Title:	Pharmacogenomic variant interpretation using large-scale experimental datasets
DNr:	Berzelius-2026-238
Project Type:	LiU Berzelius
Principal Investigator:	Yitian Zhou <yitian.zhou@liu.se>
Affiliation:	Linköpings universitet
Duration:	2026-08-27 – 2027-03-01
Classification:	30102
Keywords:

Abstract

Inter-individual variability in drug response remains a major challenge in clinical medicine, with genetic variation estimated to account for approximately 20-30% of variability in treatment efficacy and safety. Variants in drug-metabolizing enzymes, transporters, receptors, and other pharmacologically relevant proteins can substantially alter drug disposition and response. Although pharmacogenomics has established numerous clinically actionable gene-drug associations, current approaches typically assign a single functional score or phenotype to a genetic variant. This representation is fundamentally limited because the functional effect of a variant can depend strongly on the chemical substrate and, in many cases, on the presence of additional variants within the same protein. A central objective of this project is to develop a computational framework capable of predicting substrate-specific and combination-dependent effects of genetic variation. A single variant may markedly impair the metabolism or transport of one drug while having little effect on another substrate of the same protein. Furthermore, combinations of variants may produce non-additive effects through structural interactions, epistasis, or compensatory mechanisms. Systematically characterizing these effects experimentally for the enormous number of possible variant-substrate and variant-combination combinations is infeasible, motivating the development of computational models that can learn these complex relationships from large-scale functional data. The project will combine multiplexed functional genomics with artificial intelligence to generate and analyze large, multidimensional datasets. Deep mutational scanning and related high-throughput assays will provide functional measurements for thousands of genetic variants across chemically diverse substrates. These experimental data will be integrated with protein sequence, three-dimensional structural information, and molecular representations of substrates to train deep learning models. Protein language models, graph-based neural networks, and molecular representation learning will be explored to capture sequence-structure-chemical relationships and their effects on protein function. Importantly, the models will be designed to predict not only the effects of individual variants but also higher-order variant combinations, enabling computational exploration of sequence space far beyond experimentally accessible combinations. The computational workload is substantial and requires high-performance computing resources. Model development will involve repeated training of large neural networks, hyperparameter optimization, cross-validation, ensemble modelling, and systematic comparison of alternative sequence, structural, and molecular representations. Large-scale inference will subsequently be performed across millions of potential variant–substrate combinations and multi-variant genotypes, creating a computationally intensive prediction landscape that cannot be efficiently explored using conventional desktop or small-scale computing resources. HPC resources will therefore be essential for both model training and large-scale generation of predicted functional landscapes. The resulting framework will establish a scalable computational approach for moving from generalized variant annotation toward drug- and genotype-specific functional prediction. By integrating experimental functional genomics with AI and high-performance computing, the project will enable systematic interpretation of rare variants, substrate-dependent effects, and variant combinations, providing a foundation for next-generation pharmacogenomics and AI-enabled precision pharmacology.