The Bigger Fish: A Comparison of Meta-Learning QSAR Models on Low-Resourced Aquatic Toxicity Regression Tasks

Toxicological information as needed for risk assessments of chemical compounds is often sparse. Unfortunately, gathering new toxicological information experimentally often involves animal testing. Simulated alternatives, e.g., quantitative structure-activity relationship (QSAR) models, are preferred to infer the toxicity of new compounds. Aquatic toxicity data collections consist of many related tasks─each predicting the toxicity of new compounds on a given species. Since many of these tasks are inherently low-resource, i.e., involve few associated compounds, this is challenging. Meta-learning is a subfield of artificial intelligence that can lead to more accurate models by enabling the utilization of information across tasks. In our work, we benchmark various state-of-the-art meta-learning techniques for building QSAR models, focusing on knowledge sharing between species. Specifically, we employ and compare transformational machine learning, model-agnostic meta-learning, fine-tuning, and multi-task models. Our experiments show that established knowledge-sharing techniques outperform single-task approaches. We recommend the use of multi-task random forest models for aquatic toxicity modeling, which matched or exceeded the performance of other approaches and robustly produced good results in the low-resource settings we studied. This model functions on a species level, predicting toxicity for multiple species across various phyla, with flexible exposure duration and on a large chemical applicability domain.

What is the ecotoxicity of a given chemical for a given aquatic species? Predicting interactions between species and chemicals using recommender system techniques

Ecotoxicological safety assessment of chemicals requires toxicity data on multiple species, despite the general desire of minimizing animal testing. Predictive models, specifically machine learning (ML) methods, are one of the tools capable of solving this apparent contradiction as they allow to generalize toxicity patterns across chemicals and species. However, despite the availability of large public toxicity datasets, the data is highly sparse, complicating model development. The aim of this study is to provide insights into how ML can predict toxicity using a large but sparse dataset. We developed models to predict LC50-values, based on experimental LC50-data covering 2431 organic chemicals and 1506 aquatic species from the ECOTOX-database. Several well-known ML techniques were evaluated and a new ML model was developed, inspired by recommender systems. This new model involves a simple linear model that learns low-rank interactions between species and chemicals using factorization machines. We evaluated the predictive performances of the developed models based on two validation settings: 1) predicting unseen chemical-species pairs, and 2) predicting unseen chemicals. The results of this study show that ML models can accurately predict LC50-values in both validation settings. Moreover, we show that the novel factorization machine approach can match well-tuned, complex, ML approaches.

Prior Knowledge for Predictive Modeling: The Case of Acute Aquatic Toxicity

Early assessment of the potential impact of chemicals on health and the environment requires toxicological properties of the molecules. Predictive modeling is often used to estimate the property values in silico from pre-existing experimental data, which is often scarce and uncertain. One of the ways to advance the predictive modeling procedure might be the use of knowledge existing in the field. Scientific publications contain a vast amount of knowledge. However, the amount of manual work required to process the enormous volumes of information gathered in scientific articles might hinder its utilization. This work explores the opportunity of semiautomated knowledge extraction from scientific papers and investigates a few potential ways of its use for predictive modeling. The knowledge extraction and predictive modeling are applied to the field of acute aquatic toxicity. Acute aquatic toxicity is an important parameter of the safety assessment of chemicals. The extensive amount of diverse information existing in the field makes acute aquatic toxicity an attractive area for investigation of knowledge use for predictive modeling. The work demonstrates that the knowledge collection and classification procedure could be useful in hybrid modeling studies concerning the model and predictor selection, addressing data gaps, and evaluation of models’ performance.

Chi-MIC-share: a new feature selection algorithm for quantitative structure–activity relationship models

An algorithm based on an improved maximal information coefficient and a redundant allocation strategy, which can terminate feature selection automatically, is presented.

