Relationship between gene regulation network structure and prediction accuracy in high dimensional regression

Area of Science:

• Genomics
• Bioinformatics
• Systems Biology

Background:

• High-dimensional omics data, like transcriptomes, require robust estimation methods.
• Prediction accuracy in methods like lasso and PCR depends on gene regulation network structures.
• The impact of gene network structure and sample size on prediction accuracy remains under-investigated.

Purpose of the Study:

• To investigate how gene regulation network structure and sample size influence the prediction accuracy of lasso and PCR.
• To compare the performance of lasso and PCR under different network topologies and sample sizes.

Main Methods:

• Monte Carlo simulations were employed to assess prediction accuracy.
• Various gene regulation network structures (random graph, scale-free) were simulated.
• Different sample sizes were tested in conjunction with these network structures.

Main Results:

• Lasso (least absolute shrinkage and selection operator) requires a large sample size for good prediction accuracy with random gene networks.
• Principal component regression (PCR) demonstrated consistently poor prediction accuracy, irrespective of sample size.
• For scale-free gene networks, characteristic of real biological systems, lasso achieved accurate trait prediction with a relatively small number of observations.

Conclusions:

• Gene network topology significantly impacts the performance of omics data analysis methods.
• Lasso is a promising method for trait prediction from transcriptomes, particularly when gene networks exhibit scale-free properties.
• The findings suggest that efficient trait prediction from transcriptomic data is feasible with appropriate methods and consideration of network structure, even with limited sample sizes.