Transfer learning for predicting human skin sensitizers

A novel multitask learning algorithm for tasks with distinct chemical space: zebrafish toxicity prediction as an example

Data scarcity is one of the most critical issues impeding the development of prediction models for chemical effects. Multitask learning algorithms leveraging knowledge from relevant tasks showed potential for dealing with tasks with limited data. However, current multitask methods mainly focus on learning from datasets whose task labels are available for most of the training samples. Since datasets were generated for different purposes with distinct chemical spaces, the conventional multitask learning methods may not be suitable. This study presents a novel multitask learning method MTForestNet that can deal with data scarcity problems and learn from tasks with distinct chemical space. The MTForestNet consists of nodes of random forest classifiers organized in the form of a progressive network, where each node represents a random forest model learned from a specific task. To demonstrate the effectiveness of the MTForestNet, 48 zebrafish toxicity datasets were collected and utilized as an example. Among them, two tasks are very different from other tasks with only 1.3% common chemicals shared with other tasks. In an independent test, MTForestNet with a high area under the receiver operating characteristic curve (AUC) value of 0.911 provided superior performance over compared single-task and multitask methods. The overall toxicity derived from the developed models of zebrafish toxicity is well correlated with the experimentally determined overall toxicity. In addition, the outputs from the developed models of zebrafish toxicity can be utilized as features to boost the prediction of developmental toxicity. The developed models are effective for predicting zebrafish toxicity and the proposed MTForestNet is expected to be useful for tasks with distinct chemical space that can be applied in other tasks.Scieific contributionA novel multitask learning algorithm MTForestNet was proposed to address the challenges of developing models using datasets with distinct chemical space that is a common issue of cheminformatics tasks. As an example, zebrafish toxicity prediction models were developed using the proposed MTForestNet which provide superior performance over conventional single-task and multitask learning methods. In addition, the developed zebrafish toxicity prediction models can reduce animal testing.

In Silico Prediction of Skin Sensitization for Compounds via Flexible Evidence Combination Based on Machine Learning and Dempster-Shafer Theory

Skin sensitization is increasingly becoming a significant concern in the development of drugs and cosmetics due to consumer safety and occupational health problems. In silico methods have emerged as alternatives to traditional in vivo animal testing due to ethical and economic considerations. In this study, machine learning methods were used to build quantitative structure-activity relationship (QSAR) models on five skin sensitization data sets (GPMT, LLNA, DPRA, KeratinoSens, and h-CLAT), achieving effective predictive accuracies (correct classification rates of 0.688-0.764 on test sets). To address the complex mechanisms of human skin sensitization, the Dempster-Shafer theory was applied to merge multiple QSAR models, resulting in an evidence-based integrated decision model. Various evidence combinations and combination rules were explored, with the self-defined Q3 rule showing superior balance. The combination of evidence such as GPMT and KeratinoSens and h-CLAT achieved a correct classification rate (CCR) of 0.880 and coverage of 0.893 while maintaining the competitiveness of other combinations. Additionally, the Shapley additive explanations (SHAP) method was used to interpret important features and substructures related to skin sensitization. A comparative analysis of an external human test set demonstrated the superior performance of the proposed method. Finally, to enhance accessibility, the workflow was implemented into a user-friendly software named HSkinSensDS.

Multitask learning for predicting pulmonary absorption of chemicals

Pulmonary absorption is an important route for drug delivery and chemical exposure. To streamline the chemical assessment process for the reduction of animal experiments, several animal-free models were developed for pulmonary absorption research. While Calu-3 and Caco-2 cells and their derived computational models were used in estimating pulmonary permeability, the ex vivo isolated perfused lung (IPL) models are considered more clinically relevant measurements. However, the IPL experiments are resource-consuming making it infeasible for the large-scale screening of potential inhaled toxicants and drugs. In silico models are desirable for estimating pulmonary absorption. This study presented a novel machine learning method that employed an extratrees-based multitask learning approach to predict the IPL absorption rate constant (kaIPL) of various chemicals. The shared permeability knowledge was extracted by simultaneously learning three relevant tasks of Caco-2 and Calu-3 cell permeability and IPL absorption rate. Seven informative physicochemical descriptors were identified. A rigorous evaluation of the developed prediction model showed good performance with a high correlation between predictions and observations (r = 0.84) in the independent test dataset. Two case studies of inhalation drugs and respiratory sensitizers revealed the potential application of this model, which may serve as a valuable tool for predicting pulmonary absorption of chemicals.

