Prediction of the mutagenic potential of different pyrrolizidine alkaloids using LAZAR, Random Forest, Support Vector Machines, and Deep Learning Authors Verena Schöning, Christoph Helma, Philipp Boss, Jürgen Drewe **Manuscript in preparation.** Corresponding author: Prof. Dr. Jürgen Drewe, MSc Abstract ======== Pyrrolizidine alkaloids (PAs) are secondary plant metabolites of some plant families, which protect against predators and generally considered as genotoxic and mutagenic. This mutagenicity is also the point of concern in regulatory risk assessment of this substance group [EFSA 2011](#_ENREF_36)[EMA 2014](#_ENREF_38)[2016](#_ENREF_39)(; ; ). Several investigations already showed that the mutagenic potential of PAs is different, and largely depends on the structure. Since only very few of over 600 known PAs are available for *in vitro* or *in vivo* experiments, the mutagenicity of PAs in this study was estimated using four different machine learning techniques LAZAR and Deep Learning, Random Forest and Support Vector Machines. However, all models were not optimal for predicting the genotoxic potential of PAs either due to problems with the applicability domain or due to low performance. Therefore, no estimation regarding the genotoxic potential of single PAs could be made. An analysis of the genotoxic potential of different structural groups, showed promising results. For necine base and necic acid, the results fitted well with literature for three models. However, the prediction of the toxic principle of PAs, dehydropyrrolizidine was only within expectation in one model (TensorFlow-generated Deep Learning model), but not in the other four models. This study shows convincingly the need to critically review and assess the predictions obtained from machine learning approaches by internal cross-validation, but also by external validation through comparison with literature. Introduction ============ Pyrrolizidine alkaloids (PAs) are secondary plant ingredients found in many plant species as protection against predators [Hartmann & Witte 1995](#_ENREF_59)[Langel et al. 2011](#_ENREF_76)(; ). PAs are ester alkaloids, which are composed of a necine base (two fused five-membered rings joined by a nitrogen atom) and one or two necic acid (carboxylic ester arms). The necine base can have different structures and thereby divides PAs into several structural groups, e.g. otonecine, platynecine, and retronecine. The structural groups of the necic acid are macrocyclic diester, open-ring diester and monoester [Langel et al. 2011](#_ENREF_76)(). PA are mainly metabolised in the liver, which is at the same time the main target organ of toxicity [Bull & Dick 1959](#_ENREF_17)[Bull et al. 1958](#_ENREF_18)[Butler et al. 1970](#_ENREF_20)[DeLeve et al. 1996](#_ENREF_33)[Jago 1971](#_ENREF_65)[Li et al. 2011](#_ENREF_78)[Neumann et al. 2015](#_ENREF_99)(; ; ; ; ; ; ). There are three principal metabolic pathways for 1,2-unsaturated PAs [Chen et al. 2010](#_ENREF_26)(): (i) Detoxification by hydrolysis: the ester bond on positions C7 and C9 are hydrolysed by non-specific esterases to release necine base and necic acid, which are then subjected to further phase II-conjugation and excretion. (ii) Detoxification by *N*-oxidation of the necine base (only possible for retronecine-type PAs): the nitrogen is oxidised to form a PA *N*-oxides, which can be conjugated by phase II enzymes e.g. glutathione and then excreted. PA *N*-oxides can be converted back into the corresponding parent PA [Wang et al. 2005](#_ENREF_134)(). (iii) Metabolic activation or toxification: PAs are metabolic activated/ toxified by oxidation (for retronecine-type PAs) or oxidative *N*-demethylation (for otonecine-type PAs [Lin 1998](#_ENREF_82)()). This pathway is mainly catalysed by cytochrome P450 isoforms CYP2B and 3A [Ruan et al. 2014b](#_ENREF_115)(), and results in the formation of dehydropyrrolizidines (DHP, also known as pyrrolic ester or reactive pyrroles). DHPs are highly reactive and cause damage in the cells where they are formed, usually hepatocytes. However, they can also pass from the hepatocytes into the adjacent sinusoids and damage the endothelial lining cells [Gao et al. 2015](#_ENREF_48)() predominantly by reaction with protein, lipids and DNA. There is even evidence, that conjugation of DHP to glutathione, which would generally be considered a detoxification step, could result in reactive metabolites, which might also lead to DNA adduct formation [Xia et al. 2015](#_ENREF_138)(). Due to the ability to form DNA adducts, DNA crosslinks and DNA breaks 1,2-unsaturated PAs are generally considered genotoxic and carcinogenic [Chen et al. 2010](#_ENREF_26)[EFSA 2011](#_ENREF_36)[Fu et al. 2004](#_ENREF_45)[Li et al. 2011](#_ENREF_78)[Takanashi et al. 1980](#_ENREF_126)[Yan et al. 2008](#_ENREF_140)[Zhao et al. 2012](#_ENREF_148)(; ; ; ; ; ; ). Still, there is no evidence yet that PAs are carcinogenic in humans [ANZFA 2001](#_ENREF_4)[EMA 2016](#_ENREF_39)(; ). One general limitation of studies with PAs is the number of different PAs investigated. Around 30 PAs are currently commercially available, therefore all studies focus on these PAs. This is also true for *in vitro* and *in vivo* tests on mutagenicity and genotoxicity. To gain a wider perspective, in this study over 600 different PAs were assessed on their mutagenic potential using four different machine learning techniques. Materials and Methods ===================== Training dataset ---------------- For all methods, the same validated training dataset was used. The training dataset was compiled from the following sources: - Kazius/Bursi Dataset (4337 compounds, [Kazius et al. 2005](#_ENREF_71)()): > - Hansen Dataset (6513 compounds, [Hansen et al. 2009](#_ENREF_57)()): > - EFSA Dataset (695 compounds, [EFSA 2011](#_ENREF_36)()): > Mutagenicity classifications from Kazius and Hansen datasets were used without further processing. To achieve consistency between these datasets, EFSA compounds were classified as mutagenic, if at least one positive result was found for TA98 or T100 Salmonella strains. Dataset merges were based on unique SMILES (*Simplified Molecular Input Line Entry Specification*) strings of the compound structures. Duplicated experimental data with the same outcome was merged into a single value, because it is likely that it originated from the same experiment. Contradictory results were kept as multiple measurements in the database. The combined training dataset contains 8281 unique structures. Source code for all data download, extraction and merge operations is publicly available from the git repository under a GPL3 License. Testing dataset --------------- The testing dataset consisted of 602 different PAs. The compilation of the PA dataset is described in detail in [Schöning et al. (2017)](#_ENREF_119). The PAs were assigned to groups according to structural features of the necine base and necic acid. For the necine base, following groups were assigned: - Retronecine-type (1,2-unstaturated necine base) - Otonecine-type (1,2-unstaturated necine base) - Platynecine-type (1,2-saturated necine base) For the modification of necine base, following groups were assigned: - *N*-oxide-type - Tertiary-type (PAs which were neither from the *N*-oxide- nor > DHP-type) - DHP-type (dehydropyrrolizidine, pyrrolic ester) For the necic acid, following groups were assigned: - Monoester-type - Open-ring diester-type - Macrocyclic diester-type For the Random Forest (RF), Support Vector Machines (SVM), and Deep Learning (DL) models, molecular descriptors of the PAs were calculated using the program PaDEL-Descriptors (version 2.21) [Yap 2011](#_ENREF_142)[2014](#_ENREF_143)(; ). From these descriptors were chosen, which were actually used for the generation of the DL model. LAZAR ----- LAZAR (*lazy structure activity relationships*) is a modular framework for read-across model development and validation. It follows the following basic workflow: For a given chemical structure LAZAR: - searches in a database for similar structures (neighbours) with experimental data, - builds a local