cpiVAE: Robust and Interpretable Cross-Platform Proteomics Imputation.

Area of Science:

• Biochemistry
• Bioinformatics
• Genomics

Background:

• Plasma proteomic studies frequently employ diverse high-throughput platforms, leading to discordant measurements for identical proteins.
• This data discordance impedes effective cross-study integration and limits the power of meta-analyses and biomarker discovery.
• Integrating proteomics data is crucial for enhancing statistical power and deepening the understanding of proteome-phenotype relationships.

Purpose of the Study:

• To develop a deep generative model for bidirectional imputation of protein abundances between Olink and SomaScan platforms.
• To establish a method that improves cross-platform proteomics data integration and enables more powerful downstream analyses.
• To provide an interpretable and generalizable solution for harmonizing proteomics data from different experimental platforms.

Main Methods:

• Development of a cross-platform proteomics imputation variational autoencoder (cpiVAE), a deep generative model.
• Training the cpiVAE model using paired plasma proteomic measurements from the China Kadoorie Biobank (CKB) cohort.
• Evaluating cpiVAE performance against established methods like k-nearest neighbors (KNN) and Weighted Nearest Neighbors (WNN) on independent datasets.

Main Results:

• cpiVAE achieved up to 30% higher correlation between imputed and true protein values compared to KNN and WNN benchmarks.
• The model demonstrated strong generalization to an independent Atherosclerosis Risk in Communities Study (ARIC) cohort without retraining.
• Imputed protein levels accurately reflected associations with clinical phenotypes and enhanced statistical power in meta-analysis simulations.
• A post-hoc feature importance matrix provided interpretability, with extracted protein pair features overlapping with known biological interactions in the STRING database.

Conclusions:

• cpiVAE offers an accurate, generalizable, and interpretable solution for cross-platform proteomics imputation.
• The framework facilitates integrated analyses across studies utilizing different proteomics measurement platforms.
• Open-source availability of the cpiVAE framework and pre-trained model weights promotes wider adoption and data integration in proteomic research.