Bayesian Neural Networks for Molecular Discovery

Improving Uncertainty Quantification in Graph Neural Networks

Supervisor: Assoc. Prof. Mikkel N. Schmidt

Graph neural networks (GNNs) have revolutionized the analysis of molecular data and prediction of molecular properties, showcasing unparalleled predictive capabilities. However, one of the critical challenges with neural networks lies in their inability to provide reliable uncertainty estimates. This limitation becomes a significant hurdle in the domain of drug and material development, where uncertainty quantification is not just a measure of model confidence but a pivotal tool for guiding exploratory searches in vast chemical spaces.

Prior Research

Generative Models for Protein Sequence and Structure

Supervisor: Prof. Wouter Boomsma

In my research, I investigated the use of VAEs to capture the effects of mutations in protein sequences, inspired by the DeepSequence model, and also studied how the U-Net architecture can learn amino acid propensities from voxel representations of protein structures and “paint in” missing parts of the protein structure.

Next-Generation Vaccine Development Against SARS-CoV-2 (2020)

Contributed to developing a novel SARS-CoV-2 vaccine, leading to a publication in Nature Communications. The project’s success in clinical trials highlights the platform’s potential for rapid response to infectious diseases.

Publication: Capsid-like particles decorated with the SARS-CoV-2 receptor-binding domain elicit strong virus neutralization activity

Enhancing Immunogenicity Through cVLP Vaccine Platforms (2018-2020)

Supervisor: Prof. Adam F. Sander

My master’s thesis explored the impact of cVLP backbones and conjugation systems on immune responses, demonstrating significant findings for vaccine design, published in Vaccines.

Publication: Head-to-Head Comparison of Modular Vaccines Developed Using Different Capsid Virus-Like Particle Backbones and Antigen Conjugation Systems