QSAR model with these neighbours and - uses this model to predict the unknown activity of the query compound. This procedure resembles an automated version of read across predictions in toxicology, in machine learning terms it would be classified as a k-nearest-neighbour algorithm. Apart from this basic workflow, LAZAR is completely modular and allows the researcher to use any algorithm for similarity searches and local QSAR (*Quantitative structure--activity relationship*) modelling. Algorithms used within this study are described in the following sections. ### Neighbour identification Similarity calculations were based on MolPrint2D fingerprints [Bender et al. 2004](#_ENREF_8)() from the OpenBabel cheminformatics library [O\'Boyle et al. 2011](#_ENREF_104)(). The MolPrint2D fingerprint uses atom environments as molecular representation, which resembles basically the chemical concept of functional groups. For each atom in a molecule, it represents the chemical environment using the atom types of connected atoms. MolPrint2D fingerprints are generated dynamically from chemical structures and do not rely on predefined lists of fragments (such as OpenBabel FP3, FP4 or MACCs fingerprints or lists of toxicophores/toxicophobes). This has the advantage that they may capture substructures of toxicological relevance that are not included in other fingerprints. From MolPrint2D fingerprints a feature vector with all atom environments of a compound can be constructed that can be used to calculate chemical similarities. The chemical similarity between two compounds a and b is expressed as the proportion between atom environments common in both structures A ∩ B and the total number of atom environments A U B (Jaccard/Tanimoto index). $$sim = \frac{\left| A\ \cap B \right|}{\left| A\ \cup B \right|}$$ Threshold selection is a trade-off between prediction accuracy (high threshold) and the number of predictable compounds (low threshold). As it is in many practical cases desirable to make predictions even in the absence of closely related neighbours, we follow a tiered approach: - First a similarity threshold of 0.5 is used to collect neighbours, to create a local QSAR model and to make a prediction for the query compound. - If any of these steps fails, the procedure is repeated with a similarity threshold of 0.2 and the prediction is flagged with a warning that it might be out of the applicability domain of the training data. - Similarity thresholds of 0.5 and 0.2 are the default values chosen > by the software developers and remained unchanged during the > course of these experiments. Compounds with the same structure as the query structure are automatically eliminated from neighbours to obtain unbiased predictions in the presence of duplicates. ### Local QSAR models and predictions Only similar compounds (neighbours) above the threshold are used for local QSAR models. In this investigation, we are using a weighted majority vote from the neighbour's experimental data for mutagenicity classifications. Probabilities for both classes (mutagenic/non-mutagenic) are calculated according to the following formula and the class with the higher probability is used as prediction outcome. $$p_{c} = \ \frac{\sum_{}^{}\text{sim}_{n,c}}{\sum_{}^{}\text{sim}_{n}}$$ $p_{c}$ Probability of class c (e.g. mutagenic or non-mutagenic)\ $\sum_{}^{}\text{sim}_{n,c}$ Sum of similarities of neighbours with class c\ $\sum_{}^{}\text{sim}_{n}$ Sum of all neighbours ### Applicability domain The applicability domain (AD) of LAZAR models is determined by the structural diversity of the training data. If no similar compounds are found in the training data no predictions will be generated. Warnings are issued if the similarity threshold had to be lowered from 0.5 to 0.2 in order to enable predictions. Predictions without warnings can be considered as close to the applicability domain and predictions with warnings as more distant from the applicability domain. Quantitative applicability domain information can be obtained from the similarities of individual neighbours. ### Availability - LAZAR experiments for this manuscript: [https://git.in-silico.ch/pyrrolizidine](https://deref-gmx.net/mail/client/Yn0laI8dUvs/dereferrer/?redirectUrl=https%3A%2F%2Fgit.in-silico.ch%2Fpyrrolizidine) (source code, GPL3) - LAZAR framework: [https://git.in-silico.ch/lazar](https://deref-gmx.net/mail/client/v26UgZbKEpE/dereferrer/?redirectUrl=https%3A%2F%2Fgit.in-silico.ch%2Flazar) (source code, GPL3) - LAZAR GUI: [https://git.in-silico.ch/lazar-gui](https://deref-gmx.net/mail/client/QstEPrpbcqQ/dereferrer/?redirectUrl=https%3A%2F%2Fgit.in-silico.ch%2Flazar-gui) (source code, GPL3) - Public web interface: [https://lazar.in-silico.ch](https://deref-gmx.net/mail/client/Gln3hLem0DY/dereferrer/?redirectUrl=https%3A%2F%2Flazar.in-silico.ch) Random Forest, Support Vector Machines, and Deep Learning in R-project ---------------------------------------------------------------------- In comparison to LAZAR, three other models (Random Forest (RF), Support Vector Machines (SVM), and Deep Learning (DL)) were evaluated. For the generation of these models, molecular 1D and 2D descriptors of the training dataset were calculated using PaDEL-Descriptors (version 2.21) [Yap 2011](#_ENREF_142)[2014](#_ENREF_143)(; ). As the training dataset contained over 8280 instances, it was decided to delete instances with missing values during data pre-processing. Furthermore, substances with equivocal outcome were removed. The final training dataset contained 8080 instances with known mutagenic potential. The RF, SVM, and DL models were generated using the R software (R-project for Statistical Computing, *;* version 3.3.1), specific R packages used are identified for each step in the description below. During feature selection, descriptor with near zero variance were removed using '*NearZeroVar*'-function (package 'caret'). If the percentage of the most common value was more than 90% or when the frequency ratio of the most common value to the second most common value was greater than 95:5 (e.g. 95 instances of the most common value and only 5 or less instances of the second most common value), a descriptor was classified as having a near zero variance. After that, highly correlated descriptors were removed using the '*findCorrelation*'-function (package 'caret') with a cut-off of 0.9. This resulted in a training dataset with 516 descriptors. These descriptors were scaled to be in the range between 0 and 1 using the '*preProcess*'-function (package 'caret'). The scaling routine was saved in order to apply the same scaling on the testing dataset. As these three steps did not consider the outcome, it was decided that they do not need to be included in the cross-validation of the model. To further reduce the number of features, a LASSO (*least absolute shrinkage and selection operator*) regression was performed using the '*glmnet*'-function (package '*glmnet*'). The reduced dataset was used for the generation of the pre-trained models. For the RF model, the '*randomForest*'-function (package '*randomForest*') was used. A forest with 1000 trees with maximal terminal nodes of 200 was grown for the prediction. The '*svm*'-function (package 'e1071') with a *radial basis function kernel* was used for the SVM model. The DL model was generated using the '*h2o.deeplearning*'-function (package '*h2o*'). The DL contained four hidden layer with 70, 50, 50, and 10 neurons, respectively. Other hyperparameter were set as follows: l1=1.0E-7, l2=1.0E-11, epsilon = 1.0E-10, rho = 0.8, and quantile\_alpha = 0.5. For all other hyperparameter, the default values were used. Weights and biases were in a first step determined with an unsupervised DL model. These values were then used for the actual, supervised DL model. To validate these models, an internal cross-validation approach was chosen. The training dataset was randomly split in training data, which contained 95% of the data, and validation data, which contain 5% of the data. A feature selection with LASSO on the training data was performed, reducing the number of