Improving QSAR Modeling for Predictive Toxicology using Publicly Aggregated Semantic Graph Data and Graph Neural Networks

Abstract

Quantitative Structure-Activity Relationship (QSAR) modeling is the most common computational technique for predicting chemical toxicity, but a lack of methodological innovations in QSAR have led to underwhelming performance. We show that contemporary QSAR modeling for predictive toxicology can be substantially improved by incorporating semantic graph data aggregated from open-access public databases, and analyzing those data in the context of graph neural networks (GNNs). Furthermore, we introspect the GNNs to demonstrate how they can lead to more interpretable applications of QSAR, and use ablation analysis to explore the contribution of different data elements to the final models' performance.

Knowledge-guided deep learning models of drug toxicity improve interpretation

“…We also acknowledge the recent development of toxicology-focused graph databases, such as ComptoxAI, which provide extensive knowledge on relations among chemicals, genes, assays, and many other entities. 42 Such a database may help researchers generate a more comprehensive feature profile for model training and thus improve the performance of DTox. In addition to better feature profiling, we think the incorporation of context-specific knowledge may further enhance the performance of DTox.…”

“…For each nuclear receptor of interest, we extract compounds with known connections (source compounds) from two resources: DrugBank 45 and ComptoxAI. 42 The performance metric was computed as the proportion of active compounds (query) that exhibit similar structure to at least one source compound. Five thresholds of Tanimoto coefficient were adopted to define structural similarity between source and query compound: 0.8, 0.85, 0.9, 0.95, and 1.…”

Knowledge-guided deep learning models of drug toxicity improve interpretation

“…Finally, despite the incorporation of pathway ontology, DTox VNN did not significantly outperform other well-established classification algorithms, as most differences are within the 95% confidence interval of performance metrics. We noticed the recent development of toxicology-focused graph database such as ComptoxAI, which provides extensive knowledge on relations among chemicals, genes, assays, as well as many other entities 40 . Such database may help us generate more comprehensive feature profile for model training, and thus improves the predicative performance of DTox.…”

Knowledge Graph Aids Comprehensive Explanation of Drug Toxicity

“…ComptoxAI 19, 20 is a comprehensive graph knowledge base that contains curated relationships between chemicals, genes, assays, and many other entities. It contains two types of relationships linking compound and gene nodes: physical binding (with protein product) and expression-alteration (up-regulation/down-regulation).…”