Machine learning demonstrates that somatic mutations imprint invariant morphologic features in myelodysplastic...

Area of Science:

• Hematology
• Computational Biology
• Genomics

Background:

• Morphologic interpretation is standard for diagnosing myelodysplastic syndromes (MDS) but has limitations in reliability and integrating genetic data.
• The complex interplay between morphologic and genetic changes in MDS hinders clear diagnostic associations.

Purpose of the Study:

• To identify novel clinical subtypes of MDS using machine learning to uncover patterns between morphologic features and genomic events.
• To define distinct morphologic profiles within MDS patients and associate them with clinical characteristics and prognosis.

Main Methods:

• Sequencing of 1079 MDS patients to analyze bone marrow morphology and clinical features.
• Application of a machine-learning technique to identify co-occurrence patterns between 1929 identified somatic mutations and morphologic alterations.
• Validation of identified morphologic profile/genetic signature associations in an independent cohort.

Main Results:

• Five distinct morphologic profiles with unique clinical characteristics were defined.
• Higher-risk MDS patients predominantly clustered in profile 1 (77%), while lower-risk (LR) patients were distributed across profiles 2-5, characterized by specific features like pancytopenia or erythroid dysplasia.
• LR MDS patients were classified into 8 genetic signatures, with specific mutations (e.g., TET2, SF3B1) associated with distinct morphologic profiles.

Conclusions:

• Machine learning can identify nonrandom relationships between morphology and genotype in MDS, defining clinical features.
• This study provides the first comprehensive application of machine learning to elucidate interdependencies among genetic lesions, morphology, and clinical prognosis in MDS.
• The identified profiles and genetic signatures offer a novel framework for understanding MDS heterogeneity and improving patient classification.