descriptors to approximately 100. This step was repeated five times. Based on each of the five different training data, the predictive models were trained and the performance tested with the validation data. This step was repeated 10 times. Furthermore, a y-randomisation using the RF model was performed. During y-randomisation, the outcome (y-variable) is randomly permuted. The theory is that after randomisation of the outcome, the model should not be able to correlate the outcome to the properties (descriptor values) of the substances. The performance of the model should therefore indicate a by change prediction with an accuracy of about 50%. If this is true, it can be concluded that correlation between actual outcome and properties of the substances is real and not by chance [Rücker et al. 2007](#_ENREF_117)(). ![](./media/media/image1.png){width="6.26875in" height="5.486111111111111in"} Figure 1: Flowchart of the generation and validation of the models generated in R-project Deep Learning in TensorFlow --------------------------- Alternatively, a DL model was established with Python-based TensorFlow program () using the high-level API Keras () to build the models. Data pre-processing was done by rank transformation using the '*QuantileTransformer*' procedure. A sequential model has been used. Four layers have been used: input layer, two hidden layers (with 12, 8 and 8 nodes, respectively) and one output layer. For the output layer, a sigmoidal activation function and for all other layers the ReLU ('*Rectified Linear Unit*') activation function was used. Additionally, a L^2^-penalty of 0.001 was used for the input layer. For training of the model, the ADAM algorithm was used to minimise the cross-entropy loss using the default parameters of Keras. Training was performed for 100 epochs with a batch size of 64. The model was implemented with Python 3.6 and Keras. For training of the model, a 6-fold cross-validation was used. Accuracy was estimated by ROC-AUC and confusion matrix. Results ======= LAZAR ----- For 46 PAs, no prediction could be made. 26 PAs had no neighbours and 20 PAs had only one neighbour. For additional 396 PAs, the similarity threshold had to be reduced from 0.5 to 0.2 to obtain enough neighbours for a prediction. This means that these substances might not be within the applicability domain (AD). Therefore, only 160 of 602 PAs were well within the stricter AD with the similarity threshold of 0.5 and 556 PAs in the AD with the similarity threshold of 0.2. ![](./media/media/image2.png){width="5.905511811023622in" height="3.868241469816273in"} Figure 2: Genotoxic potential of the different PA groups as predicted by LAZAR, using the **similarity threshold** **of 0.5**. *Genotoxic*: percentage number of compounds per group, which were predicted to be genotoxic.\ *Not genotoxic*: percentage number of compounds per group, which were predicted to be not genotoxic\ *Outside AD*: percentage number of compounds per group, which were outside the applicability domain (AD). ![](./media/media/image3.png){width="5.905511811023622in" height="3.868241469816273in"} Figure 3: Genotoxic potential of the different PA groups as predicted by LAZAR, using the **similarity threshold of 0.2** *Genotoxic*: percentage number of compounds per group, which were predicted to be genotoxic.\ *Not genotoxic*: percentage number of compounds per group, which were predicted to be not genotoxic\ *Outside AD*: percentage number of compounds per group, which were outside the applicability domain (AD). Interestingly, using both similarity thresholds (e.g. 0.2 and 0.5), the majority of PAs in all groups except otonecine, were predicted to be not genotoxic. The following rank order for genotoxicity probability can be deduced from the results of both similarity thresholds: - Necine base: platynecine ≤ retronecine \<\< otonecine - Necic acid: monoester \< diester \< macrocyclic diester - Modification of necine base: *N*-oxide \< DHP \< tertiary PA Random Forest, Support Vector Machines, and Deep Learning --------------------------------------------------------- Applicability domain The AD of the training dataset and the PA dataset was evaluated using the Jaccard distance. A Jaccard distance of '0' indicates that the substances are similar, whereas a value of '1' shows that the substances are different. The Jaccard distance was below 0.2 for all PAs relative to the training dataset. Therefore, PA dataset is within the AD of the training dataset and the models can be used to predict the genotoxic potential of the PA dataset. y-randomisation After y-randomisation of the outcome, the accuracy and CCR are around 50%, indicating a chance in the distribution of the results. This shows, that the outcome is actually related to the predictors and not by chance. Random Forest The validation showed that the RF model has an accuracy of 64%, a sensitivity of 66% and a specificity of 63%. The confusion matrix of the model, calculated for 8080 instances, is provided in Table 1. Table 1: Confusion matrix of the RF model Predicted genotoxicity ----------------------- ------------------------ ---------- ---------- ------------- Measured genotoxicity ***PP*** ***PN*** ***Total*** ***TP*** 2274 1163 3437 ***TN*** 1736 2907 4643 ***Total*** 4010 4070 8080 PP: Predicted positive; PN: Predicted negative, TP: True positive, TN: True negative In general, the majority of PAs were considered to be not genotoxic by the RF model (Figure 4). ![](./media/media/image4.png){width="6.063194444444444in" height="3.8756944444444446in"} Figure 4: Genotoxic potential of the different PA groups as predicted by **RF model** *Genotoxic*: percentage number of compounds per group, which was predicted to be genotoxic.\ *Not genotoxic*: percentage number of compounds per group, which was predicted to be not genotoxic. From the results, the following rank orders of genotoxic potential could be deduced: - Necine base: platynecine \< retronecine \< otonecine - Necic acid: monoester (= 0%) \< diester \< macrocyclic diester - Modification of necine base: *N*-oxide = dehydropyrrolizidine (0%) \< tertiary PA Support Vector Machines The validation showed that the SVM model has an accuracy of 62%, a sensitivity of 65% and a specificity of 60%. The confusion matrix of SVM model, calculated for 8080 instances, is provided in Table 2. Table 2: Confusion matrix of the SVM model Predicted genotoxicity ----------------------- ------------------------ ---------- ---------- ------------- Measured genotoxicity ***PP*** ***PN*** ***Total*** ***TP*** 2057 1107 3164 ***TN*** 1953 2963 4916 ***Total*** 4010 4070 8080 PP: Predicted positive; PN: Predicted negative, TP: True positive, TN: True negative In the SVM model, also the majority of PAs were considered to be not genotoxic (Figure 5). ![](./media/media/image5.png){width="6.063194444444444in" height="3.9694444444444446in"} Figure 5: Genotoxic potential of the different PA groups as predicted by **SVM model** *Genotoxic*: percentage number of compounds per group, which was predicted to be genotoxic.\ *Not genotoxic*: percentage number of compounds per group, which was predicted to be not genotoxic From the results, the following rank orders of genotoxic potential could be deduced: - Necine base: otonecine \< platynecine = retronecine - Necic acid: macrocyclic diester \< monoester = diester - Modification of necine base: dehydropyrrolizidine \< tertiary PA \< *N*-oxide  Deep Learning (R-project) The validation showed that the DL model generated in R has an accuracy of 59%, a sensitivity of 89% and a specificity of 30%. The confusion matrix of the model, normalised to 8080 instances, is provided in Table 3. Table 3: Confusion matrix of the DL model (R-project) Predicted genotoxicity ----------------------- ------------------------ ---------- ---------- ------------- Measured genotoxicity ***PP*** ***PN*** ***Total*** ***TP*** 3575 435 4010 ***TN*** 2853 1217 4070 ***Total*** 6428 1652 8080 PP: Predicted positive; PN: Predicted negative, TP: True positive, TN: True negative In contrast, the majority of PAs were considered to be genotoxic by the DL model in R (Figure 6). ![](./media/media/image6.png){width="6.063194444444444in" height="3.982638888888889in"} Figure 6: Genotoxic potential of the different PA groups as predicted by **DL model (R-project)** *Genotoxic*: percentage number of compounds per group, which was predicted to be genotoxic.