Classification of cellular age using single-cell ATAC-seq data from the mouse hypothalamus
A project for Brown's Deep Learning couse taught by Prof Ritambhara Singh, implemented by apavulu2, jpatrico, and tpham43.
The Webb Lab at Brown recently published a groundbreaking single-cell RNA-seq dataset of the female mouse hypothalamus. Neurons of the hypothalamus regulate homeostatic processes and hormone production, which are dysregulated with age. Sex-specific differences in lifespan, including different hallmarks of aging, have been well documented in humans and many other species, and humans assigned female at birth have a significantly higher incidence of neurodegenerative disease. However, the mechanisms of aging in females remain understudied.
Subsequently, members of the Webb Lab have generated multiomic data of the hypothalamus at single-cell resolution in male and female mice (currently unpublished). This new data will allow us to analyze the chromatin accessibility component of aging hallmarks of neurons in the hypothalamus alongside previously published gene expression data. Moreover, the inclusion of male mice in this dataset allows us to uncover sex-specific differences in the aging hypothalamus.
Single-cell resolution ATAC-seq data consists of a very large peak by cell matrix, in which each cell is a sample and each peak is a feature. There are upwards of 250,000 features in this dataset. For this reason, we will focus on the fragments of the X chromosome (ChrX) as it is a sex chromosome that is of utmost importance in aging and to the brain. This lowers the feature number to a manageable size. The dataset will require extensive preprocessing using the Seurat and Signac libraries in R. Some bottlenecks we foresee with this dataset is the significant sparsity that comes with the territory of single-cell data. We plan to address this issue by using the cisTopic package in Python.
In this project, we seek to implement a Multi-Layer Perceptron (MLP) pipeline in which the input is the peak vector of a cell, and the output is the likelihood of the cell being from a young or aged mouse brain (binary classification task). Due to the size of the data, the complexity, and the sparsity, Deep Learning is an ideal approach. We will split our data into training and testing groups via cross-validation. We define model success in terms of accuracy- as in, the % of times our model correctly classified a cell as young or aged.
A base goal for our project within the time frame we have is to have our model perform better than random classification (correct more than half the time). A target goal is to reach 70% accuracy. A stretch goal is to reach an 80% accuracy and perform hyperparameter analysis to gain biological insight based on what the model uses for classification and how hyperparameters differ in males and females.
This project will hopefully contribute to the larger topic of sex-specific differences in the aging brain, which is crucial to understand. Females have been historically underrepresented in biomedical studies; the practice of conducting studies on only male subjects and generalizing results to all sexes has major repercussions and we must better address sex differences in all aspects of biomedicine.
Division of labor: Ananya is the point-of-contact with the Webb Lab and has the domain knowledge for the project. Tuan has expertise with preprocessing single-cell data. Jessy has the applied math background necessary for both architecture design and analysis for biological insight. All 3 of us will contribute equally to implementing the model itself.
Link to project reflection for check-in 3: https://docs.google.com/document/d/1xp4IjNIjoCbWL7LdRF2vt0_ey0l0kgC_WSm2fH4JEpY/edit?usp=sharing
Link to DL Day slides: https://docs.google.com/presentation/d/1LMWbQpH8uGLDUj-UhKA22XM2gaPneUrZhoVXa_R3yGg/edit#slide=id.gebb00ae622_0_0
Link to write-up: https://docs.google.com/document/d/1t95Gym1xtmFnaQDjRp8sgHBSSxFkeJN2wI5TuAhNCZk/edit?usp=sharing
Log in or sign up for Devpost to join the conversation.