutation profiles and single-cell transcriptomics to infer tumor origin with improving accuracy. However, these approaches remain largely investigational and require further validation in diverse clinical cohorts. Pathologists should consider AI-based cell-of-origin predictions as complementary to, not replacements for, conventional histopathology and immunophenotyping. Integration of AI insights may enhance diagnostic precision in tumors with ambiguous or mixed differentiation, thereby informing prognosis and treatment decisions.
What the evidence shows
Tumor cell-of-origin identification is a critical but sometimes elusive aspect of pathological diagnosis, especially in poorly differentiated or mixed tumors. Traditional methods rely on morphology and immunohistochemistry, which can be inconclusive in certain cases. Recent advances have leveraged AI to analyze complex molecular data for cell-of-origin prediction.
Vellanki Saicharan et al. (2026) developed Tumor-Origin.com, a machine learning platform that predicts tumor tissue of origin using somatic mutation profiles. This approach demonstrated promising accuracy across multiple tumor types, suggesting that mutation patterns contain informative signals about cell lineage. The model’s ability to classify tumors with ambiguous histology could improve diagnostic confidence and guide targeted therapy.
Single-cell RNA sequencing combined with AI has further refined cell-of-origin identification. Zhao Jimmy et al. (2020) applied single-cell analysis to neuroendocrine prostate cancer, revealing heterogeneity and distinct cells of origin within tumors. This granular resolution highlights the potential for AI to dissect tumor complexity beyond bulk tissue analysis.
Pathological studies also underscore the biological plausibility of shared origins in tumors with mixed features. For example, a cutaneous tumor exhibiting both squamous cell and neuroendocrine (Merkel cell) characteristics was shown to likely arise from a common progenitor cell (2005). Such findings support the rationale for AI models that integrate multi-omic data to predict lineage relationships.
Collectively, these studies illustrate that AI-driven cell-of-origin prediction is feasible and may complement traditional pathology by providing molecular context, especially in diagnostically challenging cases.
Caveats and uncertainty
Despite encouraging results, AI-based tumor cell-of-origin prediction faces several limitations. Most models have been trained and validated on retrospective datasets with limited sample diversity, raising concerns about generalizability to real-world clinical populations. The accuracy of predictions can vary depending on tumor type, mutation burden, and data quality.
Interpretability remains a challenge; AI models often function as “black boxes,” making it difficult to understand the biological basis for predictions or to reconcile discrepancies with histopathological findings. This limits immediate clinical trust and uptake.
Furthermore, the clinical utility of cell-of-origin predictions in guiding treatment decisions has not been prospectively demonstrated. Integration into diagnostic workflows requires standardized protocols, regulatory approvals, and demonstration of added value beyond existing methods.
Finally, single-cell approaches, while powerful, are resource-intensive and not yet widely available in routine pathology practice.
How this may change practice
As AI models for tumor cell-of-origin prediction mature, they may become valuable adjuncts in pathology, particularly for tumors with ambiguous morphology or mixed differentiation. Incorporating AI predictions could refine diagnostic accuracy, reduce diagnostic delays, and better stratify patients for targeted therapies.
Pathologists should remain informed about emerging AI tools and their validation status, critically evaluating AI outputs alongside traditional histology and immunophenotyping. Multidisciplinary collaboration will be essential to interpret AI-derived insights in the clinical context.
In the near term, AI-driven cell-of-origin prediction may be most useful in academic and reference centers with access to molecular profiling and bioinformatics expertise. Over time, streamlined platforms integrating mutation data and AI could become part of routine diagnostic panels, enhancing personalized oncology.