\ *Not genotoxic*: percentage number of compounds per group, which was predicted to be not genotoxic From the results, the following rank orders of genotoxic potential could be proposed: - Necine base: platynecine \< retronecine \< otonecine - Necic acid: monoester \< diester \< macrocyclic diester - Modification of necine base: tertiary PA = dehydropyrrolizidine \< *N*-oxide. DL model (TensorFlow) The validation showed that the DL model generated in TensorFlow has an accuracy of 68%, a sensitivity of 70% and a specificity of 46%. The confusion matrix of the model, normalised to 8080 instances, is provided in Table 4. Table 4: Confusion matrix of the DL model (TensorFlow) Predicted genotoxicity ----------------------- ------------------------ ---------- ---------- ------------- Measured genotoxicity ***PP*** ***PN*** ***Total*** ***TP*** 2851 1227 4078 ***TN*** 1825 2177 4002 ***Total*** 4676 3404 8080 PP: Predicted positive; PN: Predicted negative, TP: True positive, TN: True negative The ROC curves from the 6-fold validation are shown in Figure 7. ![C:\\Users\\JDrewe\\AppData\\Local\\Microsoft\\Windows\\INetCache\\Content.MSO\\7CFE5F13.tmp](./media/media/image7.png){width="3.825in" height="2.7327045056867894in"} Figure 7: Six-fold cross-validation of TensorFlow DL model show an average area under the ROC-curve (ROC-AUC; measure of accuracy) of 68%. In contrast to the DL generated in R, the DL model generated in TensorFlow predicted the majority of PAs as not genotoxic. ![C:\\Users\\JDrewe\\AppData\\Local\\Microsoft\\Windows\\INetCache\\Content.MSO\\4F678848.tmp](./media/media/image8.png){width="6.26875in" height="3.6993055555555556in"} Figure 8: Genotoxic potential of the different PA groups as predicted by **DL model (TensorFlow)** *Genotoxic*: percentage number of compounds per group, which was predicted to be genotoxic.\ *Not genotoxic*: percentage number of compounds per group, which was predicted to be not genotoxic The following rank orders of genotoxic potential could be proposed based on the results: - Necine base: platynecine \< otonecine \< retronecine  - Necic acid: monoester \< diester \< macrocyclic diester - Modification of necine base: tertiary PA \< *N*-oxide \<\< dehydropyrrolizidine. In summary, the validation results of the four methods are presented in the following table. Table 5 Results of the cross-validation of the four models and after y-randomisation ---------------------------------------------------------------------- Accuracy CCR Sensitivity Specificity ----------------------- ---------- ------- ------------- ------------- RF model 64.1% 64.4% 66.2% 62.6% SVM model 62.1% 62.6% 65.0% 60.3% DL model\ 59.3% 59.5% 89.2% 29.9% (R-project) DL model (TensorFlow) 68% 62.2% 69.9% 45.6% y-randomisation 50.5% 50.4% 50.3% 50.6% ---------------------------------------------------------------------- CCR (correct classification rate) Discussion ========== General model performance Based on the results of the cross-validation for all models, LAZAR, RF, SVM, DL (R-project) and DL (TensorFlow) it can be state that the prediction results are not optimal due to different reasons. The accuracy as measured during cross-validation of the four models (RF, SVM, DL (R-project and TensorFlow)) was partly low with CCR values between 59.3 and 68%, with the R-generated DL model and the TensorFlow-generated DL model showing the worst and the best performance, respectively. The validation of the R-generated DL model revealed a high sensitivity (89.2%) but an unacceptably low specificity of 29.9% indicating a high number of false positive estimates. The TensorFlow-generated DL model, however, showed an acceptable but not optimal accuracy of 68%, a sensitivity of 69.9% and a specificity of 45.6%. The low specificity indicates that both DL models tends to predict too many instances as positive (genotoxic), and therefore have a high false positive rate. This allows at least with the TensorFlow generated DL model to make group statements, but the confidence for estimations of single PAs appears to be insufficiently low. Several factors have likely contributed to the low to moderate performance of the used methods as shown during the cross-validation: 1. The outcome in the training dataset was based on the results of AMES tests for genotoxicity [ICH 2011](#_ENREF_63)(), an *in vitro* test in different strains of the bacteria *Salmonella typhimurium*. In this test, mutagenicity is evaluated with and without prior metabolic activation of the test substance. Metabolic activation could result in the formation of genotoxic metabolites from non-genotoxic parent compounds. However, no distinction was made in the training dataset between substances that needed metabolic activation before being mutagenic and those that were mutagenic without metabolic activation. LAZAR is able to handle this 'inaccuracy' in the training dataset well due to the way the algorithm works: LAZAR predicts the genotoxic potential based on the neighbours of substances with comparable structural features, considering mutagenic and not mutagenic neighbours. Based on the structural similarity, a probability for mutagenicity and no mutagenicity is calculated independently from each other (meaning that the sum of probabilities does not necessarily adds up to 100%). The class with the higher outcome is then the overall outcome for the substance. > In contrast, the other models need to be trained first to recognise > the structural features that are responsible for genotoxicity. > Therefore, the mixture of substances being mutagenic with and without > metabolic activation in the training dataset may have adversely > affected the ability to separate the dataset in two distinct classes > and thus explains the relatively low performance of these models. 2. Machine learning algorithms try to find an optimized solution in a high-dimensional (one dimension per each predictor) space. Sometimes these methods do not find the global optimum of estimates but only local (not optimal) solutions. Strategies to find the global solutions are systematic variation (grid search) of the hyperparameters of the methods, which may be very time consuming in particular in large datasets. Mutagenicity of PAs Due to the low to moderate predictivity of all models, quantitative statement on the genotoxicity of single PAs cannot be made with sufficient confidence. The predictions of the SVM model did not fit with the other models or literature, and are therefore not further considered in the discussion. Necic acid The rank order of the necic acid is comparable in the four models considered (LAZAR, RF and DL (R-project and TensorFlow). PAs from the monoester type had the lowest genotoxic potential, followed by PAs from the open-ring diester type. PAs with macrocyclic diesters had the highest genotoxic potential. The result fit well with current state of knowledge: in general, PAs, which have a macrocyclic diesters as necic acid, are considered more toxic than those with an open-ring diester or monoester [EFSA 2011](#_ENREF_36)[Fu et al. 2004](#_ENREF_45)[Ruan et al. 2014b](#_ENREF_115)(; ; ). Necine base The rank order of necine base is comparable in LAZAR, RF, and DL (R-project) models: with platynecine being less or as genotoxic as retronecine, and otonecine being the most genotoxic. In the TensorFlow-generate DL model, platynecine also has the lowest genotoxic probability, but are then followed by the otonecines and last by retronecine. These results partly correspond to earlier published studies. Saturated PAs of the platynecine-type are generally accepted to be less or non-toxic and have been shown in *in vitro* experiments to form no DNA-adducts [Xia et al. 2013](#_ENREF_139)(). Therefore, it is striking, that 1,2-unsaturated PAs of the retronecine-type should have an almost