Artificial-Intelligence, Data-Driven, Comprehensive Classification of Myeloid Neoplasms Based on Genomic, Morphological and Histological Features

Background. Recent advancements in genome characterization have transformed the study of myeloid neoplasms (MN). Accordingly, there is a shift from traditional classification schemes, based on morphological and clinical information, to next-generation systems (2022 WHO/ICC) which incorporate genomic features. However, current classifications do not clearly define the hierarchical importance of genomic vs morphological features in determining disease entities, leading to inconsistencies in clinical practice; moreover, they do not specifically address areas of overlap among different MN

Methods. In the TITAN study, we retrospectively studied 20,054 patients with MN in which clinical and morphological data, together with cytogenetics, mutational screening and outcome were available. We included 7104 AML, 8410 MDS, 2986 MDS/MPN and 1554 MF. We used MOSAIC, an AI-based framework to define unsupervised clusters according to homogeneous morphological and genomic features. Shapley Additive exPlanations (SHAP) was applied to investigate features of importance and their effects on cluster assignment

Results. We identified 34 distinct clusters and created a dashboard (https://titan-xkb3corsxq-ew.a.run.app/) where the most relevant features, clinical phenotypes, and outcomes are summarized. SHAP revealed that genomics often outweighs morphology in defining disease entities

Six clusters had splicing mutations as dominant features, albeit with different hematological phenotypes (mainly including MDS and MDS/MPN pts). The presence of additional high-risk genomic features (RUNX1/ASXL1 mutations, del(7q)/-7, abn(3q26.2) and complex karyotype) identified patients with poor survival within each category. A separate cluster including MN with splicing mutations and increased blasts (mainly including MDS and AML patients) was characterized by poor response to treatments and dismal outcome

MN with TP53 mutations constituted a distinct entity, regardless of clinical phenotype and blast count and had the poorest prognosis. Mutational and CNV analysis showed biallelic inactivation of TP53 in the majority of patients. Most monoallelic TP53 were assigned to a different cluster. Patients that evolved to AML often acquired biallelic inactivation, suggesting that TP53 allelic status may identify different disease phases within the same entity

Among morphological features, marrow fibrosis held significant hierarchical importance in defining clusters, driving patient assignment into three distinct groups characterized by 1) JAK/STAT mutations alone, 2) JAK/STAT mutations with high-risk molecular features, 3) wildtype JAK/STAT. The latter group included both MDS and MF patients who exhibited a higher risk of clonal evolution and shorter survival

Overall, 10 out of 34 clusters included patients from different MN subtypes as defined by ICC/WHO criteria, accounting for 38.1% of the entire population. This suggests that current classifications do not efficiently capture boundaries between different entities. Importantly, among 1482 patients with RNAseq data from BM progenitors available, the newly identified clusters exhibited distinct gene expression profiles, thus providing evidence for the biological consistency of the proposed classification

The 34 clusters were characterized by different risks of clonal progression and survival. Based on this finding and the observation that, in most cases, heterogeneous phenotypes contributed to each cluster, we developed a pan-MN prognostic score. This score is based on clinical and molecular parameters and is designed to be applicable to all patients, regardless of clinical labels. The performance of the score (c-index 0.75) was comparable to that of currently available disease-specific tools

Conclusion. The integrative analysis of genomic profiling and morphological features provides a proof of concept for a novel classification approach for MN, that may refine the overlapping areas among disease entities and more effectively identify patients with a homogeneous biological background. It could serve as a foundation for an innovative prognostic assessment of MN patients, independent of traditional clinical labels. The clinical implementation of pan-MN tools is expected to improve the effectiveness and quality of personalized diagnosis, treatment decisions and clinical trial eligibility