Modelling the water-plant cuticular polymer matrix membrane partitioning of diverse chemicals in multiple plant species using the support vector machine-based QSAR approach

In this study, a support vector machine (SVM) based multi-species QSAR (quantitative structure-activity relationship) model was developed for predicting the water-plant cuticular polymer matrix membrane (MX) partition coefficient, K_MXw of diverse chemicals using two simple molecular descriptors derived from the chemical structures and following the OECD guidelines. Accordingly, the Lycopersicon esculentum Mill. data were used to construct the QSAR model that was externally validated using three other plant species data. The diversity in chemical structures and end-points were verified using the Tanimoto similarity index and Kruskal-Wallis statistics. The predictive power of the developed QSAR model was tested through rigorous validation, deriving a wide series of statistical checks. The MLOGP was the most influential descriptor identified by the model. The model yielded a correlation (r²) of 0.966 and 0.965 in the training and test data arrays. The developed QSAR model also performed well in another three plant species (r² > 0.955). The results suggest the appropriateness of the developed model to reliably predict the plant chemical interactions in multiple plant species and it can be a useful tool in screening the new chemical for environmental risk assessment.

In silico prediction of pesticide aquatic toxicity with chemical category approaches

Aquatic toxicity is an important issue in pesticide development. In this study, using nine molecular fingerprints to describe pesticides, binary and ternary classification models were constructed to predict aquatic toxicity of pesticides via six machine learning methods: Naïve Bayes (NB), Artificial Neural Network (ANN), k-Nearest Neighbor (kNN), Classification Tree (CT), Random Forest (RF) and Support Vector Machine (SVM). For the binary models, local models were obtained with 829 pesticides on rainbow trout (RT) and 151 pesticides on lepomis (LP), and global models were constructed on the basis of 1258 diverse pesticides on RT and LP and 278 on other fish species. After analyzing the local binary models, we found that fish species caused influence in terms of accuracy. Considering the data size and predictive range, the 1258 pesticides were also used to build global ternary models. The best local binary models were Maccs_ANN for RT and Maccs_SVM for LP, which exhibited accuracies of 0.90 and 0.90, respectively. For global binary models, the best model was Graph_SVM with an accuracy of 0.89. Accuracy of the best global ternary model Graph_SVM was 0.81, which was a little lower than that of the best global binary model. In addition, several substructural alerts were identified including nitrobenzene, chloroalkene and nitrile, which could significantly correlate with pesticide aquatic toxicity. This study provides a useful tool for an early evaluation of pesticide aquatic toxicity in environmental risk assessment.

Ecotoxicity interspecies QAAR models from Daphnia toxicity of pharmaceuticals and personal care products

Pharmaceutical and Personal Care Products (PPCPs) became a class of contaminants of emerging concern because are ubiquitously detected in surface water and soil, where they can affect wildlife. Ecotoxicological data are only available for a few PPCPs, thus modelling approaches are essential tools to maximize the information contained in the existing data. In silico methods may be helpful in filling data gaps for the toxicity of PPCPs towards various ecological indicator organisms. The good correlation between toxicity toward Daphnia magna and those on two fish species (Pimephales promelas and Oncorhynchus mykiss), improved by the addition of one theoretical molecular descriptor, allowed us to develop predictive models to investigate the relationship between toxicities in different species. The aim of this work is to propose quantitative activity-activity relationship (QAAR) models, developed in QSARINS and validated for their external predictivity. Such models can be used to predict the toxicity of PPCPs to a particular species using available experimental toxicity data from a different species, thus reducing the tests on organisms of higher trophic level. Similarly, good QAAR models, implemented by molecular descriptors to improve the quality, are proposed here for fish interspecies. We also comment on the relevance of autocorrelation descriptors in improving all studied interspecies correlations.

QSAR modeling for predicting reproductive toxicity of chemicals in rats for regulatory purposes

