Exploring the effects of a chemical compound on a species takes a considerable experimental effort. Appropriate methods for estimating and suggesting new effects can dramatically reduce the work needed to be done by a laboratory. Here, we explore the suitability of using a knowledge graph embedding approach for ecotoxicological effect prediction. A knowledge graph has been constructed from publicly available data sets, including a species taxonomy and chemical knowledge. These knowledge sources are integrated by ontology alignment techniques. Our experimental results show that the knowledge graph and its embeddings augment the baseline models.1 (i) We have created a KG by gathering and integrating the relevant data from disparate sources. In order to discover equivalent entities we exploit internal resources, external resources (e.g., Wikidata [16]) and ontology alignment (e.g., LogMap [6, 5]). (ii) We have evaluated three KG embedding approaches (TransE [2], DistMult [18] and HolE [12]) together with a baseline based on a one-hot encoding. Out evaluation shows improvement in the metrics using KG embedding for a majority of the selected classification models. Note that, recall is preferred over precision, i.e., rather overestimate the effect of a chemical compound, than underestimate its hazardousness.
It takes immense experimental efforts to determine ecotoxicological effects a chemical compound has on a species. These effect data is available for a narrow range of compound-species pairs and a limited number of experimental test.
Here, we present a preliminary study of the benefits of using Semantic Web tools to integrate different data sources and knowledge graph (KG) approaches to improve the ecotoxicological effect prediction over a baseline. Hence, our contribution is twofold:
Knowledge graphs. We follow the RDF-based notion of KGs [
Ontology alignment. Ontology alignment is the process of finding mappings or
correspondences between a source and a target ontology or knowledge graph [
We construct the Toxicology and Risk Assessment (TERA) KG from four sources:
(i) The effect data is gathered from ECOTOX [
We learn different types of classification models, including Gaussian naive-bayes (NB), quadratic discriminant analysis (QDA), radial basis function kernel support-vector machine (SVM), and multilayer perceptron (MLP), to solve the problem described in Figure 1. The input is a compound-species pair. It is encoded either as the the concatenation
CA subClassOf type CR
CB type type type c1 c2
Affects Not affects Affects Affects s1 s2 type
SA type subClassOf
SR subClassOf
SB type c3 s3 Fig. 1: The effect prediction problem. Lowercase sj and ci are instances of species and compounds, while uppercase denote classes in the hierarchy. Solid lines are observations and dashed lines are to be predicted. i.e., does c2 affect s1?
Knowledge Graph Embedding for Ecotoxicological Effect Prediction
(a) Accuracy
(b) Precision
(c) Recall
(d) AUC
of the one-hot vectors of the compound and the species (baseline), or the concatenation
of the embedding vectors learned by the embedding model (TransE [
Results. Figure 2 shows the results of different models using different encoding methods of the input (compound-species pair). We find that two out of the four testing models, namely SVM and MLP achieve higher performance with KG embedding than with one-hot encoding. For the QDA model, KG embedding also has higher recall than onehot encoding, although the overall metrics AUC and accuracy are similar. Note that recall is more important than precision in ecotoxicological effect prediction. The only exception is the NB model, where one-hot encoding has much higher performance than KB embedding. That is because NB holds the assumption that the input variables are conditional independent. Hence, it works better on the one-hot encoding which is quite sparse. However, it is worthwhile to note that the performance of NB with one-hot encoding does not outperform the MLP and QDA models with KB embedding. Conclusion. We have created a KG called TERA that aims at covering the knowledge and data relevant to the ecotoxicological domain. We have also implemented a proofof-concept prototype for ecotoxicological effect prediction based on knowledge graph embeddings and classification models. Some of the models used can take advantage of the learned embedded features. However, simple models like NB preferred the onehot encoded vectors. The obtained results are encouraging, showing the positive impact of using KG embedding models and the benefits of having an integrated view of the different knowledge and data sources.
Future work. The main goal in the long-term future is to make the TERA-KG accessible for domain researchers and improve the effect prediction by enriching the KG. In the near future, we intend to improve the current ecotoxicological effect prediction prototype and evaluate the suitability of more sophisticated models like Graph Convolutional Networks.
Resources. The datasets, evaluation results, documentation and source codes are available from the following GitHub repository: https://github.com/Erik-BM/NIVAUC Acknowledgements. This work is supported by the grant 272414 from the Research Council of Norway (RCN), the MixRisk project (RCN 268294), the AIDA project (The Turing Institute) and the SIRIUS Centre for Scalable Data Access (RCN 237889).