Gene expression clock: an unsupervised deep learning approach for predicting circadian rhythmicity from whole genome...

Area of Science:

• Chronobiology
• Systems Biology
• Bioinformatics

Background:

• Circadian rhythms, regulated by molecular clocks, govern physiological and behavioral processes.
• Disruptions in circadian rhythms are linked to various health problems, necessitating their study.
• Analyzing gene expression data for circadian rhythms is challenging due to the lack of time-stamped samples.

Purpose of the Study:

• To propose a novel deep neural network (DNN) approach for predicting sample phases from untimed gene expression data.
• To address the challenges of identifying cyclic genes and handling small datasets in circadian rhythm research.
• To enable the study of circadian rhythms without relying on pre-identified circadian genes.

Main Methods:

• Utilized a Minimum Distortion Embedding framework for initial candidate cyclic gene screening.
• Employed greedy layer-wise pre-training of a DNN to initialize hidden layers and capture gene profile features from limited samples.
• Fine-tuned the pre-trained DNN for accurate sample phase prediction.

Main Results:

• Demonstrated accurate and robust prediction of sample phases and cyclic genes across animal and human datasets.
• Successfully identified cyclic genes without prior knowledge of conserved circadian genes.
• Validated the approach on an Alzheimer's disease (AD) dataset, identifying disrupted oscillating genes in AD patients.

Conclusions:

• The proposed DNN approach effectively predicts circadian phases and identifies cyclic genes from untimed gene expression data.
• This method offers a valuable tool for studying circadian biology, especially with limited sample sizes.
• The findings in the AD dataset highlight potential biomarkers and therapeutic targets related to circadian rhythm disruption in neurodegenerative diseases.