Development of a hierarchical support vector regression-based in silico model for the prediction of the cysteine depletion in DPRA

Topical and transdermal treatments have been dramatically growing recently and it is crucial to consider skin sensitization during the drug discovery and development process for these administration routes. Various tests, including animal and non-animal approaches, have been devised to assess the potential for skin sensitization. Furthermore, numerous in silico models have been created, providing swift and cost-effective alternatives to traditional methods such as in vivo, in vitro, and in chemico methods for categorizing compounds. In this study, a quantitative structure-activity relationship (QSAR) model was developed using the innovative hierarchical support vector regression (HSVR) scheme. The aim was to quantitatively predict the potential for skin sensitization by analyzing the percent of cysteine depletion in Direct Peptide Reactivity Assay (DPRA). The results demonstrated accurate, consistent, and robust predictions in the training set, test set, and outlier set. Consequently, this model can be employed to estimate skin sensitization potential of novel or virtual compounds.

Prediction of skin sensitization using machine learning

As global awareness of animal welfare spreads, the development of alternative animal test models is increasingly necessary. The purpose of this study was to develop a practical machine-learning model for skin sensitization using three physicochemical properties of the chemicals: surface tension, melting point, and molecular weight. In this study, a total of 482 chemicals with local lymph node assay results were collected, and 297 datasets with 6 physico-chemical properties were used to develop Random Forest (RF) model for skin sensitization. The developed model was validated with 45 fragrance allergens announced by European Commission. The validation results showed that RF achieved better or similar classification performance with f1-scores of 54% for penal, 82% for ternary, and 96% for binary compared with Support Vector Machine (SVM) (penal, 41%; ternary, 81%; binary, 93%), QSARs (ChemTunes, 72% for ternary; OECD Toolbox, 89% for binary), and a linear model (Kim et al., 2020) (41% for penal), and we recommend the ternary classification based on Global Harmonized System providing more detailed and precise information. In the further study, the proposed model results were experimentally validated with the Direct Peptide Reactivity Assay (DPRA, OECD TG 442C approved model), and the results showed a similar tendency. We anticipate that this study will help to easily and quickly screen chemical sensitization hazards.

A review on in silico prediction of the environmental risks posed by pharmaceutical emerging contaminants

Computer-aided (in silico) prediction has shown good potential to support the environmental risk assessment (ERA) of pharmaceutical emerging contaminants (PECs), allowing low-cost, animal-free, high-throughput screening of multiple potential risks posed by a wide variety of pharmaceuticals in the environment based on insufficient toxicity data. This review provided recent insights regarding the application of in silico approaches in prediction for environmental risks of PECs. Based on the review of 20 included articles from 8 countries published since 2018, we found that the researchers' interest and concern in this research topic were sharply aroused since 2021. Recently, in silico approaches have been widely used for the prediction of bioaccumulation and biodegradability, lethal endpoints, developmental toxicity, mutagenicity, other eco-toxicological effects such as ototoxicity and hematological toxicity, and human health hazards of exposure to PECs. Particular attention has been given to the simultaneous discernment of multiple environmental risks and health effects of PECs based on mechanistic data of pharmaceuticals using advanced bioinformatic methods such as transcriptomic analysis and network pharmacology prediction. In silico software platforms and databases used in the included studies were diversified, and there is currently no standardized and accepted in silico model for ERA of PECs. Date suggested that in silico prediction of the environmental risks posed by PECs is still in its infancy. Considerable critical challenges need to be addressed, including consideration of environmental exposure concentration for PECs, interactions among mixtures of PECs and other contaminants coexisting in environments, and development of in silico models specific to ERA of PECs.

