Abstract
Abstract
Graph Neural Networks (GNNs) have gained significant attention in drug discovery
due to their ability to model graph-structured biological and chemical data, such as
molecular graphs, drug–drug interaction networks, and drug–target interaction networks.
Despite their success, practical adoption of GNNs is often limited by implementation
complexity, lack of standardized benchmarking, and reduced generalization
under distribution shifts across datasets and experimental conditions.
This thesis addresses these challenges through two complementary contributions.
First, a no-code framework (GNN-NCF) is introduced to support reproducible GNN
model construction, hyperparameter optimization, training, evaluation, and inference
across multiple drug discovery tasks. The framework enables systematic benchmarking
of popular GNN architectures, including ChebConv, GCNConv, GATConv, GINConv,
and SageConv, on representative datasets from the Open Graph Benchmark
and additional drug–target interaction settings. Benchmarking results demonstrate
that GNN performance is strongly task-dependent and sensitive to dataset properties
such as class imbalance, graph density, and chemical diversity.
Second, to improve robustness under domain shift in drug–target interaction prediction,
this thesis proposes GRADDA-DTI, a gradient-aware domain adaptation method
that combines multi-level embedding alignment (MMD and CORAL) with dynamic
loss weighting (UPGrad). Experiments show that GRADDA-DTI improves crossdomain
transfer performance on benchmark datasets. Overall, this work advances
accessible and robust GNN-based workflows for drug discovery, supporting more reliable
and reproducible computational screening.