Tree-based models and cell fate choice
File(s)
Author(s)
Croydon Veleslavov, Ivan Alexander
Type
Thesis
Abstract
Lineage analysis is concerned with how pluripotent progenitors divide and differentiate to form the myriad of mature cell types we observe in complex organisms. During differentiation, these cells respond to complex spatiotemporal signals that impact cell fate choice. Small, quantitative changes in the expression of key genes can result in sister cells whose dynamics and cell fates are qualitatively distinct. With advances in single-cell transcriptomics, we can now observe in unprecedented detail how heterogeneity in gene expression between cells in a population can give rise to specialised cell populations during development.
Models that allow us to interrogate these transcriptomic differences and relate the outcomes of cell fate choice to genetic drivers carry clear appeal. Suitable models would allow us to predict cell classes from transcriptomic data alone, give probabilistic estimates for cell fates in branching trajectories and even predict how a system might behave under perturbation.
In this thesis we present a series of tree-based models trained on single-cell gene expression data to predict cell classes. Using Murine neural differentiation and Xenopus tropicalis embryogenesis as two diverse developmental settings, we show that these models perform well on withheld data and that their structure confers opportunities for knowledge discovery. Focusing on their capability for embedded feature selection, we demonstrate that these tree-based methods robustly identify features involved in regulating the cell fate boundaries of interest.
The analysis provided in this thesis considers the tension between model complexity and predictive performance. We motivate these variously complex tree-based models, working up from shallow decision stumps through to fully Bayesian ensembles. In the final case, we propose a means of incorporating prior information into Bayesian additive regression tree models through a sparsity-inducing, asymmetric Dirichlet hyperprior. We show that appropriate prior information can both improve model performance and reduce the complexity of the learnt ensemble.
Models that allow us to interrogate these transcriptomic differences and relate the outcomes of cell fate choice to genetic drivers carry clear appeal. Suitable models would allow us to predict cell classes from transcriptomic data alone, give probabilistic estimates for cell fates in branching trajectories and even predict how a system might behave under perturbation.
In this thesis we present a series of tree-based models trained on single-cell gene expression data to predict cell classes. Using Murine neural differentiation and Xenopus tropicalis embryogenesis as two diverse developmental settings, we show that these models perform well on withheld data and that their structure confers opportunities for knowledge discovery. Focusing on their capability for embedded feature selection, we demonstrate that these tree-based methods robustly identify features involved in regulating the cell fate boundaries of interest.
The analysis provided in this thesis considers the tension between model complexity and predictive performance. We motivate these variously complex tree-based models, working up from shallow decision stumps through to fully Bayesian ensembles. In the final case, we propose a means of incorporating prior information into Bayesian additive regression tree models through a sparsity-inducing, asymmetric Dirichlet hyperprior. We show that appropriate prior information can both improve model performance and reduce the complexity of the learnt ensemble.
Version
Open Access
Date Issued
2021-09
Date Awarded
2022-06
Copyright Statement
Creative Commons Attribution NonCommercial Licence
License URL
Advisor
Stumpf, Michael
Sponsor
Wellcome Trust (London, England)
Grant Number
P66556
Publisher Department
Life Sciences
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)