Toxicology knowledge graph for structural birth defects

Abstract

Background: Birth defects are functional and structural abnormalities that impact about 1 in 33 births in the United States. They have been attributed to genetic and other factors such as drugs, cosmetics, food, and environmental pollutants during pregnancy, but for most birth defects there are no known causes.

Methods: To further characterize associations between small molecule compounds and their potential to induce specific birth abnormalities, we gathered knowledge from multiple sources to construct a reproductive toxicity Knowledge Graph (ReproTox-KG) with a focus on associations between birth defects, drugs, and genes. Specifically, we gathered data from drug/birth-defect associations from co-mentions in published abstracts, gene/birth-defect associations from genetic studies, drug- and preclinical-compound-induced gene expression changes in cell lines, known drug targets, genetic burden scores for human genes, and placental crossing scores for small molecules.

Results: Using ReproTox-KG and semi-supervised learning (SSL), we scored >30,000 preclinical small molecules for their potential to cross the placenta and induce birth defects, and identified >500 birth-defect/gene/drug cliques that can be used to explain molecular mechanisms for drug-induced birth defects. The ReproTox-KG can be accessed via a web-based user interface available at https://maayanlab.cloud/reprotox-kg . This site enables users to explore the associations between birth defects, approved and preclinical drugs, and all human genes.

Conclusions: ReproTox-KG provides a resource for exploring knowledge about the molecular mechanisms of birth defects with the potential of predicting the likelihood of genes and preclinical small molecules to induce birth defects.

Fig. 1. Overview of the ReproTox-KG sources and connections.

The ReproTox-KG is made of lists of birth defects extracted from HPO and the CDC and birth-defect gene associations from HPO and Geneshot; HPO is a resource that provides an ontology of human phenotype and the human genes that have evidence to be associated with such phenotype; The CDC website has a dedicated site for listing major birth-defect terms. Using Geneshot birth-defect terms were connected to genes based on co-mentions in abstracts. The ReproTox-KG also has drug/birth-defect associations from DrugCentral, DrugShot, and other sources; To associate birth-defect terms with drugs, DrugShot was used to query birth-defect terms and drug-birth-defect association were determined based on co-mentions in abstracts. In addition, drug–gene associations were taken from the LINCS L1000 data and from drug-target knowledge. The LINCS L1000 data contain drug perturbation followed by expression for ~30,000 drugs and small molecules applied to ten human core cell lines at different concentrations and where gene expression was measured at different time points. Gene–gene associations are based on co-expression from ARCHS4; ARCHS4 contains uniformly aligned RNA-seq data from GEO and the gene–gene co-expression correlations were computed by randomly selecting thousands of RNA-seq sample and computing correlation with the Pearson’s correlation coefficient formula. Drug–drug associations within the knowledge graph are based on structural chemical similarity using RDKit, a software library that contains functions to compute the similarity between compounds based on different representations and algorithms.

Fig. 2. Overlap of drugs across categories.

Supervenn diagram of drug identifier overlap between FDA categories D and X, known placenta crossing drugs, and unique drugs and small molecules within the L1000 LINCS perturbation datasets. Drugs and compounds not represented in the L1000 perturbations are not included in the counts.

Fig. 3. Global visualization of gene expression signature similarity for LINCS drugs.

UMAP of 718,055 L1000 perturbations, colored by a FDA D and X category; b known placental crossing; c top MOAs across clusters. Clusters computed using HDBSCAN with a minimum cluster size of 40, top 25 clusters for each category, and top five MOAs of those clusters are included.

Fig. 4. Drug category and placental crossing prediction performance.

Bridge plots colored by prediction method for (a) predicting FDA D and X categories; and (b) placenta crossing. The NES are shown in the legend. Leading edges of the same bridge plots are shown on the right of each complete bridge plot.

Fig. 5. Screenshot from the ReproTox-KG user interface.

A query to identify connections between the birth defect Spina Bifida and the drug valproic acid with a limit of 25 nodes is provided as an example.

Fig. 6. Network of cliques connecting drugs, birth defects and genes.

Cliques are made of drugs that have a placenta crossing predicted rank of less than 3000 and are known to induce a birth defect based on literature evidence. These drugs are connected to the genes that their expression is affected by the drugs based on LINCS L1000 data. Finally, associations between genes and birth defects are established based on known mutations that are known to cause the same birth defect. Light blue nodes represent birth-defect terms, orange nodes represent genes, and pink nodes represent drugs and preclinical small molecules. Red lines with diamond-heads indicate an L1000 consensus drug signatures that upregulates the gene, and plungers indicate an L1000 consensus drug signature that downregulates the target gene. Gray arrowheads indicate genes that their mutations induce a birth defect, and gray diamond-heads connect drugs to the birth defects they are known to induce.