Identifying molecular biomarkers is essential for cancer diagnosis and prognosis. Traditional bioinformatics approaches often require specialized expertise, limiting interdisciplinary collaboration between computer scientists and biologists. We present an artificial intelligence (AI)-based framework that uses machine learning algorithms to identify influential miRNAs in bone marrow cancer without specialized bioinformatics tools.

miRNA expression data from 110 TCGA-LAML samples with 1,881 features were analyzed. Thirty-five classification algorithms were used to evaluate miRNA importance. We introduce the CIMPS (Consensus Importance Measure with Penalty for Similarity) method, which integrates weighted F1-score performance, feature stability, statistical significance, and model diversity to compute final importance scores.

To address overfitting in this high-dimensional setting (n=110, p=1,881), a conservative three-stage approach was used: models with F1-scores <0.88 (7 models) and >0.93 (20 models) were excluded, leaving 8 stable models (F1-scores 0.88–0.93) for analysis. Top miRNAs were mapped to target genes using Mirwalk, followed by pathway enrichment with Enrichr.

Key findings include hsa-mir-6751 (CIMPS score: 1.000) and hsa-mir-3681 (CIMPS score: 0.892), which align with critical pathways, including p53 signaling, PI3K-Akt, JAK-STAT, and hematopoietic cell lineage. The proposed CIMPS framework offers a robust, accessible approach to miRNA biomarker discovery and fosters effective collaboration between computational and biological researchers in precision oncology.