A hybrid mixture approach for clustering and characterizing biomedical data

Wait 5 sec.

by Kazeem A. Kareem, Fan DaiModel-based clustering (MBC) is widely used for identifying latent groups and characterizing heterogeneous structures in biomedical data. However, modern biomedical studies often generate datasets with large sample sizes and many variables, for which standard MBC approaches can be difficult to fit reliably and efficiently. Mixture of factor analyzers (MFA) addresses part of this challenge by representing high-dimensional covariance structures through a few underlying factors, thereby reducing the covariance dimensionality. However, current estimation methods for MFA are often computationally impractical for massive data due to the slow convergence of nested EM-type algorithms. Moreover, many existing implementations require a common number of latent factors across clusters, limiting their ability to capture group-specific structures. We propose a hybrid matrix-free computational scheme for accurate and efficient estimation of Gaussian mixture of factor analyzers, which avoids explicit construction or storage of large covariance matrices. We further develop a generalized version that allows different clusters to have varying numbers of factors, providing additional flexibility for group-specific characterization. Simulation studies show that the proposed approach achieves substantially faster convergence than the existing method, especially in high-dimensional settings, while preserving high clustering and estimation accuracy. Applications to the Wisconsin breast cancer dataset, a lymphoma gene-expression dataset, and a human brain single-cell RNA sequencing dataset demonstrate accurate identification of benign/malignant breast tumors, lymphoma subtypes, and neuronal and astrocytic cell populations. The fitted factor-loading structures reveal biologically meaningful differences in nuclear morphology for breast tumors, with malignant samples showing stronger irregularity-related variation, group-specific gene-expression heterogeneity among the three lymphoma subtypes, and distinct patterns of transcriptional variation between neurons and astrocytes, with neuronal variation reflecting a broader set of coordinated transcriptional directions and astrocytic variation concentrated in a more compact latent structure.