Public defence, Computer Science, MSc Muhammad Arslan Masood

Title of the thesis: Molecular representation learning for extreme low data in-vivo toxicity modeling

Thesis defender: Muhammad Arslan Masood 
Opponent: Professor Ola Spjuth, Uppsala University, Sweden
Custos: Professor Samuel Kaski, Aalto University School of Science

Bringing a new medicine to patients takes more than a decade of research, yet nine in ten drug candidates fail in clinical trials. About a third of these failures stem from unexpected toxicity: side effects that appear only in the first-in-human studies, unforeseen by earlier laboratory and animal tests. Every failure delays treatments patients are waiting for. 

The core problem is data. Measuring how a compound behaves inside a living body is slow and expensive, and still leans on animal experiments that raise ethical concerns and never perfectly mirror human biology. Reliable safety data is therefore scarce, a central obstacle for artificial intelligence in drug discovery. This doctoral thesis develops machine learning methods that predict toxicity accurately from only a handful of such measurements. 

The thesis contributes three methods. The first learns the "language" of chemistry from millions of molecules that need no experiments, then uses uncertainty estimates to select the most informative experiment to run next, so an accurate toxicity model can be trained from far fewer laboratory measurements. The second, VitroBERT, adapts a language model to learn both the chemistry and the biological interactions measured in laboratory experiments, improving prediction of side effects in both animals and humans. The third, BioXMol, combines molecular structure with two biological signals measured in the laboratory: changes in cellular shape and genomic activity in response to the molecules. This richer view sharpens predictions for chemically similar compounds whose toxicity in humans differs sharply, where standard chemical fingerprints perform poorly. 

Together, these methods make toxicity prediction more reliable exactly where data is scarcest. The work contributes to drug-safety screening that is faster, cheaper, and less dependent on animal testing, flagging unsafe candidates before they reach patients.

Key words: Representation learning, toxicity prediction

Thesis available for public display 7 days prior to the defence at Aalto University's public display page. 

Contact Information:
arslan.masood@aalto.fi 
https://www.linkedin.com/in/arslan-masood-26469890/