Graph Neural Networks (GNNs) are powerful tools for graph-related tasks, excelling in progressing graphstructured data while maintaining permutation invariance. However, their challenge lies in the obscurity of new node representations, hindering interpretability. This paper introduces a framework addressing this limitation by explaining GNN predictions. The proposed method takes any GNN prediction, for which it returns a concise subgraph as explanation. Utilizing Saliency Maps, an attribution gradient-based technique, we enhance interpretability by assigning importance scores to entities withing the knowledge graph via backpropagation. Evaluated on the Drug Repurposing Knowledge Graph, Graph Attention notable results with the highest recall r0a.9t9e2.oOfur framework underscores GNN eficacy and interpretability, which is crucial in complex scenarios like drug repurposing. Illustrated through an Alzheimer's disease case study, our approach provides meaningful and comprehensible explanations for GNN predictions. This work contributes to advancing the transparency and utility of GNNs in real-world applications.
In recent years, significant attention has been directed towards the development and utilization
of extensive Knowledge Graphs (KGs1)][for computational drug discovery tas2k,s3,[
Drug repurposing using KGs is often modeled as link prediction problem, where the objective is to predict the connection between a candidate drug and a target disease. Several methods, including embedding-, rule- and network-based approaches, have been employed to address this problem. While Graph Neural Networks (GNNs) have gained prominence due to their capabilities of complex reasoning over expansive KGs, their application introduces interpretability challenges. Specifically, the inherent opacity of GNNs, stemming from their aggregation SWAT4HCLS’24, Feb 26 – 29, Leiden, the Netherlands nEvelop-O strategy, hinders the traceability of computing new node representations, posing dificulties in interpreting the final predictions.
To address this interpretability challenge, Char[i5e]tinatlr.oduced an explanation ontology, categorizing various literature-derived explanation types. This paper proposes a framework generating explanations falling within the ”statistical” and ”scientific” explanation categories. Our approach draws inspiration from statistical explanation principles to enhance the interpretability of GNN predictions. Moreover, the explanations given for the case study on Alzheimer’s disease are supported by existing literature. Through empirical evidence obtained from our experiments, we aim to provide clear insights into the predictions generated by GNNs.
Our framework takes a prediction from a trained GNN as input, for which it produces an explanation in the form of a concise subgraph. By employing Saliency Maps, an attribution gradient-based method, our method assigns importance scores to nodes via backpropagation. The method generates an explanatory subgraph by summing node scores to determine the rank or importance of specific triples.
The remainder of the paper is structured as follows: Se2ctdieolnves into previous related works. Section3 describes the Drug Repurposing Knowledge Graph (DRK6G])w[ hile Section4 introduces our proposed XAI framework. Sect5iodnetails the experimental setup and reports results, followed by a discussion in Sect6io. nFinally, Section7 concludes this paper. The implementation for analysis is available in our Github repohstittoprs:y//github.com/jloe2911/ ExplainableDrugRepurposin.g
While GNNs have gained traction for their capacity to perform complex reasoning over large
KGs, their lack of interpretability poses a challenge for user trust. Jian[n7a]nadedtraels.sed
this limitation with their interpretable Deep learning-based Path-reasoning framework for drug
repurposing (iDPath), which captures the mechanisms of drug actions (MODA) by identifying the
paths from drugs to diseases in the human body. iDPath utilizes a Graph Convolutional Network
(GCN) [
Similarly, Islam et a[1l0.] introduced a workflow that leverages a set of explanation rules
and extracts explanatory paths instantiating the rules f(oCromgipvoeunnd, Disease) pairs of
interest based on the DRKG6].[ Their approach generates compact ensemble embeddings of the
entities and relations using Tran1s1E],[TransH [12], and DistMult1[
Gottlieb et a[l1.5] presented the PREDICT method, grounded in the ”Guilt by Association” (GBA) approach. PREDICT predicts novel associations between desired drugs and diseases, involving the measurements of similarities among the known drug and disease to drug-disease pairs, given a target query drug and disease. However, lacking reasoning behind the final predictions promoted Ozkan et [a1l6.] to propose XPREDICT, a KG-based explainable AI approach for drug repurposing. XPREDICT focuses on generating similarity-based explanations, calculating cosine similarity between drug-drug and disease-disease pairs. While efective in generating plausible explanations on a case study on Alzheimer’s disease through the OpenPred1i7c]t, API [ a limitation is the exclusion of relevant entities like genes and pathways in the generated explanation. To address this, we expanded our XAI framework to include gene and pathway nodes.
