An in-depth comparison of linear and non-linear joint embedding methods for bulk and single-cell multi-omics

Area of Science:

• Computational biology
• Bioinformatics
• Genomics

Background:

• Multi-omic analyses are crucial for understanding complex biological systems and predicting disease outcomes.
• Existing linear methods for joint embedding have limitations, prompting interest in non-linear neural network approaches.

Purpose of the Study:

• To conduct a head-to-head comparison of linear and non-linear joint embedding methods.
• To evaluate these methods on both bulk and single-cell multi-modal datasets.
• To provide guidance on selecting appropriate joint embedding techniques for specific downstream tasks.

Main Methods:

• Comparison of linear and non-linear joint embedding techniques.
• Utilized both bulk and single-cell multi-modal datasets.
• Evaluated performance on survival analysis (bulk data) and cell type classification (single-cell data).

Main Results:

• Non-linear methods demonstrated a significant advantage in imputing missing data modalities compared to linear methods.
• Concatenating principal components served as a strong baseline when all modalities were available.
• Joint embeddings from non-linear methods improved performance when only one modality was available at test time.
• Imputed omics profiles from neural methods were sufficiently realistic for downstream classification tasks with minimal performance loss.

Conclusions:

• Non-linear joint embedding methods, particularly product-of-experts, offer superior performance for multi-omic data integration and imputation.
• The choice of embedding method impacts downstream task performance, with non-linear approaches showing benefits in scenarios with missing data or limited modalities.
• Neural joint embedding methods facilitate realistic data imputation, enabling robust predictions in biological analyses.