On tower and checkerboard neural network architectures for gene expression inference

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in genomics

Background:

• The L1000 platform enables cost-effective gene expression profiling by measuring ~1,000 landmark genes.
• Computational methods, such as Deep Neural Network-based Gene Expression inference (D-GEX), infer the expression of thousands of additional genes.

Purpose of the Study:

• To develop novel D-GEX architectures that enhance gene expression inference accuracy.
• To improve computational efficiency and reduce the memory footprint of gene expression inference models.

Main Methods:

• Proposed two novel D-GEX architectures utilizing network partitioning into towers.
• Implemented a checkerboard architecture with skip connections and five towers.
• Optimized the training protocol for improved inference performance.

Main Results:

• Achieved a significant improvement in gene expression inference quality.
• Reduced the average mean absolute error from 0.134 to 0.128.
• Demonstrated enhanced inference accuracy without increasing the number of trainable parameters.

Conclusions:

• The proposed D-GEX architectures increase gene expression inference accuracy.
• The novel models do not increase the model's memory footprint, overcoming a key limitation.
• This advancement facilitates more efficient and accurate gene expression profiling.