The input to our framework is the DRK6G], [initially developed for the COVID-19 drug repurposing project. DRKG is a comprehensive KG that interconnects various biological entities, including compounds, diseases, genes, and side efects. It collects approxima9t7eltyhousand entities and ove5rmillion edges. Curated from six diferent data sources, namely STRING, DrugBank, Hetionet, Global Network of Biomedical Relationships (GNBR), IntAct and Drug Gene Interaction database (DGIdb), DRKG compri1s0e7s diferent edge-types. Given the goal of drug repurposing to discover new uses for existing drugs, we approach the challenge as a link prediction task, aiming to predict missing edges between Compounds and Diseases: ⟶ . Figure1 illustrates the entire pipeline, encompassing two major steps.
This section outlines the proposed framework for explaining predictions generated by GNNs. The initial phase of the pipeline focuses on training a GNN-based link prediction model. The fundamental concept behind a GNN layer is to compress a set of vectors into a single one. This involves a two-step process: (i) message computation and (ii) aggregation. During the first step, each node calculates a message, represented as a node embedding. Mathematically, for each node ∈ { () ∪ }, the message at layeirs computed as () = () (ℎ(−1) ). Moving to the second step, each node aggregates the messages from its neighbours. Foℎr()miaslly, determined by () ({ () , ∈ () } , () ). Consequently, th e-th GNN layer takes node embeddings from the node itseℎl(f,−1) , and those from neighbouring nodℎe(s∈−, 1()) , producing the node embeddingℎ() . In this study, various GNN variants were employed, including Graph Convolutional Network (GCN8]),[GraphSAGE [18], and Graph Attention Network (GA1T9)].[ The key distinction lies in how these variants aggregate information across the diferent layers.
In the second phase of the pipeline, the process involves utilizing a trained GNN-based link prediction model and a predictiô nto generate an explanat i on⊆ in the form of a compact subgraph derived from the input graph. To achieve this, we employed Saliency M20a],pasn[ attribution gradient-based method proficient in assigning importance scores to entities within the KG via the process of backpropagation.
By highlighting or masking specific nodes to reveal their contributions to the final prediction, we can easily extract a concise subgraph comprising the most important nodes that significantly influenced the prediction(s). Specifically, we applied Saliency Maps for each predicti ô,n∀ ∈ to obtain importance scores for all nodes influential in determining the outcome of that particular prediction ̂ . Following the computation of gradients of the output concerning the input (i.e., node embedding), the attribution value using Saliency Maps alo nℎgdtihmeension for an input ∈ ℝ is defined as the absolute value of the gradieSnatli:ency () = | () | where () denotes the output of a GNN model for in p.ut
Utilizing the scores assigned by Saliency Maps to each node in the network, we can construct an explanatory subgraph employing Algori1t.hTmhe algorithm generates explanatory subgraphs based on node attribution scores computed by Saliency Maps (SM). It starts by ranking Gene entities based on SM scores. For each Gene, various relations, including participation in pathways, association with diseases and binding with compounds are extracted. These relations are ranked based on the cumulative scores assigned to their respective nodes. In the end, top triples are included in the final explanatory subgraph.
Hence, the proposed methodology explains GNN predictions by encompassing two primary phases. The initial phase involves training a GNN-based link prediction model. Variants like GCN, GraphSAGE, and GAT are employed to compute node embeddings. The second phase leverages the trained GNN-based link prediction model to generate explanations for GNN predictions employing Saliency Maps. Saliency Maps highlight the most influential nodes contributing to the final predictions. The algorithm defined in the framework uses these scores Algorithm 1 Algorithm used to generate explanations. to rank paths involving genes, pathways, diseases, and compounds, which are then used to construct an explanatory subgraph.
