Project Overview 

This project applied unsupervised machine learning to analyze human retinal transcriptome data, aiming to identify genes governed by circadian rhythms. The ultimate goal is to create a "diurnal atlas" of the human retina, which can guide the development of circadian therapeutics—optimizing drug administration timing—for treating eye diseases like Age-related Macular Degeneration (AMD).

Objective & Motivation 

Age-related Macular Degeneration (AMD) is a leading cause of irreversible vision loss. Biological processes, including those in the retina, follow a 24-hour cycle known as the circadian rhythm. Aligning medical treatments with this natural clock (chronotherapy) can significantly enhance drug efficacy and reduce side effects.

The primary objective was to identify which genes in the human retina exhibit circadian behavior and to understand how this behavior changes in diseased states. This research forms the foundation for timing AMD treatments to be most effective.

Human eye with a schematic enlargement of retina & Outline of rhythmic functions in human circadian pacemaker over the 24h cycle 

Methodology 

• 1. Data Sourcing: The analysis was performed on four publicly available human retinal transcriptome datasets. The initial model was trained using a set of 659 "seed genes" known to have daily oscillatory patterns, derived from 30 human samples.

  2. Machine Learning Model: We employed CYCLOPS, an unsupervised autoencoder algorithm designed specifically to identify rhythmic signals in biological data without needing time-stamped samples for every individual.

  3. Analysis & Gene Extraction:

     • The CYCLOPS algorithm was applied to the transcriptome datasets to estimate the rhythmic phase of all genes.

     • This allowed us to extract genes that follow a 24-hour rhythm ("clock genes").

     • We then compared the rhythmicity of these genes across four sets of macular samples, ranging from healthy (control) to various stages of macular disease.

Key Findings & Results 

• Clock Gene Identification: Successfully extracted approximately 34 core clock genes from the retinal datasets that demonstrated significant circadian rhythmicity.

• Rhythm Disruption in Disease: A crucial finding was that the strength of the circadian rhythm in these genes progressively weakened in the more diseased macular samples. This suggests a strong link between the disruption of the eye's natural clock and the progression of macular degeneration.

• Proof of Concept: The project successfully demonstrated that machine learning can effectively model and predict circadian gene behavior from static, large-scale biological datasets.

Conclusion & Impact 

This project serves as a critical step toward creating a comprehensive map of the human retinal diurnal transcriptome. Such an atlas would be an invaluable resource for ophthalmology, enabling researchers and clinicians to explore and implement circadian therapeutic strategies. By understanding the optimal time to administer drugs, we can move towards more personalized and effective treatments for debilitating eye diseases.