Collaborative Profile-QSAR: A Natural Platform for Building Collaborative Models among Competing Companies.

Massively multitask bioactivity models that transfer learning between thousands of assays have been shown to work dramatically better than separate models trained on each individual assay. In particular, the applicability domain for a given model can expand from compounds similar to those tested in that specific assay to those tested across the full complement of contributing assays. If many large companies would share their assay data and train models on the superset, predictions should be better than what each company can do alone. However, a company's compounds, targets, and activities are among their most guarded trade secrets. Strategies have been proposed to share just the individual collaborators' models, without exposing any of the training data. Profile-QSAR (pQSAR) is a two-level, multitask, stacked model. It uses profiles of level-1 predictions from single-task models for thousands of assays as compound descriptors for level-2 models. This work describes its simple and natural adaptation to safe collaboration by model sharing. Broad model sharing has not yet been implemented across multiple large companies, so there are numerous unanswered questions. Novartis was formed from several mergers and acquisitions. In principle, this should allow an internal simulation of model sharing. In practice, the lack of metadata about the origins of compounds and assays made this difficult. Nevertheless, we have attempted to simulate this process and propose some findings: multitask pQSAR is always an improvement over single-task models; collaborative multitask modeling did not improve predictions on internal compounds; collaboration did improve predictions for external compounds but far less than the purely internal multitask modeling for internal compounds; collaborative models for external compounds increasingly improve as overlap between compound collections increases; combining profiles from inside and outside the company is not best, with internal predictions better using only the inside profile and external using only the outside profile, but a consensus of models using all three profiles is best on external compounds and a good compromise on internal compounds. We anticipate similar results from other model-sharing approaches. Indeed, since collaborative pQSAR through model sharing is mathematically identical to pQSAR using actual shared data, we believe our conclusions should apply to collaborative modeling by any current method even including the unlikely scenario of directly sharing all chemical structures and assay data.

All-Assay-Max2 pQSAR: Activity Predictions as Accurate as Four-Concentration IC50s for 8558 Novartis Assays.

Profile-quantitative structure-activity relationship (pQSAR) is a massively multitask, two-step machine learning method with unprecedented scope, accuracy, and applicability domain. In step one, a "profile" of conventional single-assay random forest regression models are trained on a very large number of biochemical and cellular pIC50 assays using Morgan 2 substructural fingerprints as compound descriptors. In step two, a panel of partial least squares (PLS) models are built using the profile of pIC50 predictions from those random forest regression models as compound descriptors (hence the name). Previously described for a panel of 728 biochemical and cellular kinase assays, we have now built an enormous pQSAR from 11 805 diverse Novartis (NVS) IC50 and EC50 assays. This large number of assays, and hence of compound descriptors for PLS, dictated reducing the profile by only including random forest regression models whose predictions correlate with the assay being modeled. The random forest regression and pQSAR models were evaluated with our "realistically novel" held-out test set, whose median average similarity to the nearest training set member across the 11 805 assays was only 0.34, comparable to the novelty of compounds actually selected from virtual screens. For the 11 805 single-assay random forest regression models, the median correlation of prediction with the experiment was only rext2 = 0.05, virtually random, and only 8% of the models achieved our standard success threshold of rext2 = 0.30. For pQSAR, the median correlation was rext2 = 0.53, comparable to four-concentration experimental IC50s, and 72% of the models met our rext2 > 0.30 standard, totaling 8558 successful models. The successful models included assays from all of the 51 annotated target subclasses, as well as 4196 phenotypic assays, indicating that pQSAR can be applied to virtually any disease area. Every month, all models are updated to include new measurements, and predictions are made for 5.5 million NVS compounds, totaling 50 billion predictions. Common uses have included virtual screening, selectivity design, toxicity and promiscuity prediction, mechanism-of-action prediction, and others. Several such actual applications are described.

In Silico Prediction of Drug-Induced Liver Injury Based on Adverse Drug Reaction Reports.

