Basic Science and Pathogenesis.

Area of Science:

• Genomics
• Proteomics
• Bioinformatics
• Neurodegenerative Diseases

Background:

• Growing availability of large genetic and proteomic datasets necessitates scalable, integrated omics analysis methods for Alzheimer's disease (AD).
• Eight distinct genomic and proteomic analysis tools have been developed for the AD community, accessible via the Global Research and Imaging Platform (GRIP).
• These tools aim to enhance accuracy and interpretability in diverse AD cases, covering network discovery, confounder removal, ancestry inference, variant annotation, and data harmonization.

Purpose of the Study:

• To introduce and highlight two novel methods: generative PCR (gPCR) for phenotypically relevant network detection and FairPCA for confounder effect removal.
• To demonstrate the utility of gPCR in learning AD-associated proteomic networks.
• To showcase FairPCA's effectiveness in removing confounding factors like sex from omics data and in genome-wide association studies (GWAS).

Main Methods:

• gPCR, an improvement over supervised variational autoencoders (SVAEs), generates phenotypically relevant latent components for omics data.
• FairPCA, an adversarial learning algorithm, ensures principal components are orthogonal to undesired confounders and scales to large datasets using randomization.
• Both methods were evaluated on proteomic data (SomaScan, TMT-MS) from 300 AD samples and in simulated GWAS.

Main Results:

• gPCR achieved superior predictive performance (AUC=0.94) compared to traditional PCR (AUC=0.62) and matched elastic net regression (AUC=0.94) while identifying relevant protein networks.
• FairPCA demonstrated improved AD status prediction (AUC=0.83) over traditional PCA (AUC=0.64) when removing sex as a confounder.
• In GWAS simulations, FairPCA significantly improved the accuracy of effect size estimation (cosine similarity from 0.71 to 0.99).

Conclusions:

• gPCR preserves predictive power while enhancing generative capabilities for omics analysis.
• FairPCA effectively simplifies the removal of confounding factors in large-scale omics datasets.
• These tools, along with others developed for GRIP, are poised to significantly accelerate scientific discovery in Alzheimer's disease research.