To evaluate the proposed framework, we employed GC8]N,G[raphSAGE [18], and GAT [19] on the DRKG 6[]. The assessment of GNNs was based on Precision, Recall, and Hits@k metrics.
In order to train the GNNs, we implemented a train-test s8p0l%itfoorf training samples and20% for testing samples. The node embeddings were set with a dimension si1z0e0otfo represent the features and relationships within the graph structures. For optimization, the Adam optimizer was selected to eficiently update model weights and enhance convergence during the training process. To iteratively refine the model’s parameters, we con5d00ucted epochs, enabling the network to learn intricate patterns and relationships within the data.
Precision measures the ratio of true positives to predicted positives, ofering insights into the cost of false positives. Conversely, recall calculates the ratio of true positives to actual positives, indicating the cost of false negatives. Hits@k measures how frequently the correct head or tail is among the toppredictions. Table1 highlights that GAT outperformed, achieving a Hits@5 score of0.451 and a Hits@10 score o0f.672. The trade-of between precision and recall is evident, with a high emphasis on recall in drug repurposing scenarios to minimize falsely predictedtreat-links betweendrugs anddiseases. Notably, GraphSAGE achieved the highest recall rate 0o.f992.
We applied our explainability framework to interpret predictions generated by GNNs in a case study focused on Alzheimer’s disease. The analysis involved a trained GNN, namely GraphSAGE having the highest recall rat0e.99o2f, and a disease-filtered KG, where only Alzheimer’s disease-related information was retained. In particular, we excluded all triples (Compound, treats, Disease) where the disease difers from Alzheimer’s. Notably, our focus was solely on explaining correct predictions, disregarding classification errors made by the trained GNN.
To assess the eficacy of our approach, we generated explanatory subgraphs for specific instances, such as (Donepezil, treats, Alzheimer) and (Memantine, treats, Alzheimer2). Figure illustrates the explanatory subgraph for Donepezil. The literature reveals that the primary goal of Alzheimer’s drugs, including Donepezil, is to maintain elevated acetylcholine (ACh) levels, thereby compensating for the loss of functioning cholinergic bra2in1,c2e2l]l.sF[igure2 emphasized the crucial role of Donepezil binding to acetylcholinesterase (AChE) and butyrylcholinesterase (BChE). Both BChE and AChE are involved in ACh metabolism and thus are important for the cholinergic function in the b2r3a].in [
Similarly, we explored the explanatory subgraph for (Memantine, treats, Alzheimer) as depicted in Figure3. Memantine, commonly prescribed for moderate to severe Alzheimer’s disease, is believed to help prevent excess levels of the substance glutamate from damaging the brai1n. Thus, according to our explainability framework and existing literature, important genes associated with Memantine’s eficacy include glutamate receptor and acetylcholinesterase. This comprehensive interpretation provides valuable insights into the complex mechanisms underlying the treatment of Alzheimer’s disease with these medications.
For experts equipped with biological knowledge, evaluating AI-based model explanations, particularly those generated by GNNs, involves several key considerations. First, assessing the accuracy and consistency of predictions and explanations is crucial, comparing their performance to existing methods or baselines. Second, scrutinizing the biological validity and plausibility of the returned subgraphs becomes paramount, examining how well they align with known or novel associations between nodes and target predictions. The comprehensibility and simplicity of the explanations play a pivotal role, ofering insights into how efectively they aid users in grasping the reasoning and logic behind predictions, along with the underlying 1https://www.ncbi.nlm.nih.gov/books/NBK279356/ mechanisms and disease processes. Finally, evaluating the actionability and applicability of the explanations is essential, gauging their capacity to support users in making well-informed decisions, whether related to diagnosis, prognosis, treatment, or prevention of the disease.
While explanations from KGs currently have limitations, relying on simple associations and bindings between drugs, genes, and diseases, there is room for improvement. Existing frameworks often overlook crucial aspects such as causal mechanisms, dose-response relationships, and clinical outcomes. However, insights from biological explanations guide us towards refining our explanation.
