Learning biologically-interpretable latent representations for gene expression data: Pathway Activity Score Learning...

Area of Science:

• Bioinformatics
• Computational Biology
• Systems Biology

Background:

• High-dimensional molecular gene-expression datasets pose challenges for analysis.
• Existing dimensionality reduction methods often lack direct biological interpretability.
• Identifying meaningful biological patterns in complex datasets is crucial.

Purpose of the Study:

• To introduce Pathway Activity Score Learning (PASL), a novel dimensionality reduction algorithm.
• To develop a method that generates interpretable features representing pathway activity.
• To enhance predictive performance and biological insight from gene expression data.

Main Methods:

• Developed the Pathway Activity Score Learning (PASL) algorithm for dimensionality reduction.
• Constructed features directly interpretable as pathway activity scores.
• Trained PASL on a large corpus (50,000 samples) to create a universal gene expression feature dictionary.
• Validated the dictionary on 35,643 held-out samples and applied it to 165 diverse disease datasets.
• Utilized the AutoML tool JADBio to assess information retention post-transformation.

Main Results:

• PASL demonstrates superior predictive performance compared to the state-of-the-art method (PLIER) on breast cancer and leukemia datasets.
• The universal feature dictionary built by PASL is validated for reconstruction accuracy.
• Predictive information is retained in the PASL-transformed feature space across diverse diseases.
• PASL's latent space offers straightforward biological interpretation, unlike traditional methods like PCA.

Conclusions:

• PASL provides an effective and interpretable approach to dimensionality reduction in gene expression data.
• The algorithm enhances predictive modeling and biological understanding of complex molecular data.
• PASL facilitates the creation of a universal feature representation applicable across various biological contexts.