The experimental determination of multi-generation reproductive toxicity of chemicals involves high costs and a large number of animal studies over a long period of time. Computational toxicology offers possibilities to overcome such difficulties. In this study, we have established ensemble machine learning (EML) based quantitative structure-activity relationship models for predicting the reproductive toxicity potential (LOAEL) of structurally diverse chemicals in accordance with the OECD guidelines. Accordingly, decision tree forest (DTF) and decision tree boost (DTB) QSAR models were developed using a novel dataset composed of the toxicity endpoints for 334 chemicals. Relevant structural features of chemicals responsible for toxicity potential were identified and used in QSAR modeling. The generalization and prediction abilities of the constructed QSAR models were evaluated by internal and external validation procedures and by deriving several stringent statistical criteria parameters. In the test set, the two models (DTF and DTB) yielded R2 of 0.856 and 0.945, between the experimental and predicted endpoint toxicity values. The models were also evaluated for predictive use through the most recent criteria based on root mean squared error (RMSE) and mean absolute error (MAE). The values of various statistical validation coefficients derived for the test data were above their respective threshold limits and thus put a high confidence in this analysis. The applicability domains of the constructed QSAR models were defined using the leverage and standardization approaches. The results suggest that the proposed QSAR models can reliably predict the reproductive toxicity potential of diverse chemicals and can be useful tools for screening new chemicals for safety assessment.

In silico prediction of the developmental toxicity of diverse organic chemicals in rodents for regulatory purposes

The experimental determination of the developmental toxicity potential (LEL) of chemicals is not only tedious, time and resource intensive, but it also involves unethical tests on animals. In this study, we have established quantitative structure activity relationship (QSAR) models for predicting the developmental toxicity potential of chemicals in rodents following the OECD guidelines. Accordingly, decision tree forest (DTF) and decision tree boost (DTB) based local (L-QSAR), global (G-QSAR) and interspecies quantitative structure activity-activity relationship (ISC QSAAR) models were developed for estimating the LEL (lowest effective level) dose of chemicals for developmental toxicity in rats and rabbits. The structural features of chemicals responsible for developmental toxicity in rodents were extracted and used in QSAR/QSAAR analysis. The external predictive power of the developed models was evaluated through the internal and external validation procedures. In test data, the L-QSAR models (DTF, DTB) yielded R2 values of >0.846 (rat) and >0.906 (rabbit), whereas in G-QSAR, the correlation value was >0.870 between the measured and predicted endpoint values. In ISC QSAAR models, the R2 values in test data were 0.830 (DTF) and 0.927 (DTB), respectively. Values of various statistical validation coefficients derived from the test data (except rm2 in DTF based rat L-QSAR and ISC QSAAR models) were above their respective threshold limits, thus putting a high confidence in this analysis. The prediction quality of the developed QSAR/QSAAR models was also assessed using the mean absolute error (MAE) criteria and found good. The applicability domains of the constructed models were defined using the descriptor range, leverage, and standardization approaches. The results suggest that the developed QSAR/QSAAR models can reliably predict the developmental toxicity potential of structurally diverse chemicals in rodents, generating useful toxicity data for risk assessment in humans.

Modeling the toxicity of chemical pesticides in multiple test species using local and global QSTR approaches

The safety assessment processes require the toxicity data of chemicals in multiple test species and thus, emphasize the need for computational methods capable of toxicity prediction in multiple test species. Pesticides are designed toxic substances and find extensive applications worldwide. In this study, we have established local and global QSTR (quantitative structure-toxicity relationship) and ISC QSAAR (interspecies correlation quantitative structure activity-activity relationship) models for predicting the toxicities of pesticides in multiple aquatic test species using the toxicity data in crustacean (Daphnia magna, Americamysis bahia, Gammarus fasciatus, and Penaeus duorarum) and fish (Oncorhynchus mykiss and Lepomis macrochirus) species in accordance with the OECD guidelines. The ensemble learning based QSTR models (decision tree forest, DTF and decision tree boost, DTB) were constructed and validated using several statistical coefficients derived on the test data. In all the QSTR and QSAAR models, Log P was an important predictor. The constructed local, global and interspecies QSAAR models yielded high correlations (R2) of >0.941; >0.943 and >0.826, respectively between the measured and model predicted endpoint toxicity values in the test data. The performances of the local and global QSTR models were comparable. Furthermore, the chemical applicability domains of these QSTR/QSAAR models were determined using the leverage and standardization approaches. The results suggest for the appropriateness of the developed QSTR/QSAAR models to reliably predict the aquatic toxicity of structurally diverse pesticides in multiple test species and can be used for the screening and prioritization of new pesticides.