comparable genotoxic potential in the LAZAR and DL (R-project) model. In literature, otonecine-type PAs were shown to be more toxic than those of the retronecine-type [Li et al. 2013](#_ENREF_80)(). Modifications of necine base The group-specific results of the TensorFlow-generated DL model appear to reflect the expected relationship between the groups: the low genotoxic potential of *N*-oxides and the highest potential of dehydropyrrolizidines [Chen et al. 2010](#_ENREF_26)(). In the LAZAR model, the genotoxic potential of dehydropyrrolizidines (DHP) (using the extended AD) is comparable to that of tertiary PAs. Since, DHP is regarded as the toxic principle in the metabolism of PAs, and known to produce protein- and DNA-adducts [Chen et al. 2010](#_ENREF_26)(), the LAZAR model did not meet this expectation it predicted the majority of DHP as being not genotoxic. However, the following issues need to be considered. On the one hand, all DHP were outside of the stricter AD of 0.5. This indicates that in general, there might be a problem with the AD. In addition, DHP has two unsaturated double bounds in its necine base, making it highly reactive. DHP and other comparable molecules have a very short lifespan, and usually cannot be used in *in vitro* experiments. This might explain the absence of suitable neighbours in LAZAR. Furthermore, the probabilities for this substance groups needs to be considered, and not only the consolidated prediction. In the LAZAR model, all DHPs had probabilities for both outcomes (genotoxic and not genotoxic) mainly below 30%. Additionally, the probabilities for both outcomes were close together, often within 10% of each other. The fact that for both outcomes, the probabilities were low and close together, indicates a lower confidence in the prediction of the model for DHPs. In the DL (R-project) and RF model, *N*-oxides have a by far more genotoxic potential that tertiary PAs or dehydropyrrolizidines. As PA *N*-oxides are easily conjugated for extraction, they are generally considered as detoxification products, which are *in vivo* quickly renally eliminated [Chen et al. 2010](#_ENREF_26)(). On the other hand, *N*-oxides can be also back-transformed to the corresponding tertiary PA [Wang et al. 2005](#_ENREF_134)(). Therefore, it may be questioned, whether *N*-oxides themselves are generally less genotoxic than the corresponding tertiary PAs. However, in the groups of modification of the necine base, dehydropyrrolizidine, the toxic principle of PAs, should have had the highest genotoxic potential. Taken together, the predictions of the modifications of the necine base from the LAZAR, RF and R-generated DL model cannot -- in contrast to the TensorFlow DL model - be considered as reliable. Overall, when comparing the prediction results of the PAs to current published knowledge, it can be concluded that the performance of most models was low to moderate. This might be contributed to the following issues: 1. In the LAZAR model, only 26.6% PAs were within the stricter AD. With the extended AD, 92.3% of the PAs could be included in the prediction. Even though the Jaccard distance between the training dataset and the PA dataset for the RF, SVM, and DL (R-project and TensorFlow) models was small, suggesting a high similarity, the LAZAR indicated that PAs have only few local neighbours, which might adversely affect the prediction of the mutagenic potential of PAs. 2. All above-mentioned models were used to predict the mutagenicity of PAs. PAs are generally considered to be genotoxic, and the mode of action is also known. Therefore, the fact that some models predict the majority of PAs as not genotoxic seems contradictory. To understand this result, the basis, the training dataset, has to be considered. The mutagenicity of in the training dataset are based on data of mutagenicity in bacteria. There are some studies, which show mutagenicity of PAs in the AMES test [Chen et al. 2010](#_ENREF_26)(). Also, [Rubiolo et al. (1992)](#_ENREF_116) examined several different PAs and several different extracts of PA-containing plants in the AMES test. They found that the AMES test was indeed able to detect mutagenicity of PAs, but in general, appeared to have a low sensitivity. The pre-incubation phase for metabolic activation of PAs by microsomal enzymes was the sensitivity-limiting step. This could very well mean that this is also reflected in the QSAR models. Conclusions =========== In this study, an attempt was made to predict the genotoxic potential of PAs using five different machine learning techniques (LAZAR, RF, SVM, DL (R-project and TensorFlow). The results of all models fitted only partly to the findings in literature, with best results obtained with the TensorFlow DL model. Therefore, modelling allows statements on the relative risks of genotoxicity of the different PA groups. Individual predictions for selective PAs appear, however, not reliable on the current basis of the used training dataset. This study emphasises the importance of critical assessment of predictions by QSAR models. This includes not only extensive literature research to assess the plausibility of the predictions, but also a good knowledge of the metabolism of the test substances and understanding for possible mechanisms of toxicity. In further studies, additional machine learning techniques or a modified (extended) training dataset should be used for an additional attempt to predict the genotoxic potential of PAs. References ========== []{#_ENREF_4 .anchor} []{#_ENREF_8 .anchor} []{#_ENREF_17 .anchor} []{#_ENREF_18 .anchor} []{#_ENREF_20 .anchor} []{#_ENREF_26 .anchor} []{#_ENREF_33 .anchor} []{#_ENREF_36 .anchor} []{#_ENREF_38 .anchor} []{#_ENREF_39 .anchor} []{#_ENREF_45 .anchor} []{#_ENREF_48 .anchor} []{#_ENREF_57 .anchor} []{#_ENREF_59 .anchor} []{#_ENREF_63 .anchor} []{#_ENREF_65 .anchor} []{#_ENREF_71 .anchor} []{#_ENREF_76 .anchor} []{#_ENREF_78 .anchor} []{#_ENREF_80 .anchor} []{#_ENREF_82 .anchor} []{#_ENREF_99 .anchor} []{#_ENREF_104 .anchor} []{#_ENREF_115 .anchor} []{#_ENREF_116 .anchor} []{#_ENREF_117 .anchor} []{#_ENREF_119 .anchor} []{#_ENREF_126 .anchor} []{#_ENREF_134 .anchor} []{#_ENREF_138 .anchor} []{#_ENREF_139 .anchor} []{#_ENREF_140 .anchor} []{#_ENREF_142 .anchor} []{#_ENREF_143 .anchor} []{#_ENREF_148 .anchor} Aguer C, Gambarotta D, Mailloux RJ, Moffat C, Dent R, et al. 2011. Galactose enhances oxidative metabolism and reveals mitochondrial dysfunction in human primary muscle cells. PLoS One 6:e28536Ahmed SN, Siddiqi ZA. 2006. Antiepileptic drugs and liver disease. Seizure 15:156-64Aleo MD, Luo Y, Swiss R, Bonin PD, Potter DM, Will Y. 2014. Human drug-induced liver injury severity is highly associated with dual inhibition of liver mitochondrial function and bile salt export pump. Hepatology (Baltimore, Md) 60:1015-22ANZFA. 2001. Pyrrolizidine alkaloids in food. A Toxicological Review and Risk Assessment. ed. Authority, ANZF, pp. 1-16Armstrong SJ, Zuckerman AJ, Bird RG. 1972. Induction of morphological changes in human embryo liver cells by the pyrrolizidine alkaloid lasiocarpine. British journal of experimental pathology 53:145-9Barysz M, Jashari G, Lall RS, Srivastava AK, Trinajstic N. 1983. On the distance matrix of molecules containing heteroatoms. In *Chemical Applications of Topology and Graph Theory*, pp. 222-30. Amsterdam, The Netherlands: ElsevierBasak SC, Harriss DK, Magnuson VR. Comparative Study of Lipophilicity \versus\ Topological Molecular Descriptors in Biological Correlations. Journal of Pharmaceutical Sciences 73:429-37Bender A, Mussa HY, Glen RC, Reiling S. 2004. Molecular similarity searching using atom environments, information-based feature selection, and a naive Bayesian classifier. J Chem Inf Comput Sci 44:170-8Benichou C, Danan G, Flahault A. 1993. Causality assessment of adverse reactions to drugs\--II. An original model for validation of drug causality assessment methods: case reports with positive rechallenge. J Clin Epidemiol 46:1331-6Bergmeir C, Benítez JM. 2012. Neural Networks in R Using the Stuttgart Neural Network Simulator: RSNNS. Journal of Statistical