CRISPR dependency profiles distinguish pediatric from adult cancer cell lines via explainable machine learning: Identification of GPX4 and proliferative genes as therapeutic targets

Abstract Pediatric and adult cancers diverge profoundly in their molecular architecture, yet whether these differences extend to genome-scale genetic dependencies remains poorly characterized, with prior studies examining single subtypes or individual algorithms rather than a systematic, interpretable comparison. To address this gap, we tested whether Clustered Regularly Interspaced Short Palindromic Repeats (CRISPR)-Cas9 loss-of-function dependency profiles can discriminate pediatric from adult cancer cell lines, and applied explainable artificial intelligence (XAI) to identify the genes driving this distinction. Chronos-normalized CRISPR GeneEffect scores for 18,531 genes across 1208 cell lines (242 pediatric, 966 adult) were analyzed; class imbalance was corrected with the Synthetic Minority Over-sampling Technique (SMOTE) within each training fold, differentially dependent genes were identified by Mann-Whitney U tests with Benjamini-Hochberg False Discovery Rate (FDR) correction, and 200 features were retained by SelectKBest. Eleven classifiers were benchmarked under 5-fold cross-validation across 11 metrics, and predictions were interpreted with SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) In this first (to our knowledge) genome-scale, multi-classifier comparison Light Gradient Boosting Machine (LightGBM) achieved the best discrimination (Area Under the Receiver Operating Characteristic Curve [ROC-AUC] = 0.909, accuracy = 88.7%, Matthews Correlation Coefficient [MCC] = 0.632, Brier score = 0.089), with the leading ensemble and margin-based models reaching ROC-AUC values near 0.90. Among 3602 differentially dependent genes (FDR < 0.05), Integrin-Linked Kinase (ILK), Talin-1 (TLN1), and Aurora Kinase A (AURKA) showed greater dependency in pediatric lines, whereas Glutathione Peroxidase 4 (GPX4) and Insulin-like Growth Factor 2 messenger RNA-Binding Protein 1 (IGF2BP1) were comparatively more essential in adult lines. SHAP highlighted CHMP4B, IGF2BP1, GPX4, and LDB1, and the representative true-positive LIME case emphasized IGF2BP1, PAQR6, and CDS2. By exposing dependencies that adult-centric precision-oncology programs overlook, this work provides a framework for age-specific target discovery and prioritizes the pediatric-enriched ILK–TLN1 mechanotransduction axis for experimental validation.