Covering Hierarchical Dirichlet Mixture Models on binary data to enhance genomic stratifications in onco-hematology

Area of Science:

• Onco-hematology
• Statistical modeling
• Genomics

Background:

• Genomically-driven classification systems are advancing blood cancer research.
• Personalized medicine in onco-hematology relies on enhanced patient stratification.
• Hierarchical Dirichlet Mixture Models (HDMM) are commonly used for clustering genomic data in onco-hematology.

Purpose of the Study:

• To propose alternative approaches for characterizing HDMM components and assigning patients.
• To improve patient stratification in onco-hematology beyond traditional methods.
• To evaluate the efficacy of Multivariate Fisher's Non-Central Hypergeometric (MFNCH) distributions for HDMM component estimation.

Main Methods:

• Utilizing Hierarchical Dirichlet Mixture Models (HDMM) for genomic data clustering.
• Proposing two methods for HDMM component parameter estimation: multinomial and MFNCH-based.
• Developing a patient assignment strategy focused on identifying the most likely component for each patient.

Main Results:

• The MFNCH-based approach for patient assignment demonstrated comparable or superior performance to the multinomial approach on simulated data.
• The study illustrates the MFNCH-based approach's effectiveness on real Acute Myeloid Leukemia data.
• MFNCH-based approach offers a balance between rigorous statistical characterization and clinical knowledge refinement.

Conclusions:

• The MFNCH-based approach provides a valuable alternative for patient stratification in onco-hematology.
• This method enhances the application of HDMM in blood cancer genomics.
• The findings contribute to advancing personalized medicine in onco-hematology.