Praxis-BGM: Clustering of Omics Data Using Semi-Supervised Transfer Learning for Gaussian Mixture Models via...

Area of Science:

• Computational biology
• Machine learning
• Bioinformatics

Background:

• High-dimensional omics data present challenges for model-based clustering due to limited sample sizes.
• Gaussian mixture models (GMMs) often suffer from instability and poor generalization with complex mixture structures in omics data.
• Existing methods struggle with small sample sizes and complex data structures in omics analysis.

Purpose of the Study:

• To develop a novel framework, Praxis-BGM, for robust GMM clustering in high-dimensional omics data.
• To enable semi-supervised transfer learning by incorporating informative prior GMMs from large-scale reference data.
• To improve clustering performance, stability, and biological interpretability in small-sample omics settings.

Main Methods:

• Developed Praxis-BGM, a natural-gradient variational inference framework for GMMs.
• Utilized the Variational Online Newton (VON) algorithm for natural-gradient updates.
• Implemented the framework in JAX for efficient, scalable computation.

Main Results:

• Praxis-BGM demonstrated improved posterior clustering performance and stability in simulations and real-world applications.
• Successfully applied to breast cancer transcriptomics for subtype recovery and single-cell transcriptomics for cell-type label transfer.
• Showcased enhanced biological interpretability even with partially mismatched prior models.

Conclusions:

• Praxis-BGM offers a computationally efficient and scalable solution for GMM clustering with high-dimensional omics data.
• The framework effectively leverages transfer learning from reference data to enhance clustering in small-sample target datasets.
• Praxis-BGM provides a valuable tool for advancing omics data analysis, subtype discovery, and cell-type classification.