Standard machine learning approaches outperform deep representation learning on phenotype prediction from...

Area of Science:

• Molecular diagnostics
• Genomics
• Bioinformatics

Background:

• Predicting health outcomes from gene expression is crucial for molecular diagnostics but remains challenging.
• Developing robust and reproducible predictive signatures for phenotypes has not been achieved across most disease areas.
• This study comprehensively analyzes prediction tasks across multiple diseases, including ulcerative colitis, atopic dermatitis, diabetes, and various cancer subtypes.

Purpose of the Study:

• To systematically investigate factors influencing gene expression-based phenotype prediction.
• To evaluate the impact of gene subsets, normalization methods, and prediction algorithms.
• To explore deep representation learning for enhancing predictive performance on large transcriptomics datasets.

Main Methods:

• Analysis of 24 binary/multiclass prediction problems and 26 survival analysis tasks.
• Systematic investigation of gene subsets, normalization techniques (e.g., centered log-ratio transformation), and prediction algorithms.
• Application of deep representation learning on large transcriptomics compendia (GTEx, TCGA).

Main Results:

• Combining large numbers of genes consistently outperformed single-gene methods.
• Unsupervised and semi-supervised representation learning did not consistently improve out-of-sample performance.
• L2-regularized regression methods applied to centered log-ratio transformed transcript abundances provided the best overall predictive analyses.

Conclusions:

• Proper normalization and state-of-the-art regularized regression are key for transcriptomics-based phenotype prediction.
• Breakthrough performance may depend on reducing sequencing data errors and integrating multi-omics data (single-cell sequencing, proteomics).
• Improved utilization of prior knowledge is also suggested as a critical factor for advancing predictive capabilities.