Reconfiguring the online tool of SkinSensPred for predicting skin sensitization of pesticides

Adverse outcome pathway (AOP)-based computational models provide state-of-the-art prediction for human skin sensitizers and are promising alternatives to animal testing. However, little is known about their applicability to pesticides due to scarce pesticide data for evaluation. Moreover, pesticides traditionally have been tested on animals without human data, making validation difficult. Direct application of AOP-based models to pesticides may be inappropriate since their original applicability domains were designed to maximize reliability for human response prediction on diverse chemicals but not pesticides. This study proposed to identify a consensus chemical space with concordant human responses predicted by the SkinSensPred online tool and animal testing data to reduce animal testing. The identified consensus chemical space for non-sensitizers achieved high concordance of 85% and 100% for the cross-validation and independent test, respectively. The reconfigured SkinSensPred can be applied as the first-tier tool for identifying non-sensitizers to reduce. animal testing for pesticides by 19.6%.

SkinSensPred as a Promising in Silico Tool for Integrated Testing Strategy on Skin Sensitization

Skin sensitization is an important regulatory endpoint associated with allergic contact dermatitis. Recently, several adverse outcome pathway (AOP)-based alternative methods were developed to replace animal testing for evaluating skin sensitizers. The AOP-based assays were further integrated as a two-out-of-three method with good predictivity. However, the acquisition of experimental data is resource-intensive. In contrast, an integrated testing strategy (ITS) capable of maximizing the usage of laboratory data from AOP-based and in silico methods was developed as defined approaches (DAs) to both hazard and potency assessment. There are currently two in silico models, namely Derek Nexus and OECD QSAR Toolbox, evaluated in the OECD Testing Guideline No. 497. Since more advanced machine learning algorithms have been proposed for skin sensitization prediction, it is therefore desirable to evaluate their performance under the ITS framework. This study evaluated the performance of a new ITS DA (ITS-SkinSensPred) adopting a transfer learning-based SkinSensPred model. Results showed that the ITS-SkinSensPred has similar or slightly better performance compared to the other ITS models. SkinSensPred-based ITS is expected to be a promising method for assessing skin sensitization.

Computational prediction of Calu-3-based in vitro pulmonary permeability of chemicals

Pulmonary is a potential route for drug delivery and exposure to toxic chemicals. The human bronchial epithelial cell line Calu-3 is generally considered to be a useful in vitro model of pulmonary permeability by calculating the apparent permeability coefficient (Papp) values. Since in vitro experiments are time-consuming and labor-intensive, computational models for pulmonary permeability are desirable for accelerating drug design and toxic chemical assessment. This study presents the first attempt for developing quantitative structure-activity relationship (QSAR) models for addressing this goal. A total of 57 chemicals with Papp values based on Calu-3 experiments was first curated from literature for model development and testing. Subsequently, eleven descriptors were identified by a sequential forward feature selection algorithm to maximize the cross-validation performance of a voting regression model integrating linear regression and nonlinear random forest algorithms. With applicability domain adjustment, the developed model achieved high performance with correlation coefficient values of 0.935 and 0.824 for cross-validation and independent test, respectively. The preliminary results showed that computational models could be helpful for predicting Calu-3-based in vitro Pulmonary Permeability of Chemicals. Future works include the collection of more data for further validating and improving the model.

Ensemble learning for predicting ex vivo human placental barrier permeability

Background

The placental barrier protects the fetus from exposure to some toxicants and is vital for drug development and risk assessment of environmental chemicals. However, in vivo experiments for assessing the placental barrier permeability of chemicals is not ethically acceptable. Although ex vivo placental perfusion methods provide good alternatives for the assessment of placental barrier permeability, the application to a large number of test chemicals could be time- and resource-consuming. Computational prediction models for ex vivo placental barrier permeability are therefore desirable.

Methods

A total of 87 chemicals and corresponding 1444 physicochemical properties were divided into training and test datasets. Three types of algorithms including linear regression, random forest, and ensemble models were applied to develop prediction models for ex vivo placental barrier permeability.

Results

Among the tested models, the ensemble model integrating the previous two methods performed best for predicting ex vivo human placental barrier permeability with correlation coefficients of 0.887 and 0.825 when considering the applicability domain. An additional test on seven newly curated chemicals from the literature showed a good correlation coefficient of 0.879 which was further improved to 0.921 by considering the variation of experiments.

