Exploring the e ects a chemical compound has on a species takes a considerable experimental e ort. Appropriate methods for estimating and suggesting new e ects can dramatically reduce the work needed to be done by a laboratory. In this PhD research we aim at exploring the suitability of using a knowledge graph embedding approach for ecotoxicological e ect prediction. A knowledge graph is being constructed from publicly available data sets, including a species taxonomy and chemical classi cation and similarity. We use ontology alignment techniques to integrate the e ect data into the knowledge graph. Our preliminary experimental results show that the knowledge graph based approach improves the selected baselines.
(ii) Using the knowledge graph together with machine learning techniques to predict e ects. The objectives of this task are twofold: (a) Limit the search space for the laboratory (binary prediction). (b) Predict e ects outright with a margin of error (regression). 2
In this section we introduce some preliminaries and give insights into the current state of the art e orts applying semantic web technologies within the eld of toxicology and risk assessment.
Use case. Ecotoxicology is a multidisciplinary eld that studies the ecological and toxicological e ects of chemical pollutants on populations, communities and ecosystems. Risk assessment is the result of the intrinsic hazards of a substance combined with an estimate of the environmental exposure (i.e., Hazard + Exposure = Risk).
Figure 1 shows a risk assessment pipeline. Exposure is data gathered from the environment, while e ects are hypothesis that are tested in a laboratory. These two data sources are used to calculate risk, which is used to nd (further) susceptible species and the mode of action (MoA) or type of impact a compound would have over those species. Results from the MoA analysis are used as new e ect hypothesis.
E ect prediction. Estimating the e ect a compound has on a species is a large
research eld within ecotoxicology. Currently, state-of-the-art solutions such as
Quantitative Structure-Activity Relationship (QSAR) models ( e.g., [
There is emerging work in improving the usability of ecotoxicological data
by mapping to knowledge graphs or ontologies, e.g., [
Currently, mapping ecotoxicological data to di erent sources are under
construction. The ECOTOX web search interface4 now contains mappings to a
external taxonomy source [
We are not aware of e orts toward mapping taxonomic classes, e.g., genus, family, etc. which can reveal inconsistencies in the datasets.
Embedding models. Knowledge graph embedding [
There is previous work investigating modelling of chemical e ects, e.g., [
The relevance of the research to be conducted during the PhD can be summarized as follows: (i) Manually integrating background knowledge into risk assessment systems is cumbersome since a common vocabulary does not exists. Our approach will reduce the time spent organizing data, and increase the number of case studies than can be conducted. A common vocabulary will enhance the interoperability between several risk assessment systems, increasing the con dence in the assessments. (ii) The e ect data used in risk assessment models is the result of time-consuming laboratory work. By using machine learning techniques with background knowledge, in the form of a knowledge graph, we aim at being able to limit the search space for new tests to be analysed in the laboratory. For example, we aim at recommending the top-ten compounds to test on a speci c species, rather than conducting experiments using thousands of possible compounds. (iii) Design and implementation of a fully- edged recommender system to predict the level of e ect on a species. For example, DEET (pesticide) has the potential to kill 50% of the population of the common house y. Such generalization using the available data and knowledge is the main target of the research, which aims at reducing to a minimum further laboratory analysis. 4
This work aims to address the following questions: a. Can the disparate data sources used in ecotoxicological risk assessment be integrated into a knowledge graph to improve accessibility? b. Can the knowledge graph be used to improve (or diversify) ecotoxicological e ect prediction over current state-of-the-art models? The hypothesis associated with the above questions are: A. It is possible to integrate disparate data sources in a toxicological knowledge graph using Semantic Web tools.
B. Extrapolation of e ect data increase the reach of risk assessment systems while remaining accurate. 5
This section will describe the approaches used to investigate the hypothesis above. The evaluation of the hypothesis is described in Section 7. Hypothesis A. There are multiple sources, varying from tabular, SPARQL endpoints, REST APIs, and RDF formats, each with its own vocabulary, that needs to be integrated to enable a uni ed data access. The main sources of data are:
T ransform
T ransform N CBI Split
TERA-KG
SP ARQL Import
(i) E ect data (ECOTOX [
Includes limited metadata linked to proprietary identi ers for compounds
and species.
