Title:	Predicting chemotherapy sensitivity using graph neural networks based on deep mutational scanning
DNr:	Berzelius-2024-400
Project Type:	LiU Berzelius
Principal Investigator:	Ming Xiao <mingx@kth.se>
Affiliation:	Kungliga Tekniska högskolan
Duration:	2024-11-01 – 2025-05-01
Classification:	10203
Keywords:

Abstract

Understanding protein function is crucial for advancing our knowledge of biological systems and for developing targeted interventions. We previously designed a transformer-based variational auto-encoder for predicting the effect of genetic variants in coding regions. We tested our model on deep mutational scanning assays and established a new benchmark for 2 out of 26 pharmacogene-related proteins. Several aspects of the method remain to be explored. First, we did not manage to get good performances with rich multi-modal latent spaces despite extensively experimenting on mixture of Gaussian as well as Vamp priors. Second, we intended to interpret the modes of the Vamp prior by attempting to fold the input space prototypes to investigate their folding structure. Third, we wish to understand why our current approach does not benefit from the addition of the protein structure information. Fourth, we aim at reporting performances on novel drug transporter variant datasets designed in collaboration with the team at the pharmacology department. The characterization of variants effect on drug transporters may provide a powerful means of predicting patient response to therapeutic treatments, decreasing the burden of treatment course such as chemotherapy.