Background: Cancer imposes a significant global economic and social burden, emphasizing the urgent need for early diagnostic tools. Although various biomarkers, particularly microRNAs (miRNAs), have been identified, clinical validation remains insufficient. In this study, we characterized a novel serum-based 12-miRNA panel, S12micro, developed through machine learning algorithms for pan-cancer diagnosis. Material and methods: Serum miRNA profiles from 8 public datasets comprising 20 794 samples (13 161 healthy controls and 7633 patients across 13 cancers) were processed with a standardized pipeline. We evaluated 113 pipelines generated from 12 ML algorithms (feature selection + classifier) under strict separation of development and validation. The final operating threshold was fixed on the training cohort and applied unchanged to all validation cohorts. KEGG enrichment was performed on target genes of the 12 miRNAs. Results: S12micro (12 miRNAs) achieved the highest mean area under the curve (AUC) across large-scale validation datasets ( N ≥ 1000; mean AUC = 0.987). In small-scale validation cohorts ( N < 1000), S12micro showed AUCs between 0.900 and 1.000 in 31/33 datasets. Beyond ROC analysis, we also evaluated the model using precision-recall curve analysis, decision curve analysis, and the Youden Index. Against previously published regression-specified serum miRNA signatures, S12micro showed superior average performance in the validation datasets. KEGG enrichment analysis revealed enrichment of canonical cancer pathways among predicted targets of S12micro. In-house RT-qPCR feasibility testing across four cancer types supported technical deployability (AUCs ≥ 0.760). Analysis of The Cancer Genome Atlas (TCGA) datasets suggested that these 12 miRNAs are much less abundant in solid tumors than in serum, supporting a serum-oriented diagnostic potential. Conclusion: S12micro is a promising, noninvasive serum signature that discriminates cancers from healthy controls across diverse cohorts. While encouraging, real-world specificity versus benign diseases and cross-platform robustness require prospective, multicenter validation with harmonized preanalytics and normalization.
Long et al. (Mon,) studied this question.