Integrated machine learning, cosmic signatures, and AI-driven genomic profiling of highly heterogeneous blast crisis CML at single patient level resolution: Implication as a versatile risk stratification and precision oncology approach for refractory

Abstract Abstract Blast crisis chronic myeloid leukemia (BC-CML) is an aggressive, therapy-resistant phase with poor outcomes. We applied a pipeline combining whole-exome sequencing (WES), COSMIC mutational signatures, and machine learning (ML)-based clustering with drug repurposing to stratify patients and guide therapy. In 19 BC-CML, WES revealed higher mutational burden and distinct genomic features. ML identified three BC-CML subtypes, each associated with unique signatures and actionable pathways. Drug mapping prioritized FDA-approved agents for targeted therapy. This versatile model supports risk-based precision oncology in refractory and relapsed (R&R) cancers. Introduction BC-CML remains a biologically complex, clinically refractory disease stage despite tyrosine kinase inhibitors (TKIs), due to secondary oncogenic mutations and genomic instability [1]. Genomic profiling has revealed pan-cancer mutations converging on pathways in solid and high-grade hematologic tumors [2]. Combining WES with ML enables patient stratification and targeted therapy discovery [3]. We present an integrated WES–ML–COSMIC pipeline to identify BC-CML subgroups and therapies, with potential applicability to R&R malignancies. Methods We studied 157 CML patients (123 chronic-phase (CP),15 accelerated-phase (AP) and 19 BC-CML) after ethical approval [4]. Peripheral blood mononuclear cells were isolated, DNA extracted and sequenced on Illumina NovaSeq. Reads were aligned to GRCh38 using BWA-MEM [5], variants called with GATK and annotated with VEP and COSMIC [6]. ML clustering and PCA identified genomic subgroups [7], and mutational signatures were analyzed using SigProfilerExtractor [8]. Drug mapping used PanDrugs to prioritize FDA-approved agents [9]. Results We identified over 2,500 somatic mutations, with BC-CML showing a 54% higher burden than earlier phases. Missense mutations predominated, with hotspots on chromosomes 1, 7, 17, and 19. ML stratification revealed three BC-CML clusters: Cluster 1 (BRCA2/TP53 mutations) characterized by homologous recombination deficiency; Cluster 2 (IDH1/2, TET2) driven by epigenetic dysregulation; and Cluster 3 (JAK2, CSF3R) linked to cytokine signaling and oxidative stress. COSMIC signatures further distinguished clusters: Cluster 1 was enriched for Signatures 3 and 5, Cluster 2 for Signatures 1 and 2, and Cluster 3 for Signatures 13 and 18. Drug mapping aligned therapies to these clusters: PARP inhibitors and MDM2 antagonists for Cluster 1, IDH inhibitors and hypomethylating agents for Cluster 2, and JAK inhibitors for Cluster 3. Statistical analysis confirmed significant inter-cluster differences in mutational load and signature scores. These findings demonstrate that BC-CML is genetically heterogeneous, with distinct, actionable pathways in each subgroup. Discussion Our integrated WES–ML–COSMIC framework reveals the genetic heterogeneity of BC-CML and provides a clinically actionable model for precision oncology [2]. The three subgroups reflect divergent pathogenic mechanisms—genomic instability, epigenetic dysregulation, and cytokine-driven stress—each targetable with existing therapies [2]. Incorporating COSMIC signatures with actionable mutations enables precise stratification, supporting rational use of PARP inhibitors in BRCA/TP53-deficient cases, IDH inhibitors in epigenetically altered subtypes, and JAK inhibitors in cytokine-driven subtypes [3]. This stratification aligns with emerging regulatory priorities for biomarker-guided therapy in rare malignancies [9]. By mapping mutations to FDA-approved agents, this pipeline supports clinical translation of repurposed therapies for refractory BC-CML [4]. While limited by BC-CML sample size and requiring external validation, these findings illustrate the feasibility of genomic stratification and drug repurposing as a versatile, risk-based precision oncology framework applicable to other refractory and relapsed cancers [9]. ReferencesKwon HJ, et al. Mol Cancer. 2025;24(1):114.Cruz-Rodriguez N, Deininger MW. Blood. 2025;145(9):931-43.Herráiz-Gil S, et al. Appl Sci. 2025;15(5):2798.Awada H, et al. Cancers (Basel). 2023;15(8):2248.Li H, Durbin R. Bioinformatics. 2009;25(14):1754-60.DePristo MA, et al. Nat Genet. 2011;43(5):491-8.Pedregosa F, et al. J Mach Learn Res. 2011;12:2825-30Sondka Z, et al. Nucleic Acids Res. 2023;52(D1):D1210-D7.Mao Y, et al. Mol Cancer. 2025;24(1):123.