GraphPINE: Graph importance propagation Neural Network for interpretable drug response prediction
Keywords: Graph Neural Networks, Information Propagation, Drug Response Prediction, Explainability
Abstract: Explainability is necessary for tasks that require a clear reason for a given result such as finance or biomedical research. Recent explainability methodologies have focused on attention, gradient, and Shapley value methods. These do not handle data with strong associated prior knowledge and fail to constrain explainability results by relationships that may exist between predictive features.
We propose a GraphPINE, a novel graph neural network (GNN) architecture that leverages domain-specific prior knowledge for node importance score initialization. Use cases in biomedicine necessitate generating hypotheses related to specific nodes. Commonly, there is a manual post-prediction step examining literature (i.e., prior knowledge) to better understand features. While node importance can be obtained for gradient and attention-based methods after prediction, these node importances lack complementary prior knowledge; GraphPINE seeks to overcome this limitation. GraphPINE differs from other GNNs with gating methods that utilize an LSTM-like sequential format such that we introduce an importance propagation layer that unifies 1) updates for feature matrix and node importances, jointly and 2) uses GNN-based graph propagation of feature values. This initialization and updating mechanism allows for more informed feature learning and improved graph representation.
We apply GraphPINE to cancer drug response prediction using pharmacogenomics data (i.e., both drug screening and gene data collected by several assays) for ~5K gene nodes included in a gene-gene input graph with drug-target interaction (DTI) knowledge graph as initial importance. The gene-gene graph and DTIs were taken from literature curated prior knowledge sources and weighted by the literature information. GraphPINE demonstrates competitive performance and achieves a PR-AUC of 0.894 and ROC-AUC of 0.796 across 952 drugs. To highlight the interpretability aspect of our work, we provide the ability to generate sub-graphs of node importances. While our use case is related to biology, our work is generally applicable to tasks where information is separately known about feature relationships. Code: https://anonymous.4open.science/r/GraphPINE-40DE
Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 12503
Loading