Integrating Knowledge Graphs with Graph Neural Networks for Explainable AML Drug Response Prediction : A Digital Twin-Inspired Framework for Patient Specific Mechanistic Therapy Modeling

Abstract

There has been recent use of Artificial Intelligence in drug response prediction, but many of those models are black boxes. It is crucial to provide justifiable reasons for the results produced by an AI model, as such explainability can lead to greater trust from doctors and patients. Such results can be used as inputs for trustworthy personalized medical practices. This research proposes an explainable Graph Neural Network model, combined with a literature-derived biomedical knowledge graph and the Beat AML cohort, for predicting drug response in patients with Acute Myeloid Leukemia (AML), a rapidly progressing and heterogeneous hematological malignancy. Here, as a knowledge graph, we use an extracted subgraph of AML from the PubMed Knowledge Graph (PKG) 2.0, which connects tens of millions of papers from the biomedical literature. Constraining the model with well-established biological relationships aims to mitigate the risk of overfitting inherent in the limited patient samples of the Beat AML dataset, which has high-dimensional transcriptomic data. The ultimate goal is to build a “Why” explanation layer and create literature-grounded explanation subgraphs that provide transparent biological evidence for the predictions generated by the AI model. These explanations are further compared using previously reported AML signaling mechanisms and therapeutic response patterns from the literature. The proposed framework utilized a Graph Attention Network v2 (GATv2) architecture together with graph attention analysis and perturbation-based explainability methods to identify biologically relevant signaling regions contributing to therapeutic response prediction. The model achieved an AUC score of 0.7038 while generating patient-specific mechanistic explanations for AML drug response prediction.