Conclusion

In this study, the first valid predicting model for ex vivo human placental barrier permeability was developed following the OECD guideline. The model is expected to be useful for assessing the human placental barrier permeability and can be integrated with developmental toxicity prediction models for investigating the toxic effects of chemicals on the fetus.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12859-022-04937-y.

A machine learning-driven approach for prioritizing food contact chemicals of carcinogenic concern based on complementary in silico methods

Carcinogenicity is one of the most critical endpoints for the risk assessment of food contact chemicals (FCCs). However, the carcinogenicity of FCCs remains insufficiently investigated. To fill the data gap, the application of standard experimental methods for identifying chemicals of carcinogenic concerns from a large set of FCCs is impractical due to their resource-intensive nature. In contrast, computational methods provide an efficient way to quickly screen chemicals with carcinogenic potential for subsequent experimental validation. Since every model was developed based on a limited number of training samples, the use of single models for carcinogenicity assessment may not cover the complex mechanisms of carcinogenesis. This study proposed a novel machine learning-based weight-of-evidence (WoE) model for prioritizing chemical carcinogenesis. The WoE model can nonlinearly integrate complementary computational methods of structural alerts, quantitative structure-activity relationship models and in silico toxicogenomics models into a WoE-score. Compared to the best single method, the WoE model gained 8% and 19.7% improvement in the area under the receiver operating characteristic curve (AUC) value and chemical coverage, respectively. The prioritization of 1623 FCCs concludes 44 chemicals of high carcinogenic concern. The machine learning-based WoE approach provides a fast and comprehensive way for prioritizing chemicals of carcinogenic concern.

Semi-correlations as a tool to model for skin sensitization

Semi-correlation specifically assesses the correlation between a binary variable and a continuous variable. Semi-correlations were applied to develop binary models for various endpoints. We applied the semi-correlation to develop models of two kinds of skin sensitization one related to animals (local lymph node assay LLNA) and one to human beings (direct peptide reactivity assay DPRA and/or human cell line activation test h-CLAT). The models refer to binary classification for a two-level strategy: the first level (analysis of all compounds) is used in the format "sensitizer or non-sensitizer", and the second level (only sensitizers) is a further classification in the format "strong or weak sensitizer". The ranges of statistical characteristics of the models depend on the endpoint, LLNA or DPRA/h-CLAT: for the first level, sensitivity: 0.69-0.88, specificity: 0.75-0.89, accuracy: 0.77-0.87, Matthew's correlation coefficient (MCC): 0.54-0.57 and for the second level, sensitivity: 0.70-1.0, specificity: 0.78-0.83, accuracy: 0.77-0.87, MCC: 0.54-0.76. Thus, the described approach can be applied to building up models of the skin sensitization potency.

In silico Prediction of Skin Sensitization: Quo vadis?

Skin direct contact with chemical or physical substances is predisposed to allergic contact dermatitis (ACD), producing various allergic reactions, namely rash, blister, or itchy, in the contacted skin area. ACD can be triggered by various extremely complicated adverse outcome pathways (AOPs) remains to be causal for biosafety warrant. As such, commercial products such as ointments or cosmetics can fulfill the topically safe requirements in animal and non-animal models including allergy. Europe, nevertheless, has banned animal tests for the safety evaluations of cosmetic ingredients since 2013, followed by other countries. A variety of non-animal in vitro tests addressing different key events of the AOP, the direct peptide reactivity assay (DPRA), KeratinoSens™, LuSens and human cell line activation test h-CLAT and U-SENS™ have been developed and were adopted in OECD test guideline to identify the skin sensitizers. Other methods, such as the SENS-IS are not yet fully validated and regulatorily accepted. A broad spectrum of in silico models, alternatively, to predict skin sensitization have emerged based on various animal and non-animal data using assorted modeling schemes. In this article, we extensively summarize a number of skin sensitization predictive models that can be used in the biopharmaceutics and cosmeceuticals industries as well as their future perspectives, and the underlined challenges are also discussed.