Software 46:1-26Bishop-Bailey D, Thomson S, Askari A, Faulkner A, Wheeler-Jones C. 2014. Lipid-metabolizing CYPs in the regulation and dysregulation of metabolism. Annu Rev Nutr 34:261-79Blower PE, Cross KP. 2006. Decision Tree Methods in Pharmaceutical Research. Current topics in medicinal chemistry 6:31-9Boelsterli UA, Lee KK. 2014. Mechanisms of isoniazid-induced idiosyncratic liver injury: emerging role of mitochondrial stress. Journal of gastroenterology and hepatology 29:678-87Bramer M. 2013. Principles of Data Mining. p. 444: Springer-VerlagBreimann L. 2001. Random Forests. Machine Learning 45:5-32Breimann L. 2003. Manual-Setting Up, Using, And Understanding Random Forests V4.0.1-33Bull LB, Dick AT. 1959. The chronic pathological effects on the liver of the rat of the pyrrolizidine alkaloids heliotrine, lasiocarpine and their N-oxides. J Path Bact 78:483-502Bull LB, Dick AT, McKenzie JS. 1958. The actue toxic effects of heliotrine and lasiocarpine, and their N-oxides, on the rat. J Path Bact 75:17-25Burden FR. 1989. Molecular identification number for substructure searches. Journal of Chemical Information and Computer Sciences 29:225-7Butler WH, Mattocks AR, Barnes JM. 1970. Lesions in the liver and lungs of rats given pyrrole derivates of pyrrolizidine alkaloids. J Path 100:169-75Chai J, He Y, Cai SY, Jiang Z, Wang H, et al. 2012. Elevated hepatic multidrug resistance-associated protein 3/ATP-binding cassette subfamily C 3 expression in human obstructive cholestasis is mediated through tumor necrosis factor alpha and c-Jun NH2-terminal kinase/stress-activated protein kinase-signaling pathway. Hepatology 55:1485-94Chalhoub WM, Sliman KD, Arumuganathan M, Lewis JH. 2014. Drug-induced liver injury: what was new in 2013? Expert Opin Drug Metab Toxicol 10:959-80Chawla NV, Bowyer KW, Hall LO. 2002. SMOTE: Synthetic Minority Over-sampling Technique. Journal of Artificial Intelligence Research 16:321--57Chen M, Borlak J, Tong W. 2013. High lipophilicity and high daily dose of oral medications are associated with significant risk for drug-induced liver injury. Hepatology (Baltimore, Md) 58:388-96Chen M, Suzuki A, Thakkar S, Yu K, Hu C, Tong W. 2016. DILIrank: the largest reference drug list ranked by the risk for developing drug-induced liver injury in humans. Drug Discov Today 21:648-53Chen T, Mei N, Fu PP. 2010. Genotoxicity of pyrrolizidine alkaloids. J Appl Toxicol 30:183-96Crabtree HG. 1928. The carbohydrate metabolism of certain pathological overgrowths Biochem J 22:1289-98Daly AK, Donaldson PT, Bhatnagar P, Shen Y, Pe\'er I, et al. 2009. HLA-B\*5701 genotype is a major determinant of drug-induced liver injury due to flucloxacillin. Nature genetics 41:816-9Danan G, Benichou C. 1993. Causality assessment of adverse reactions to drugs\--I. A novel method based on the conclusions of international consensus meetings: application to drug-induced liver injuries. J Clin Epidemiol 46:1323-30Dar AC, Shokat KM. 2011. The evolution of protein kinase inhibitors from antagonists to agonists of cellular signaling. Annu Rev Biochem 80:769-95de Wildt SN, Kearns GL, Leeder JS, van den Anker JN. 1999. Cytochrome P450 3A: ontogeny and drug disposition. Clin Pharmacokinet 37:485-505DeLeve LD, Ito Y, Bethea NW, McCuskey MK, Wang X, McCuskey RS. 2003. Embolization by sinusoidal lining cells obstructs the microcirculation in rat sinusoidal obstruction syndrome. Am J Physiol Gastrointest Liver Physiol 284:G1045--G52DeLeve LD, Wang X, Kuhlenkamp JF, Kaplowitz N. 1996. Toxicity of Azathioprine and Monocrotaline in Murine Sinusoidal Endothelial Cells and Hepatocytes: The Role of Glutathione and Relevance to Hepatic Venoocclusive Disease. Hepatology 23:589-99Dong H, Haining RL, Thummel KE, Rettie AE, Nelson SD. 2000. Involvement of human cytochrome P450 2D6 in the bioactivation of acetaminophen. Drug Metab Dispos 28:1397-400Doostdar H, Grant MH, Melvin WT, Wolf CR, Burke MD. 1993. The effects of inducing agents on cytochrome P450 and UDP-glucuronyltransferase activities in human HEPG2 hepatoma cells. Biochemical pharmacology 46:629-35EFSA. 2011. Scientific Opinion on Pyrrolizidine alkaloids in food and feed. EFSA Journal 9:1-134Ekins S, Williams AJ, Xu JJ. 2010. A predictive ligand-based Bayesian model for human drug-induced liver injury. Drug Metab. Dispos. 38:2302-8EMA. 2014. EMA/HMPC/893108/2011: Public statement on the use of herbal medicinal products containing toxic, unsaturated pyrrolizidine alkaloids (PAs).1-24EMA. 2016. EMA/HMPC/328782/2016: Public statement on contamination of herbal medicinal products/traditional herbal medicinal products with pyrrolizidine alkaloids.1-11Fashe MM, Juvonen RO, Petsalo A, Vepsalainen J, Pasanen M, Rahnasto-Rilla M. 2015. In silico prediction of the site of oxidation by cytochrome P450 3A4 that leads to the formation of the toxic metabolites of pyrrolizidine alkaloids. Chem Res Toxicol 28:702-10Field RA, Stegelmeier BL, Colegate SM, Brown AW, Green BT. 2015. An in vitro comparison of the cytotoxic potential of selected dehydropyrrolizidine alkaloids and some N-oxides. Toxicon 97:36-45Fleming I. 2014. The pharmacology of the cytochrome P450 epoxygenase/soluble epoxide hydrolase axis in the vasculature and cardiovascular disease. Pharmacol Rev 66:1106-40Fonti V. 2017. *Feature Selection using LASSO*. Research paper. VU Amsterdam. 26 pp.Fu PP, Chou MW, Churchwell M, Wang Y, Zhao Y, et al. 2010. High-Performance Liquid Chromatography Electrospray Ionization Tandem Mass Spectrometry for the Detection and Quantitation of Pyrrolizidine Alkaloid-Derived DNA Adducts in Vitro and in Vivo. Chem Res Toxicol 23:637--52Fu PP, Xia Q, Lin G, Chou MW. 2004. Pyrrolizidine alkaloids\--genotoxicity, metabolism enzymes, metabolic activation, and mechanisms. Drug Metab Rev 36:1-55Galeotti N, Vivoli E, Bilia AR, Vincieri FF, Ghelardini C. 2010. St. John\'s wort reduces neuropathic pain through a hypericin-mediated inhibition of the protein kinase Cgamma and epsilon activity. Biochem Pharmacol 79:1327-36Ganesan S, Tekwani BL, Sahu R, Tripathi LM, Walker LA. 2009. Cytochrome P(450)-dependent toxic effects of primaquine on human erythrocytes. Toxicol Appl Pharmacol 241:14-22Gao H, Ruan JQ, Chen J, Li N, Ke CQ, et al. 2015. Blood pyrrole-protein adducts as a diagnostic and prognostic index in pyrrolizidine alkaloid-hepatic sinusoidal obstruction syndrome. Drug Des Devel Ther 9:4861-8Gitlin N. 1980. Salicylate hepatotoxicity: the potential role of hypoalbuminemia. J Clin Gastroenterol 2:281-5Gordon GJ, Coleman WB, Grisham JW. 2000. Bax-mediated apoptosis in the livers of rats after partial hepatectomy in the retrorsine model of hepatocellular injury. Hepatology 32:312-20Gradhand U, Lang T, Schaeffeler E, Glaeser H, Tegude H, et al. 2008. Variability in human hepatic MRP4 expression: influence of cholestasis and genotype. Pharmacogenomics J 8:42-52Gramatica P, Corradi M, Consonni V. 2000. Modelling and prediction of soil sorption coefficients of non-ionic organic pesticides by molecular descriptors. Chemosphere 41:763-77Greene N, Fisk L, Naven RT, Note RR, Patel ML, Pelletier DJ. 2010. Developing structure-activity relationships for the prediction of hepatotoxicity. Chemical Research in Toxicology 23:1215-22Guo YX, Xu XF, Zhang QZ, Li C, Deng Y, et al. 2015. The inhibition of hepatic bile acids transporters Ntcp and Bsep is involved in the pathogenesis of isoniazid/rifampicin-induced hepatotoxicity. Toxicology mechanisms and methods 25:382-7Hall LH, Kier LB. 1995. Electrotopological State Indices for Atom Types: A Novel Combination of Electronic, Topological, and Valence State Information. Journal of Chemical Information and Computer Sciences 35:1039-45Hammann F, Schoning V, Drewe J. 2018. Prediction of clinically relevant drug-induced liver injury from structure using machine learning. J Appl Toxicol Hansen K, Mika S, Schroeter T, Sutter A, ter Laak A, et al. 2009. Benchmark data set for in silico prediction of Ames mutagenicity. J Chem Inf Model 49:2077-81Hartmann T, Ehmke A, Eilert U, yon Borstel K, Thcuring C. 1989. Sites of synthesis, translocation and accumulation of pyrrolizidine alkaloid N-oxides