(ii) Compound data from di erent sources. Hierarchies available through
downloadable RDF les and SPARQL endpoints (PubChem [
Improving the knowledge graph can be done with several sources. First, a
dataset containing biological activity, e.g., Chemical ontology (CO) [
Another aspect important to e ect prediction is the habitat of the species,
e.g., [
Hypothesis B. The prediction task at hand is depicted in Figure 3. Initially,
we use a naive approach, which is to assume that similar compounds has a
comparable e ect on the same species and vice versa. The state of the art in
risk assessment systems implement akin solutions. However, it is not clear what
constitutes similarity in this context. The similarity between compounds are
quanti able using di erent methods, however, similarity does not imply
similar biological activity [
CA CB type type type type c1 c2 c3
Affects Not affects
Affects Affects s1 s2 s3 type type type
SA SB subClassOf subClassOf
SR We have evaluated three (plus variants) prediction models based on the e ect data and the TERA knowledge graph. Note that currently the TERA knowledge graph has been created with the bare minimum of sources required to integrate the e ect data with external metadata for compounds and species. Selected results are shown in Table 3. Prediction models used in this preliminary evaluation: (i) A nearest-neighbour approach. A compound-species pair can inherent an e ect if another compound or species is close in the knowledge graph. This method provides a useful baseline. However, the performance of this method is far from ideal, as it either will over or underestimate e ects based on the number of neighbours considered. (ii) A zero-background-knowledge multi-layer perceptron (MLP) model was applied to the e ects data. This model is able to learn simple relations, e.g., (a) Accuracy for the MLP prediction models.
(b) Recall for the MLP prediction models.
s1 and s2 is e ected by c3, therefore, c3 is toxic and will e ect s3. However,
when this model is presented with previously unseen compound-species
pairs, it cannot rely on background knowledge, and hence, the prediction
will be highly awed.
(iii) Using knowledge graph embedding ([
The obtained predictions are promising and show the potential usefulness of the machine learning models in our setting and the bene ts of using the TERA knowledge graph. As mentioned before, we favour recall with respect to precision. One the one hand, false positives are not necessarily harmful, while overlooking the hazard of a chemical may have important consequences. On the other hand, due to the limited experiments in terms of concentration (i.e., e ect data may not be complete), some chemicals may look less toxic than others while they may still be hazardous. At the same time the adoption of a RDF-based knowledge graph enables the use of an extensive range of Semantic Web infrastructure that is currently available (e.g., reasoning engines, ontology alignment systems, SPARQL query engines). 7
In this section, we introduce the evaluation plan for the success of this project. We can divide the evaluation of both research questions into qualitative and quantitative measures.
The value of the knowledge graph in toxicology research is uncertain at this stage. The knowledge graph must provide value for the researchers. We can ensure this by evaluating the quality of the knowledge graph. Our de nition of quality is that the knowledge graph should have high levels of: (i) Coverage. The sources included in the knowledge graph must cover the areas of interest. The coverage also relates to the degree of successful mappings between the sources. There will be a trade-o between completeness and correctness of the mappings. (ii) Integration. The ease of integrating the various sources. This involves aligning and mapping to attain a consolidated knowledge graph. (iii) Functionality. The ability of the knowledge graph to be integrated into the risk assessment systems. This includes keeping the exibility of Semantic Web technology without a commitment to a schema. We can add new triples and extend the knowledge graph, without the need of major changes. (iv) Embedding enrichment. The semantic enrichment the knowledge graph gives the embeddings compared to embeddings learned from e ects.
The quantitative evaluation of the knowledge graph generation is tightly related to that of evaluating the e ect prediction. We can evaluate the ability to make good e ect predictions quite easily. This can be done with precision, recall, accuracy, etc. for binary e ects or with mean squared error, R2-score, etc. for regression. However, the value of these predictions is integrating them into the risk assessment pipeline. The evaluation metrics must be inline with the ability our methods have to enhance risk assessment. We will compare environmental case studies results before and after the use of our modelling results. Since there is no ground truth data for risk assessment we rely on domain experts to determine if our contributions adds value to the assessments.
Risk assessments has currently large margins of errors (experimental errors etc.), and we may introduce new sources of error with our e ect predictions. However, we are con dent that errors can also be reduced by greater data coverage. These are di erent types of errors and part of the evaluation process will be to nd the optimal trade-o between them.
The current preliminary results uses random dataset splits for training and testing the models. We aim at introducing highly selective datasets that can test predictive performance in di erent scenarios. We will also try a completely clean test, where we recommend compound-species pairs to be tested in the lab. This will obviously be limited by the available compounds and test species of the particular laboratory.
The methodologies and knowledge graphs will be publicly available such that feedback from the community can help us evaluate and improve our contributions. 8
The conducted work falls into one of the main research lines of toxicology
research to enhance the generation of hypothesis to be tested in the laboratory [
Knowledge graph embedding models have been applied in general purpose
link discovery and knowledge graph completion tasks [
This PhD project is supported by grant 272414 from the Research Council of Norway. I would like to thank my supervisors, Ernesto Jimenez-Ruiz (The Alan Turing Institute and University of Oslo), Raoul Wolf (Norwegian Institute for Water Research), and Knut Erik Tollefsen (Norwegian Institute for Water Research) for their feedback on this work. In addition, I would also like to thank Jiaoyan Chen (University of Oxford), Martin Giese (University of Oslo) and Zo a C. Rudjord (Norwegian Institute for Water Research) for their contribution in di erent stages of this PhD research.