PLASMA: Partial LeAst Squares for Multiomics Analysis.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• High-throughput "omics" technologies generate vast datasets, necessitating advanced methods for multi-omics data integration.
• Supervised learning methods for multi-omics integration, particularly for survival prediction, are less developed than unsupervised approaches.

Purpose of the Study:

• To introduce PLASMA, the first supervised algorithm capable of predicting time-to-event outcomes from multi-omics data.
• To address challenges in multi-omics data integration, including handling datasets with missing omics assays for some samples.

Main Methods:

• PLASMA employs a two-layer partial least squares (PLS) approach to identify components correlating with outcomes.
• It constructs a joint Cox proportional hazards model using these selected components.
• The algorithm can learn from multi-omics data even when samples have been assayed on only a subset of omics data.

Main Results:

• PLASMA successfully predicted risk stratification in stomach adenocarcinoma (STAD) and esophageal adenocarcinoma (ESCA) patient data.
• The model demonstrated superior performance compared to individually trained omics models and the unsupervised Multi-Omics Factor Analysis (MOFA) method.
• A negative comparison using dissimilar ESCA squamous cell carcinomas showed no significant risk separation (p = 0.57), validating PLASMA's specificity.

Conclusions:

• PLASMA provides a powerful supervised approach for multi-omics data integration and survival prediction.
• The identified predictive factors within the PLASMA model are biologically justifiable.
• This method advances the field of personalized medicine by enabling more accurate patient risk stratification.