in Senecio vulgaris L. Planta 177:98-107Hartmann T, Witte L. 1995. Chemistry, Biology and Chemoecology of the Pyrrolizidine Alkaloids. In *Alkaloids: Chemical and Biological Perspectives*, ed. Pelletier, pp. 155-233. Pergamon, London, New YorkHessel S, Gottschalk C, Schumann D, These A, Preiss-Weigert A, Lampen A. 2014. Structure-activity relationship in the passage of different pyrrolizidine alkaloids through the gastrointestinal barrier: ABCB1 excretes heliotrine and echimidine. Mol Nutr Food Res 58:995-1004Hunt CM, Westerkam WR, Stave GM. 1992. Effect of age and gender on the activity of human hepatic CYP3A. Biochemical pharmacology 44:275-83Ibanez L, Perez E, Vidal X, Laporte JR, Grup d\'Estudi Multicenteric d\'Hepatotoxicitat Aguda de B. 2002. Prospective surveillance of acute serious liver disease unrelated to infectious, obstructive, or metabolic diseases: epidemiological and clinical features, and exposure to drugs. J Hepatol 37:592-600ICH. 2011. Guideance on genotoxicity testing and data interpretation for pharmaceuticals intended for human use S2(R1). p. 29Iyer VV, Yang H, Ierapetritou MG, Roth CM. 2010. Effects of glucose and insulin on HepG2-C3A cell metabolism. Biotechnol Bioeng 107:347-56Jago MV. 1971. Factors affecting the chronic hepatotoxicity of pyrrolizidine alkaloids. The Journal of Pathology 105:1-11Jeon JY, Sparreboom A, Baker SD. 2017. Kinase Inhibitors: The Reality Behind the Success. Clin Pharmacol Ther 102:726-30Jeong W, Doroshow JH, Kummar S. 2013. United States Food and Drug Administration approved oral kinase inhibitors for the treatment of malignancies. Curr Probl Cancer 37:110-44Ji L, Chen Y, Liu T, Wang Z. 2008. Involvement of Bcl-xL degradation and mitochondrial-mediated apoptotic pathway in pyrrolizidine alkaloids-induced apoptosis in hepatocytes. Toxicol Appl Pharmacol 231:393-400Jornil J, Nielsen TS, Rosendal I, Ahlner J, Zackrisson AL, et al. 2013. A poor metabolizer of both CYP2C19 and CYP2D6 identified by mechanistic pharmacokinetic simulation in a fatal drug poisoning case involving venlafaxine. Forensic Sci Int 226:e26-31Kalthoff S, Ehmer U, Freiberg N, Manns MP, Strassburg CP. 2010. Interaction between oxidative stress sensor Nrf2 and xenobiotic-activated aryl hydrocarbon receptor in the regulation of the human phase II detoxifying UDP-glucuronosyltransferase 1A10. J Biol Chem 285:5993-6002Kazius J, McGuire R, Bursi R. 2005. Derivation and validation of toxicophores for mutagenicity prediction. J Med Chem 48:312-20Khan D, Khan AU. 2016. Descriptors and their selection methods in QSAR analysis: paradigm for drug design. Drug Discov Today 21:1291-302Kim HY, Stermitz FR, Molyneux RJ, Wilson DW, Taylor D, Coulombe RA, Jr. 1993. Structural influences on pyrrolizidine alkaloid-induced cytopathology. Toxicol Appl Pharmacol 122:61-9Kock K, Ferslew BC, Netterberg I, Yang K, Urban TJ, et al. 2014. Risk factors for development of cholestatic drug-induced liver injury: inhibition of hepatic basolateral bile acid transporters multidrug resistance-associated proteins 3 and 4. Drug Metab Dispos 42:665-74Lammert C, Einarsson S, Saha C, Niklasson A, Bjornsson E, Chalasani N. 2008. Relationship between daily dose of oral medications and idiosyncratic drug-induced liver injury: search for signals. Hepatology 47:2003-9Langel D, Ober D, Pelser PB. 2011. The evolution of pyrrolizidine alkaloid biosynthesis and diversity in the Senecioneae. Phytochemistry Reviews 10:3-74Lasser KE, Allen PD, Woolhandler SJ, Himmelstein DU, Wolfe SM, Bor DH. 2002. Timing of new black box warnings and withdrawals for prescription medications. JAMA 287:2215-20Li N, Xia Q, Ruan J, Fu PP, Lin G. 2011. Hepatotoxicity and Tumorigenicity Induced by Metabolic Activation of Pyrrolizidine Alkaloids in Herbs. Current Drug Metabolism 12Li X, Cameron MD. 2012. Potential role of a quetiapine metabolite in quetiapine-induced neutropenia and agranulocytosis. Chem Res Toxicol 25:1004-11Li YH, Kan WL, Li N, Lin G. 2013. Assessment of pyrrolizidine alkaloid-induced toxicity in an in vitro screening model. J Ethnopharmacol 150:560-7Lima A, Bernardes M, Azevedo R, Medeiros R, Seabra V. 2015. Pharmacogenomics of Methotrexate Membrane Transport Pathway: Can Clinical Response to Methotrexate in Rheumatoid Arthritis Be Predicted? Int J Mol Sci 16:13760-80Lin G. 1998. Microsomal Formation of a Pyrrolic Alcohol Glutathione Conjugate of ClivorineFirm Evidence for the Formation of a Pyrrolic Metabolite of an Otonecine-Type Pyrrolizidine Alkaloid. Drug Metabolism and Disposition 26:181-4Lindigkeit R, Biller A, Buch M, Schiebel H-M, Boppré M, Hartmann T. 1997. The two faces of pyrrolizidine alkaloids: the role of the tertiary amine and its N-oxide in chemical defense of insects with acquired plant alkaloids. Eur J Biochem 245Makhlouf HA, Helmy A, Fawzy E, El-Attar M, Rashed HA. 2008. A prospective study of antituberculous drug-induced hepatotoxicity in an area endemic for liver diseases. Hepatol Int 2:353-60Marin-Hernandez A, Rodriguez-Enriquez S, Vital-Gonzalez PA, Flores-Rodriguez FL, Macias-Silva M, et al. 2006. Determining and understanding the control of glycolysis in fast-growth tumor cells. Flux control by an over-expressed but strongly product-inhibited hexokinase. FEBS J 273:1975-88Marroquin LD, Hynes J, Dykens JA, Jamieson JD, Will Y. 2007. Circumventing the Crabtree effect: replacing media glucose with galactose increases susceptibility of HepG2 cells to mitochondrial toxicants. Toxicol Sci 97:539-47Mattocks AR. 1986. *Chemistry and Toxicology of Pyrrolizidine Alkaloids*: Academic PressMeharena HS, Chang P, Keshwani MM, Oruganty K, Nene AK, et al. 2013. Deciphering the structural basis of eukaryotic protein kinase regulation. PLoS Biol 11:e1001680Merz KH, Schrenk D. 2016. Interim relative potency factors for the toxicological risk assessment of pyrrolizidine alkaloids in food and herbal medicines. Toxicol Lett 263:44-57Miners JO, Birkett DJ. 1998. Cytochrome P4502C9: an enzyme of major importance in human drug metabolism. British Journal of Clinical Pharmacology 45:525-38Mingard C, Paech F, Bouitbir J, Krahenbuhl S. 2018. Mechanisms of toxicity associated with six tyrosine kinase inhibitors in human hepatocyte cell lines. J Appl Toxicol 38:418-31Mingatto FE, Dorta DJ, dos Santos AB, Carvalho I, da Silva CH, et al. 2007. Dehydromonocrotaline inhibits mitochondrial complex I. A potential mechanism accounting for hepatotoxicity of monocrotaline. Toxicon 50:724-30Mitchell JB. 2014. Machine learning methods in chemoinformatics. Wiley Interdiscip Rev Comput Mol Sci 4:468-81Morgan RE, Trauner M, van Staden CJ, Lee PH, Ramachandran B, et al. 2010. Interference with bile salt export pump function is a susceptibility factor for human liver injury in drug development. Toxicol Sci 118:485-500Muegge I, Mukherjee P. 2016. An overview of molecular fingerprint similarity search in virtual screening. Expert Opin Drug Discov 11:137-48Najibi A, Heidari R, Zarifi J, Jamshidzadeh A, Firoozabadi N, Niknahad H. 2016. Evaluating the Role of Drug Metabolism and Reactive Intermediates in Trazodone-Induced Cytotoxicity toward Freshly-Isolated Rat Hepatocytes. Drug Res (Stuttg) 66:592-6Nantasenamat C, Isarankura-Na-Ayudhya C, Naenna T, Prachayasittikul V. 2009. A Practical Overview of Quantitative Structure-Activity Relationship. EXCLI Journal 8:74-88National Cancer Institute. 2006. Common Terminology Criteria for Adverse Events v3.0 (CTCAE). ed. Program, CTENeumann MG, Cohen LB, Opris M, Nanau R, Jeong H. 2015. Hepatotoxicity of Pyrrolizidine Alkaloids. J Pharm Pharm Sci 18:825-43Newby D, Freitas AA, Ghafourian T. 2015. Decision trees to characterise the roles of permeability and solubility on the prediction of oral absorption. Eur J Med Chem 90:751-65Niederer C, Behra R, Harder A, Schwarzenbach RP, Escher BI. 2004. Mechanistic approaches for evaluating the toxicity of reactive organochlorines and epoxides in green algae. Environmental Toxicology and Chemistry 23:697-704NTP. 1978. Bioassay of lasiocarpine for possible carcinogenicity. pp. 1-82NTP. 2003. Toxicology and Carcinogenesis Studies of Riddelliine (CAS No. 23246-96-0) in F344/N Rats And B6c3F~1~ Mice (Gavage Studies). ed. Health, NIoO\'Boyle NM, Banck M, James CA, Morley C, Vandermeersch T, Hutchison GR. 2011. Open Babel: An open chemical toolbox. J Cheminform 3:33Open Babel community. 