Drug-induced liver injury (DILI) is a major cause of drug attrition. Currently existing Quantitative Structure-Activity Relationship models have limited predictive capabilities for DILI. Furthermore, their practical applications were limited by lack of new hepatotoxicity data. In this study, we first collected and curated a novel set of 122 DILI-positive and 932 DILI-negative drugs from online adverse drug reports using proportional reporting ratios as the signal detection method. Second, three strategies (under-sampling the majority class, synthetic minority over-sampling technique, and adjusting decision threshold approach) were employed to develop predictive classification models to cope with the unbalanced dataset. Random forest (RF) models using CDK, MACCS, and Mold2 descriptors based on the under-sampling and over-sampling strategies afforded correct classification ratio (CCR) of ∼0.77 and 0.78, respectively. Recursive RF models based on the last strategy tremendously reduced modeling descriptors (at most 95.4% for Mold2) while apparently improved the predictability with a consensus CCR of 0.84 (sensitivity of 0.88 and specificity of 0.79). Structural analysis showed that pyrimidine derivatives, purine derivatives, and halogenated hydrocarbon were critical for drugs' hepatotoxicity. The reporting frequency of many drugs was gender-dependent (eg, antiviral and anti-cancer drugs for males and antibacterial drugs for females) as well as age-dependent (eg, antiviral and anti-cancer drugs for the middle age group of 20-29, 30-39, and 40-49). Approximately 84% of total cases were reported during the first 6 months of administration. The curated hepatotoxicity dataset along with the predictive classification models presented here should provide insight into future studies of DILI.

[Pollution Characteristics and Source Analysis of Polycyclic Aromatic Hydrocarbons in Agricultural Soils from Shandong].

Polycyclic aromatic hydrocarbons (PAHs) are ubiquitous environmental contaminants that originate mainly from anthropogenic sources. PAHs have elicited much concern because they exhibit strong toxic, carcinogenic, and mutagenic properties. Agricultural soil is at risk of PAH contamination mainly caused by atmospheric depositions, wastewater irrigation, or organic substances and biowaste applied as fertilizers. The surface agricultural soils were collected from Shandong in July 2015, and measured for 16 US EPA priority PAHs using high performance liquid chromatography with UV and fluorescence detector. The content and composition of PAHs were analyzed. The differences of PAHs between soils from the field for growing crops and from vegetable greenhouses, and between soils from point sources and from non-point sources were compared. The sources of PAHs were determined with methods of ratio between PAHs and positive matrix factorization (PMF), and the risks of PAHs were assessed. The results showed that the total content of 16 PAHs (∑16PAHs) ranged from 111.5 ng·g-1 to 2744.1 ng·g-1, with the mean of 556.3 ng·g-1. The content of 3-ring PAHs was relatively high, with the mean of 201.5 ng·g-1; while the contents of 2-ring and 6-ring PAHs were relatively low, with the mean of 39.3 ng·g-1 and 43.4 ng·g-1, respectively. According to the contamination classification in Poland, 71% of the samples in Shangdong were weakly contaminated. Compared with other areas in China, the content of PAHs in the agricultural soils in Shandong was in the middle range. Acenaphthene, fluorine, and fluoranthene were the major PAH compounds, accounting for more than 10% of the total PAHs; while the contribution of indeno (1,2,3-cd) pyrene was low. The content of ∑16PAHs and contribution of 7 carcinogenic PAHs were significantly higher in soils polluted by point sources than those in soils from non-point sources. Moreover, the contribution of PAHs with 2-3 rings was significantly higher in soils from non-point sources, while the contribution of PAHs with 4-6 rings was significantly higher in soils polluted by point sources. There was no significant difference in soils from vegetable greenhouses and from adjacent field soils, and the contribution of PAHs with 3-4 rings was high. The PAH isomer pair ratios of Ant/(Ant+Phe), Flu/(Flu+Pyr), BaA/(BaA+Chr), and InP/(InP+BP) were utilized as molecular indices to elucidate the possible PAH sources, and the results suggested that the PAHs in the soils were mainly from combustion. To quantitatively assess the contribution of various sources to PAH contamination, PMF was used to analyze the sources. The sources of PAHs were combustion of coal biomass, oil combustion from traffic, coking, and petroleum pollution, with contribution of 42.7%, 19.3%, 22.8% and 15.2%, respectively. Toxic equivalency factors were used to evaluate PAH contamination in the soils, and the carcinogenicity of other PAHs relative to BaP was quantified to estimate the BaP-equivalent concentration (TEQBaP). The TEQBaPof 16 PAHs (∑16TEQBap) in soils from non-point sources and vegetable greenhouses was 31.69 and 44.47 ng·g-1, respectively, which were below the safe value in Canadian soil quality guidelines. However, the ∑16TEQBap in some field soils from point sources exceeded the safe value, indicating that there were potential risks in the soils from point sources in Shandong.

