The effect of non-linear signal in classification problems using gene expression.

Area of Science:

• Computational biology
• Genomics
• Machine learning

Background:

• Predictive modeling of transcriptomic data faces conflicting approaches: complex non-linear models (e.g., neural networks) versus simpler, interpretable linear models.
• The choice depends on whether biological systems are best represented by high-dimensional complexity or simple predictive boundaries.

Purpose of the Study:

• To compare the performance of multi-layer neural networks and logistic regression for prediction tasks using transcriptomic data.
• To investigate the presence and impact of non-linear signals in gene expression data.
• To determine the necessity of linear models as baselines in predictive modeling of biological systems.

Main Methods:

• Comparative analysis of multi-layer neural networks and logistic regression.
• Utilized GTEx and Recount3 datasets for prediction tasks.
• Employed Limma to remove linear predictive signals and assess the impact on model performance.

Main Results:

• Non-linear signals were identified in predicting tissue and metadata sex labels from expression data.
• Removing linear signals impaired linear models but not non-linear ones.
• Non-linear signals did not consistently guarantee superior performance of neural networks over logistic regression.

Conclusions:

• Multi-layer neural networks can be valuable for gene expression data predictions.
• Including a linear baseline model is crucial for robust predictive modeling.
• Effective predictive dividing lines in high-dimensional biological data may not always be non-linear.