2011. *Molecular fingerprints and similarity searching --- Open Babel v2.3.1 documentation. Openbabel.org*. , Dececmber 31, 2018Paech F, Bouitbir J, Krahenbuhl S. 2017. Hepatocellular Toxicity Associated with Tyrosine Kinase Inhibitors: Mitochondrial Damage and Inhibition of Glycolysis. Front Pharmacol 8:367Parkinson A, Mudra DR, Johnson C, Dwyer A, Carroll KM. 2004. The effects of gender, age, ethnicity, and liver cirrhosis on cytochrome P450 enzyme activity in human liver microsomes and inducibility in cultured human hepatocytes. Toxicol Appl Pharmacol 199:193-209Pellinen P, Honkakoski P, Stenback F, Niemitz M, Alhava E, et al. 1994. Cocaine N-demethylation and the metabolism-related hepatotoxicity can be prevented by cytochrome P450 3A inhibitors. Eur J Pharmacol 270:35-43Regev A, Seeff LB, Merz M, Ormarsdottir S, Aithal GP, et al. 2014. Causality assessment for suspected DILI during clinical phases of drug development. Drug Saf 37 Suppl 1:S47-56Rendic S. 2002. Summary of information on human CYP enzymes: human P450 metabolism data. Drug Metab Rev 34:83-448Reuben A, Koch DG, Lee WM, Acute Liver Failure Study G. 2010. Drug-induced acute liver failure: results of a U.S. multicenter, prospective study. Hepatology 52:2065-76Rodrigues AC. 2010. Efflux and uptake transporters as determinants of statin response. Expert Opin Drug Metab Toxicol 6:621-32Roskoski R, Jr. 2015. A historical overview of protein kinases and their targeted small molecule inhibitors. Pharmacol Res 100:1-23Ruan J, Liao C, Ye Y, Lin G. 2014a. Lack of metabolic activation and predominant formation of an excreted metabolite of nontoxic platynecine-type pyrrolizidine alkaloids. Chem Res Toxicol 27:7-16Ruan J, Yang M, Fu P, Ye Y, Lin G. 2014b. Metabolic activation of pyrrolizidine alkaloids: insights into the structural and enzymatic basis. Chem Res Toxicol 27:1030-9Rubiolo P, Pieters L, Calomme M, Bicchi C, Vlietinck A, Vanden Berghe D. 1992. Mutagenicity of pyrrolizidine alkaloids in the Salmonella typhimurium/mammalian microsome system. Mutat Res 281:143-7Rücker C, Rücker G, Meringer M. 2007. y-Randomization and Its Variants in QSPR/QSAR. J. Chem. Inf. Model. 47:2345-57Schoental R, Head MA. 1957. Progression of liver lesions produced in rats by temporary treatment with pyrrolizidine (senecio) alkaloids, and the effects of betaine and high casein diet. Br J Cancer 11:535-44Schöning V, Hammann F, Peinl M, Drewe J. 2017. Editor\'s Highlight: Identification of Any Structure-Specific Hepatotoxic Potential of Different Pyrrolizidine Alkaloids Using Random Forests and Artificial Neural Networks. Toxicol Sci 160:361-70Shah RR, Morganroth J, Shah DR. 2013. Hepatotoxicity of tyrosine kinase inhibitors: clinical and regulatory perspectives. Drug Saf 36:491-503Spjuth O, Alvarsson J, Berg A, Eklund M, Kuhn S, et al. 2009. Bioclipse 2: A scriptable integration platform for the life sciences. BMC Bioinformatics 10:1-5Spjuth O, Helmus T, Willighagen EL, Kuhn S, Eklund M, et al. 2007. Bioclipse: an open source workbench for chemo- and bioinformatics. BMC Bioinformatics 8:1-10Srinivas N, Sandeep KS, Anusha Y, Devendra BN. 2014. In Vitro Cytotoxic Evaluation and Detoxification of Monocrotaline (Mct) Alkaloid: An In Silico Approach. International Invention Journal of Biochemistry and Bioinformatics 2:20-9Stine JG, Chalasani NP. 2017. Drug Hepatotoxicity: Environmental Factors. Clin Liver Dis 21:103-13Stine JG, Lewis JH. 2011. Drug-induced liver injury: a summary of recent advances. Expert Opin Drug Metab Toxicol 7:875-90Takanashi H, Umeda M, Hirono I. 1980. Chromosomal aberrations and mutations in cultured mammalidan cells induced by pyrrolizidine alkaloids. Mutation Research 78:67-77Takeda M, Okamoto I, Nakagawa K. 2015. Pooled safety analysis of EGFR-TKI treatment for EGFR mutation-positive non-small cell lung cancer. Lung Cancer 88:74-9Tamta H, Pawar RS, Wamer WG, Grundel E, Krynitsky AJ, Rader JI. 2012. Comparison of metabolism-mediated effects of pyrrolizidine alkaloids in a HepG2/C3A cell-S9 co-incubation system and quantification of their glutathione conjugates. Xenobiotica 42:1038-48Teh LK, Bertilsson L. 2012. Pharmacogenomics of CYP2D6: molecular genetics, interethnic differences and clinical importance. Drug Metab Pharmacokinet 27:55-67Teo YL, Ho HK, Chan A. 2013. Risk of tyrosine kinase inhibitors-induced hepatotoxicity in cancer patients: a meta-analysis. Cancer Treat Rev 39:199-206Teo YL, Ho HK, Chan A. 2015. Formation of reactive metabolites and management of tyrosine kinase inhibitor-induced hepatotoxicity: a literature review. Expert Opin Drug Metab Toxicol 11:231-42Thompson RA, Isin EM, Ogese MO, Mettetal JT, Williams DP. 2016. Reactive Metabolites: Current and Emerging Risk and Hazard Assessments. Chem Res Toxicol 29:505-33Walker K, Ginsberg G, Hattis D, Johns DO, Guyton KZ, Sonawane B. 2009. Genetic polymorphism in N-Acetyltransferase (NAT): Population distribution of NAT1 and NAT2 activity. Journal of toxicology and environmental health. Part B, Critical reviews 12:440-72Wang YP, Yan J, Fu PP, Chou MW. 2005. Human liver microsomal reduction of pyrrolizidine alkaloid N-oxides to form the corresponding carcinogenic parent alkaloid. Toxicol Lett 155:411-20Weininger D. 1988. SMILES, a chemical language and information system. 1. Introduction to methodology and encoding rules. J Chem Inf Comput Sci 28:31-6Westerink WM, Schoonen WG. 2007. Phase II enzyme levels in HepG2 cells and cryopreserved primary human hepatocytes and their induction in HepG2 cells. Toxicol In Vitro 21:1592-602Wu P, Nielsen TE, Clausen MH. 2015. FDA-approved small-molecule kinase inhibitors. Trends Pharmacol Sci 36:422-39Xia Q, Ma L, He X, Cai L, Fu PP. 2015. 7-glutathione pyrrole adduct: a potential DNA reactive metabolite of pyrrolizidine alkaloids. Chem Res Toxicol 28:615-20Xia Q, Zhao Y, Von Tungeln LS, Doerge DR, Lin G, et al. 2013. Pyrrolizidine alkaloid-derived DNA adducts as a common biological biomarker of pyrrolizidine alkaloid-induced tumorigenicity. Chem Res Toxicol 26:1384-96Yan J, Xia Q, Chou MW, Fu P. 2008. Metabolic activation of retronecine and retronecine N-oxide -- formation of DHP-derived DNA adducts. Toxicology and Industrial Health 24Yang X, Li W, Sun Y, Guo X, Huang W, et al. 2017. Comparative Study of Hepatotoxicity of Pyrrolizidine Alkaloids Retrorsine and Monocrotaline. Chem Res Toxicol 30:532-9Yap CW. 2011. PaDEL-descriptor: an open source software to calculate molecular descriptors and fingerprints. Journal of computational chemistry 32:1466-74Yap CW. 2014. *Descriptors*. , 27.10.2016Yu K, Geng X, Chen M, Zhang J, Wang B, et al. 2014a. High daily dose and being a substrate of cytochrome P450 enzymes are two important predictors of drug-induced liver injury. Drug Metab Dispos 42:744-50Yu K, Geng X, Chen M, Zhang J, Wang B, et al. 2014b. High daily dose and being a substrate of cytochrome P450 enzymes are two important predictors of drug-induced liver injury. Drug Metab. Dispos. 42:744-50Zanger UM, Turpeinen M, Klein K, Schwab M. 2008. Functional pharmacogenetics/genomics of human cytochromes P450 involved in drug biotransformation. Anal Bioanal Chem 392:1093-108Zhang J, Sheng Y, Shi L, Zheng Z, Chen M, et al. 2017. Quercetin and baicalein suppress monocrotaline-induced hepatic sinusoidal obstruction syndrome in rats. Eur J Pharmacol 795:160-8Zhao Y, Xia Q, Gamboa da Costa G, Yu H, Cai L, Fu PP. 2012. Full structure assignments of pyrrolizidine alkaloid DNA adducts and mechanism of tumor initiation. Chem Res Toxicol 25:1985-96Zheng Z, Shi L, Sheng Y, Zhang J, Lu B, Ji L. 2016. Chlorogenic acid suppresses monocrotaline-induced sinusoidal obstruction syndrome: The potential contribution of NFkappaB, Egr1, Nrf2, MAPKs and PI3K signals. Environ Toxicol Pharmacol 46:80-9Zhu XW, Xin YJ, Ge HL. 2015. Recursive Random Forests Enable Better Predictive Performance and Model Interpretation than Variable Selection by LASSO. J Chem Inf Model 55:736-46