Controllable Edge-Type-Specific Interpretation in Multi-Relational Graph Neural Networks for Drug Response Prediction

TL;DR

CETExplainer introduces a post-hoc, edge-type-controlled interpretability method for multi-relational GNNs applied to cancer drug response prediction. By constructing a directed heterogeneous graph and optimizing a joint objective that maximizes mutual information between predictions and subgraphs while enforcing a structure-score penalty, it yields fine-grained, biologically meaningful explanations that emphasize clinically relevant edge types. The authors also establish a ground-truth-based evaluation framework and demonstrate superior explanation quality and stability against strong baselines on real-world data, highlighting potential for more trustworthy CDR predictions. This work advances interpretable GNNs in biomedical contexts by enabling deliberate attention to edge types that encode important biological relationships, with practical implications for personalized oncology decisions.

Abstract

Graph Neural Networks have been widely applied in critical decision-making areas that demand interpretable predictions, leading to the flourishing development of interpretability algorithms. However, current graph interpretability algorithms tend to emphasize generality and often overlook biological significance, thereby limiting their applicability in predicting cancer drug responses. In this paper, we propose a novel post-hoc interpretability algorithm for cancer drug response prediction, CETExplainer, which incorporates a controllable edge-type-specific weighting mechanism. It considers the mutual information between subgraphs and predictions, proposing a structural scoring approach to provide fine-grained, biologically meaningful explanations for predictive models. We also introduce a method for constructing ground truth based on real-world datasets to quantitatively evaluate the proposed interpretability algorithm. Empirical analysis on the real-world dataset demonstrates that CETExplainer achieves superior stability and improves explanation quality compared to leading algorithms, thereby offering a robust and insightful tool for cancer drug prediction.

Figures (6)

• Figure 1: Process of Constructing a Directed Heterogeneous Network for Link Prediction. The multi-omics data of cell lines is input into a DNN layer to obtain cell line representations. Similarly, SMILES data is processed through a GNN layer to derive drug representations. Cell similarity triples and drug similarity triples are obtained using cosine similarity. These triples, together with resistance and sensitivity triples, form a directed heterogeneous network. This network is then input into an R-GCN layer to perform link prediction, enabling the identification of new links.
• Figure 2: Main Architecture of CETExplaine. $G_o$ represents the original directed heterogeneous graph, and the "new link" refers to the prediction result the link prediction task. $G_c$ is the combination of Go and "new link". $G_n$ is the subgraph form by $k$-hop neighborhood nodes and edges that extracted from $G_c$. $G_s$ is the explanation subgraph that derive from $G_n$. $Y_o$ and $Y_s$ are the prediction results obtained by input $G_o$ and $G_s$ to RGCN layers. The objective is to guide the explanation model to prioritize crucial edges within the interpretive process.
• Figure 3: Three situations for constructing GT. Sit. 1 demonstrates that if $i$ is sensitive to $J$, and $J$ is similar to $j$, then $i$ is sensitive to $j$. Sit. 2 shows that if $I$ is sensitive to $j$, and $I$ is similar to $i$, then $i$ is sensitive to $j$. Sit. 3 demonstrates that if $I$ is sensitive to $J$, $I$ is similar to $i$, and $J$ is similar to $j$, then $i$ is sensitive to $j$.
• Figure 4: (a) Results showing the quantitatively evaluate CETExplainer using three proposed metrics and compare it with the baseline models, ExplaiNE and GNNExplainer.(b) Res represents resistance, Sen represents sensitivity, Dsim represents drug similarity, and Csim represents cell similarity. CETExplainer improves the proportion of sensitivity, which is more important for explanations and predictions compared to the other two models, and maintains a more balanced distribution of various edge types. This results in an improved quality of the explanations.
• Figure 5: Qualitative evaluation of the three explanation models on triplet instances 1266 and 1989.