Investigating the effect of dependence between conditions with Bayesian Linear Mixed Models for motif activity...

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Gene regulatory networks control cellular identity and behavior through transcription factors (TFs) binding to DNA.
• Modeling TF influence on gene expression typically assumes linear relationships and sample independence, often using Ridge Regression.
• The independence assumption may be violated in biological samples from the same source or similar experimental conditions, potentially leading to signal detection loss.

Purpose of the Study:

• To investigate whether a Bayesian model accommodating sample correlations improves the accuracy of inferring TF motif activities compared to standard linear models.
• To evaluate the performance of a Bayesian Linear Mixed Model against Ridge Regression under different noise structures.

Main Methods:

• Extended Ridge Regression to a Bayesian Linear Mixed Model to incorporate dependencies between samples.
• Conducted simulation studies to compare the Bayesian Linear Mixed Model with Ridge Regression.
• Analyzed four real biological datasets to assess model performance on empirical data.

Main Results:

• The Bayesian Linear Mixed Model outperformed Ridge Regression in simulations when noise was uncorrelated.
• No performance gain was observed when noise shared a similar covariance structure with motif signals, with a mathematical explanation provided.
• Analysis of real datasets indicated that linear models explained at most ~40% of the signal, and the Bayesian model offered no advantage due to covariance structure similarity.

Conclusions:

• While Bayesian models can improve motif activity inference by accounting for sample correlations, their advantage is contingent on the noise structure.
• The effectiveness of incorporating sample correlations depends on whether the noise covariance mirrors the signal covariance.
• Real-world biological data may exhibit complex covariance structures limiting the benefits of advanced models over simpler linear approaches.