Causal Discovery in High-dimensional, Multicollinear Datasets.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• High-throughput genomic sequencing generates vast datasets with numerous biological features.
• Discovering causal mechanisms of disease from complex data is challenging due to high dimensionality and multicollinearity.
• Existing causal discovery algorithms struggle with these inherent data complexities.

Purpose of the Study:

• To evaluate a novel non-orthogonal, general empirical Bayes matrix factorization approach for causal discovery in biomedical research.
• To address the limitations of traditional causal discovery algorithms, specifically the curse of dimensionality and multicollinearity.
• To identify latent factors and biological pathways associated with disease outcomes and clinical features.

Main Methods:

• Applied a non-orthogonal, general empirical Bayes matrix factorization method to causal discovery algorithms.
• Evaluated the strategy on simulated datasets to assess performance.
• Tested the approach on real-world breast cancer and SARS-CoV-2 (COVID-19) datasets.

Main Results:

• The strategy successfully inferred interpretable latent factors from observed variables.
• In breast cancer data, identified survival-associated latent factors and enriched pathways linked to clinical features.
• In the SARS-CoV-2 dataset, accurately predicted COVID-19 status and ICU admission, associating factors with known biological pathways.

Conclusions:

• The empirical Bayes matrix factorization approach offers a promising solution for causal discovery in high-dimensional biomedical data.
• This method enhances the interpretability of latent factors and their biological relevance.
• The findings demonstrate the utility of this strategy for understanding disease mechanisms and clinical outcomes in cancer and infectious diseases.