Pre-training and Sparse Log-Contrast Regression in Compositional Microbiome Studies
This project explores the potential of pre-training techniques to improve predictive modeling in microbiome research, with a focus on irritable bowel disease (IBD). By leveraging large-scale microbiome data, the study aims to enhance the accuracy of health outcome predictions, addressing the challenges of compositional data analysis. This approach could pave the way for more reliable insights into the gut microbiome’s role in complex diseases. You can read the full proposal here
.

Inferring interbacterial interactions in a defined gut bacterial community
The gut microbiome is extremely diverse and interactions among the individual bacterial community members are not limited to pairs, but can occur in groups of larger size where for example the presence and activity of one strain affects the way two others are interacting with each other. The purpose of this thesis is to computationally investigate the ecological interactions of a synthetic gut bacterial community with respect to metabolic outcomes with state-of-the-art statistical and machine learning approaches. You can read the full proposal here
.

Prediction of gene expression in response to chemical stress
The purpose of this M.Sc. thesis proposal is to evaluate the effectiveness of different representations of molecular structures at predicting the transcriptional response of human gut pathogens to different types of chemical stress. To achieve this task, we will rely on data generated as a part of the StressRegNet consortium, where the expression of specific genes in Salmonella enterica and Campylobacter jejuni have been measured in response to various chemical compounds.
The student will compare commonly used chemical representations such as the extended connectivity fingerprint (ECFP4), and chemical descriptors, to our previously described pre-trained chemical representation MolE. You can read the full proposal here
.

Interpretable predictions of bacterial stress response to chemical pertubations.
The purpose of this project is to develop a strategy for identifying chemical features that are predictive of a given outcome. Specifically, the student will work with the data generated by the StressRegNet consortium to identify the chemical features that best explain the observed changes in gene expression
You can read the full proposal here
.

Interpretable chemical representations via sparsified Graph Attention Networks
The main objective of this M.Sc. thesis proposal is to explore the use of the SparseMax function in the graph attention network (GAT) framework to generate an interpretable representation of molecular structures to predict a certain outcome. Benchmarks will include various regression and classification tasks from MoleculeNet. We will also apply the model to the task of predicting antimicrobial activity in the microbiome.
The student will determine if the use of SparseMax improves predictive performance over regular GATs. You can read the full proposal here
.

Inference for high-dimensional microbiome regression analysis
The primary objective of this thesis is to implement and benchmark the FDR-controlled variable selection procedure for sparse log-contrast models, which is a constrained sparse regression model. This procedure is designed to identify microbial predictors while controlling for false discoveries, thereby enhancing the reliability of the results. You can read the full proposal here
.
“Biomedical Statistics and Data Science”