Network depth affects inference of gene sets from bacterial transcriptomes using denoising autoencoders.

Area of Science:

• Computational Biology
• Genomics
• Systems Biology

Background:

• Publicly available bacterial gene expression datasets are growing, necessitating automated methods for hypothesis generation.
• Denoising autoencoders (DAEs) are neural networks used for inferring coordinated gene regulation from expression data by encoding datasets into a reduced layer.

Purpose of the Study:

• To generalize DAEs for deep networks and explore the impact of network architecture on gene set inference.
• To develop and validate a DAE-based pipeline for extracting gene sets from transcriptomic data in Escherichia coli.
• To identify genes uniquely induced during human colonization in uropathogenic E. coli.

Main Methods:

• Developed a DAE-based pipeline for gene set extraction from E. coli transcriptomic data.
• Validated the method by comparing inferred gene sets with known biological pathways.
• Investigated the effects of varying DAE network depth and width on gene set recovery and biological inference.

Main Results:

• Increasing network depth in DAEs led to more concise gene set definitions.
• Network width adjustment presented a trade-off between generalizability and biological inference accuracy.
• Applied the optimized pipeline to identify genes specifically induced during human colonization in an independent uropathogenic E. coli dataset.

Conclusions:

• The DAE architecture significantly influences the biological insights derived from gene expression data.
• The developed pipeline effectively extracts biologically relevant gene sets and can identify condition-specific gene induction.
• This approach advances automated hypothesis generation in bacterial gene regulation studies.