Pre-Training and Ensembling of Deep Neural Networks for Target Gene Expression Prediction From Landmark Genes

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Gene expression profiling is crucial in biological and biomedical research.
• The L1000 assay efficiently measures landmark gene expression to predict target gene expression.
• Deep neural networks (DNNs) show promise for predicting gene expression from landmark genes.

Purpose of the Study:

• To improve the accuracy of DNNs for gene expression prediction using a novel pre-training and ensemble method.
• To enhance the performance of gene expression inference and multi-target regression.

Main Methods:

• Applied autoencoder-based pre-training to both the input and output of DNNs.
• Utilized multiple autoencoders with different random seeds for pre-training.
• Implemented an ensemble approach combining pre-trained DNNs.

Main Results:

• The proposed pre-training and ensemble method significantly outperformed state-of-the-art methods in accuracy.
• Pre-training improved DNN optimization by favorably positioning them in the function space.
• Pre-training diversified gene associations and highlighted more biologically relevant genes, especially with deeper networks.

Conclusions:

• The developed method offers superior accuracy for gene expression inference compared to existing techniques.
• Autoencoder pre-training and ensembling enhance DNN performance and biological relevance.
• The approach is potentially applicable to broader multi-target regression problems beyond gene expression.