ºÚÁÏÍø

Title of the thesis: Deep learning methods for molecular biology: peptide and protein analysis and drug-like molecule generation

Thesis defender: Alexandru Dumitrescu 
Opponent: Professor Victor Greiff, University of Oslo, Norway
Custos: Associate Professor Harri Lähdesmäki, Aalto University School of Science

This thesis develops computational methods that can help scientists understand biological processes involved in disease and speed up the development of new treatments. Computational models can learn from large datasets and analyse and explore the molecular machinery of cells at a much greater scale than laboratory experiments alone. 

Biological data is often affected by noise arising from imperfect instruments and experimental bias, and can also be incomplete or heavily skewed towards current scientific interest. For example, T cell data largely comprises interactions with antigens from only a small number of distinct pathogens. Addressing these challenges often requires extensive adaptation of computational methods. 

The thesis develops four deep learning methods for medically relevant problems. The first method aids drug discovery by generating novel drug-like molecules as 3D objects. It represents molecules as fields, demonstrating advantages over competing methods in capturing molecular conformation statistics that are important for drug safety and efficacy. 

The thesis then presents three methods for peptide and protein analysis. The first proposes a novel solution for signal peptide analysis. Signal peptides form on newly synthesised proteins and direct their transport across cellular membranes. The second develops a method for analysing how chemical modifications of amino acids affect peptide interactions with major histocompatibility complex (MHC) proteins, an important disease signalling mechanism in the immune system. It further shows how one such modification in peptides from patients with rheumatoid arthritis alters peptide–MHC interactions, demonstrating its potentially disruptive effects on immune pathways in autoimmunity. Lastly, the thesis explores the potential of computational tools to support prediction of T cell specificity. T cells are white blood cells that survey jawed vertebrates for diseased cells. 

The thesis concludes that successful deep learning applications to biomedical problems often require adaptation to the underlying characteristics of the data. Such tailored approaches can make biological research more efficient and help translate growing amounts of molecular data into knowledge relevant to medicine.

Keywords: Deep learning, diffusion models, transformers, adaptive immunity, protein secretion, T cells, small molecules, TCR, peptide-MHC interaction, signal peptides, post-translational modifications

Thesis available for public display 7 days prior to the defence at .