Improving Regulatory Network Reconstruction Through Topological Priors, Robust Hyperparameter Exploration, and Multi-Task Learning

Abstract

Regulatory network reconstruction is an ongoing field of research that biologists have been pressing with considerable effort. Although several computational methods have been investigated, inferred networks still severely lag behind actual biology. Nevertheless, the methods that have been developed thus far do not begin to scratch the surface of possible approaches. To this end, we examine three advanced methods for improving the accuracy of computationally learned networks.

The first contribution of this thesis is an integrative network inference framework for structural priors. We promote the presence of a network motif by maintaining vertex and edge relationships and biasing the likelihood of an edge to favor such motifs. We apply our method in a yeast dataset and observe a slight increase in accuracy.

In our second approach, we look at a sampling technique to estimate hyperparameters and network structure in a high-dimensional setting. We implement the method in an existing framework and find improved results in terms of network reconstruction. However, we also evaluate the convergence properties for automatically tuned hyperparameters and find inconclusive results.

Our third approach repurposes an existing multi-task learning framework to learn context-specific networks in single-cell RNA-Sequencing data. We look at multiple partitioning schemes to cluster cell populations and compare networks in terms of pairwise similarity.

For each approach, we demonstrate the accuracy of network reconstruction by evaluating against a series of gold standards derived from literature and experiments. Additionally, we report the benefits and drawbacks of each method to analyze its efficacy.