Computational identification of preservatives with potential neuronal cytotoxicity

Preservatives play a vital role in cosmetics by preventing microbiological contamination for keeping products safe to use. However, a few commonly used preservatives have been suggested to be neurotoxic. Cytotoxicity to neuronal cells is commonly used as the first-tier assay for assessing chemical-induced neurotoxicity. Given the time and resources required for chemical screening, computational methods are attractive alternatives over experimental approaches in prioritizing chemicals prior to further experimental evaluations. In this study, we developed a Quantitative Structure-Activity Relationships (QSAR) model for the identification of potential neurotoxicants. A set of 681 chemicals was utilized to construct a robust prediction model using oversampling and Random Forest algorithms. Within a defined applicability domain, the independent test on 452 chemicals showed a high accuracy of 87.7%. The application of the model to 157 preservatives identified 15 chemicals potentially toxic to neuronal cells. Three of them were further validated by in vitro experiments. The results suggested that further experiments are desirable for assessing the neurotoxicity of the identified preservatives with potential neuronal cytotoxicity.

Prediction of human fetal-maternal blood concentration ratio of chemicals

Background

The measurement of human fetal-maternal blood concentration ratio (logFM) of chemicals is critical for the risk assessment of chemical-induced developmental toxicity. While a few in vitro and ex vivo experimental methods were developed for predicting logFM of chemicals, the obtained experimental results are not able to directly predict in vivo outcomes.

Methods

A total of 55 chemicals with logFM values representing in vivo fetal-maternal blood ratio were divided into training and test datasets. An interpretable linear regression model was developed along with feature selection methods. Cross-validation on training dataset and prediction on independent test dataset were conducted to validate the prediction model.

Results

This study presents the first valid quantitative structure-activity relationship model following the Organisation for Economic Co-operation and Development (OECD) guidelines based on multiple linear regression for predicting in vivo logFM values. The autocorrelation descriptor AATSC1c and information content descriptor ZMIC1 were identified as informative features for predicting logFM. After the adjustment of the applicability domain, the developed model performs well with correlation coefficients of 0.875, 0.850 and 0.847 for model fitting, leave-one-out cross-validation and independent test, respectively. The model is expected to be useful for assessing human transplacental exposure.

Leveraging complementary computational models for prioritizing chemicals of developmental and reproductive toxicity concern: an example of food contact materials

The evaluation of developmental and reproductive toxicity of food contact materials (FCMs) is an important task for food safety. Since traditional experiments are both time-consuming and labor-intensive, only a small number of FCMs have sufficient toxicological data for evaluating their effects on human health. While computational methods such as structural alerts and quantitative structure-activity relationships can serve as first-line tools for the identification of chemicals of high toxicity concern, models with binary outputs and unsatisfied accuracy and coverage prevent the use of computational methods for prioritizing chemicals of high concern. This study proposed a genetic algorithm-based method to develop a weight-of-evidence (WoE) model leveraging complementary methods of structural alerts, quantitative structure-activity relationships and in silico toxicogenomics models for chemical prioritization. The WoE model was applied to evaluate 623 food contact chemicals and identify 26 chemicals of high toxicity concern, where 13 chemicals have been reported to be developmental or reproductive toxic and further experiments are suggested for the remaining 13 chemicals without toxicity data related to developmental and reproductive effects. The proposed WoE model is potentially useful for prioritizing chemicals of high toxicity concern and the methodology may be applied to toxicities other than developmental and reproductive toxicity.

Skin Doctor: Machine Learning Models for Skin Sensitization Prediction that Provide Estimates and Indicators of Prediction Reliability

The ability to predict the skin sensitization potential of small organic molecules is of high importance to the development and safe application of cosmetics, drugs and pesticides. One of the most widely accepted methods for predicting this hazard is the local lymph node assay (LLNA). The goal of this work was to develop in silico models for the prediction of the skin sensitization potential of small molecules that go beyond the state of the art, with larger LLNA data sets and, most importantly, a robust and intuitive definition of the applicability domain, paired with additional indicators of the reliability of predictions. We explored a large variety of molecular descriptors and fingerprints in combination with random forest and support vector machine classifiers. The most suitable models were tested on holdout data, on which they yielded competitive performance (Matthews correlation coefficients up to 0.52; accuracies up to 0.76; areas under the receiver operating characteristic curves up to 0.83). The most favorable models are available via a public web service that, in addition to predictions, provides assessments of the applicability domain and indicators of the reliability of the individual predictions.