Handling Ill-Conditioned Omics Data With Deep Probabilistic Models.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• High-throughput technologies generate high-dimensional omics data, posing challenges for machine learning due to the limited number of samples versus features.
• Dimensionality reduction is crucial for extracting relevant information and projecting data into a lower-dimensional space.
• Probabilistic latent space models are gaining traction for their ability to capture data structure and uncertainty.

Purpose of the Study:

• To develop a general classification and dimensionality reduction method for omics datasets.
• To address the challenges of missing data and the limited sample size relative to feature count in omics data.
• To propose a novel semi-supervised Bayesian latent space model for enhanced data analysis.

Main Methods:

• Introduced the Deep Bayesian Logistic Regression (DBLR) model, a semi-supervised Bayesian latent space model.
• DBLR infers a low-dimensional embedding driven by the target label and learns a global weight vector for predictions.
• Incorporated a probabilistic regularization method to mitigate overfitting, leveraging the model's semi-supervised nature.

Main Results:

• The DBLR model generated more informative low-dimensional representations compared to state-of-the-art methods.
• DBLR demonstrated superior performance in classification tasks across synthetic and real-world datasets.
• The model effectively handled missing data entries, a common issue in omics datasets.

Conclusions:

• The proposed DBLR model offers a robust solution for dimensionality reduction and classification of high-dimensional omics data.
• DBLR's ability to handle missing data and prevent overfitting makes it a valuable tool in bioinformatics.
• The model's performance indicates its potential for advancing omics data analysis and biological discovery.