Understanding drug interactions at the molecular level, like the binding of Donepezil to enzymes such as CYP2D6, opens avenues for more nuanced explanations. While this impacts pharmacokinetics and pharmacodynamics, it does not directly elucidate the therapeutic action on Alzheimer’s disease. Conversely, the binding of Donepezil to ACHE, an enzyme in acetylcholine metabolism, is a crucial mechanism explaining its positive impact on cognitive function in Alzheimer’s patients. Similarly, Memantine’s binding to CYP2B6, an enzyme in drug metabolism, raises considerations about pharmacokinetics and pharmacodynamics. Memantine’s binding to GRIN1, a subunit of the NMDA receptor, presents a critical mechanism. By blocking the NMDA receptor, Memantine protects neurons from excitotoxicity caused by glutamate overstimulation, explaining how it reduces cognitive decline and improves overall function in Alzheimer’s patients.
In essence, refining associations with a focus on relevant targets presents an exciting avenue for enhancing explanations related to the therapeutic efects of drugs like Donepezil and Memantine in Alzheimer’s treatment. This not only aligns with expert biological insights but also contributes to the wider acceptance and utility of explanations in the domain.
While the proposed framework demonstrates notable strengths in interpretable graph-based predictions, it is essential to recognize certain limitations. Firstly, the framework assumes a static and deterministic nature of input graphs, potentially hindering its applicability to dynamic, temporal, or uncertain graph characteristics. Secondly, the focus on local or subgraph-level explanations might limit the framework’s ability to capture global aspects of the graph, including its topology, properties, or embeddings. Furthermore, the framework’s single-subgraph approach may restrict its capacity to provide diverse or alternative explanations for the same prediction, highlighting considerations for its application in scenarios requiring comprehensive and varied explainability. Another limitation of this work is that there was no formal evaluation conducted for the explanations.
To propel the capabilities of our framework and address its identified limitations, future work could encompass diverse directions. One avenue involves extending the framework to accommodate various graph types, including directed, weighted, or attributed graphs, enabling a comprehensive evaluation of its performance and interpretability across diferent graph datasets. Additionally, investigating the integration of domain knowledge, such as biological pathways, gene ontology, or disease phenotypes, into the framework holds potential for enhancing biological interpretability and relevance in the explanations. Furthermore, conducting user studies with domain experts, such as biologists, clinicians, or patients, can provide valuable insights into the usefulness, trustworthiness, and user satisfaction of the framework. Such studies can shed light on potential challenges and limitations, guiding refinements and improvements for future iterations of the framework.
In this paper, we presented an XAI framework that combines GNNs for link prediction on the
DRKG [
NIPS’17, Curran Associates Inc., Red Hook, NY, USA, 2017, p. 1025–1035. [19] P. Veličković, G. Cucurull, A. Casanova, A. Romero, P. Liò, Y. Bengio, Graph attention networks, in: International Conference on Learning Representations, 201h8t.tUpRs:L: //openreview.net/forum?id=rJXMpikC.Z [20] S. Karen, V. Andrea, Z. Andrew, Deep inside convolutional networks: Visualising image classification models and saliency maps, 2013. URLh:ttps://arxiv.org/abs/1312.60.3d4oi:10. 48550/ARXIV.1312.6034. [21] M. Saeedi, F. Mehranfar, Challenges and approaches of drugs such as memantine, donepezil, rivastigmine, and aducanumab in the treatment, control and management of alzheimer’s disease, Recent patents on biotechnology 16 (2022) 102–121. [22] J. Knowles, Donepezil in alzheimer’s disease: An evidence-based review of its impact on clinical and economic outcomes, 2006. URhLt: tps://www.ncbi.nlm.nih.gov/pmc/articles/ PMC3321665/. [23] S. Sokolow, X. Li, L. Chen, K. D. Taylor, J. I. Rotter, R. A. Rissman, P. S. Aisen, L. G.
Apostolova, Deleterious efect of butyrylcholinesterase k-variant in donepezil treatment of mild cognitive impairment, 2017. URhL:ttps://www.ncbi.nlm.nih.gov/pmc/articles/ PMC5534361/#:~:text=Donepezil%20also%20inhibits%20butyrylcholinesterase%20(BChE, cholinergic%20function%20in%20the%20brain.