Generative Principal Component Regression via Variational Inference

Area of Science:

• Integrative biology
• Computational neuroscience
• Systems proteomics

Background:

• Latent variable models like Principal Component Analysis (PCA) are used to detect biological networks but struggle with noisy or low-variance phenotypes.
• Principal Component Regression (PCR) often underperforms traditional methods due to poor incorporation of phenotype information.
• Supervised Variational Autoencoders (SVAEs) improve prediction but introduce discrepancies between encoder and posterior distributions, potentially leading to flawed conclusions.

Purpose of the Study:

• Introduce generative principal component regression (gPCR), a novel objective for linear latent variable models.
• Enforce consistency between encoder and posterior distributions while maintaining predictive accuracy.
• Improve network detection and interpretation for phenotype-related biological studies.

Main Methods:

• Developed gPCR, a new objective function for linear latent variable models.
• Ensured consistency between encoder and posterior distributions.
• Validated gPCR using synthetic data, electrophysiology datasets, and a proteomics study of Alzheimer's disease.

Main Results:

• gPCR matches standard regression performance while retaining network interpretability.
• gPCR learns more realistic loadings and improves target selection compared to PCA and SVAEs on synthetic data.
• gPCR demonstrates enhanced predictive power and better integration of phenotype signals in electrophysiology data.
• gPCR recovers biologically coherent networks in Alzheimer's disease proteomics data, aligning with prior findings.

Conclusions:

• gPCR offers a robust approach for detecting phenotype-responsible networks across diverse scientific disciplines.
• The method enhances predictive accuracy and provides more reliable biological insights than existing techniques.
• gPCR is particularly valuable for complex datasets where phenotype information is crucial for network discovery.