Dominant COSMIC Mutational Signatures and Unsupervised Machine Learning Define Mechanism-Based Therapeutic Stratification in Blast-Crisis CML: Implications in Precision Medicine of Relapsed and Refractory Cancers

Background/Objectives: Blast-crisis chronic myeloid leukemia (BC-CML) represents a genomically unstable, therapeutically refractory phase driven by clonal evolution beyond BCR::ABL1 dependence. Clinical outcomes remain poor, with median survival under one year. Mutation-level analyses inadequately resolve biological heterogeneity or inform rational patient-specific therapy. We therefore implemented a mutational process–anchored framework integrating genomics, machine learning (ML), and artificial intelligence (AI)–guided drug prioritization at single-patient resolution. Methods: Whole-exome sequencing, COSMIC mutational signature deconvolution, unsupervised ML clustering, and AI-guided pharmacogenomic prioritization were applied to 224 patients with CML (190 chronic phase, 15 accelerated phase, 19 blast crisis) and 19 healthy controls. Results: Mutation burden increased significantly with disease progression, and blast crisis exhibited complex co-mutation architectures involving TP53, BRCA2, IDH1/2, DNMT3A, JAK2, and CSF3R. While gene-level patterns were heterogeneous, dominant mutational processes converged across cases. COSMIC signature extraction stratified BC-CML into three biologically coherent clusters: homologous recombination deficiency/genomic instability (Signatures 3/5), epigenetic–metabolic dysregulation (Signatures 1/2), and oxidative stress–associated cytokine signaling mutagenesis (Signatures 13/18). These process-defined states mechanistically link DNA repair failure, chromatin remodeling, and inflammatory signaling to blast transformation. Integration with an AI-guided pharmacogenomic framework translated signature biology into rational, cluster-specific therapeutic prioritization, including PARP inhibition, IDH targeting, CDK4/6 blockade, JAK pathway inhibition, and redox-modulating strategies. Dominant signature resolution reduced multidrug ambiguity and supported single-patient therapeutic parsimony. External concordance assessment across independent pharmacogenomic datasets supported predicted vulnerabilities. Conclusions: This study establishes a mechanistically interpretable, process-anchored precision oncology framework for BC-CML linking genomic architecture to actionable therapeutic decision-making warranting clinical validation. The approach can further be implemented for AL-AI-prioritized precision treatment of other relapsed, refractory and metastatic cancers.