Mixture cytotoxicity assessment of ionic liquids and heavy metals in MCF-7 cells using mixtox.

Ionic liquids (ILs) are widely used as extractants for heavy metals. However, the effect of mixtures of ILs and heavy metals is rarely understood. In this study, we tested the cytotoxicity of four ILs, four heavy metals and their mixtures on human MCF-7 cells in 96-well microplates. The toxicity of single compounds in MCF-7 cells ranges from 3.07 × 10(-6) M for Cu(II) to 2.20 × 10(-3) M for 1-ethyl-3-methylimidazolium tetrafluoroborate. The toxicity of heavy metals in MCF-7 is generally higher than the toxicity of ILs. A uniform experimental design was used to simulate environmentally realistic mixtures. Two classical reference models (concentration addition and independent action) were used to predict their mixture. The experiments to evaluate the toxicity of the mixture revealed antagonism among four ILs and four heavy metals in MCF-7 cells. Pearson correlation analysis showed that Ni(II) and 1-dodecyl-3-methylimidazolium chloride are positively correlated with the extent of antagonism, while 1-hexyl-3-methylimidazolium tetrafluoroborate showed a negative correlation. Data analysis was conducted in the R package mixtox, which integrates features such as curve fitting, experimental design, and mixture toxicity prediction. The international community of toxicologists is welcome to use this package and provide feedback as suggestions and comments.

Recursive Random Forests Enable Better Predictive Performance and Model Interpretation than Variable Selection by LASSO.

Variable selection is of crucial significance in QSAR modeling since it increases the model predictive ability and reduces noise. The selection of the right variables is far more complicated than the development of predictive models. In this study, eight continuous and categorical data sets were employed to explore the applicability of two distinct variable selection methods random forests (RF) and least absolute shrinkage and selection operator (LASSO). Variable selection was performed: (1) by using recursive random forests to rule out a quarter of the least important descriptors at each iteration and (2) by using LASSO modeling with 10-fold inner cross-validation to tune its penalty λ for each data set. Along with regular statistical parameters of model performance, we proposed the highest pairwise correlation rate, average pairwise Pearson's correlation coefficient, and Tanimoto coefficient to evaluate the optimal by RF and LASSO in an extensive way. Results showed that variable selection could allow a tremendous reduction of noisy descriptors (at most 96% with RF method in this study) and apparently enhance model's predictive performance as well. Furthermore, random forests showed property of gathering important predictors without restricting their pairwise correlation, which is contrary to LASSO. The mutual exclusion of highly correlated variables in LASSO modeling tends to skip important variables that are highly related to response endpoints and thus undermine the model's predictive performance. The optimal variables selected by RF share low similarity with those by LASSO (e.g., the Tanimoto coefficients were smaller than 0.20 in seven out of eight data sets). We found that the differences between RF and LASSO predictive performances mainly resulted from the variables selected by different strategies rather than the learning algorithms. Our study showed that the right selection of variables is more important than the learning algorithm for modeling. We hope that a standard procedure could be developed based on these proposed statistical metrics to select the truly important variables for model interpretation, as well as for further use to facilitate drug discovery and environmental toxicity assessment.

Benefits from hazards: mixture hormesis induced by [emim]Cl despite its individual inhibitions.

The threshold model based on monotonic concentration-response curves (CRCs) is unsuitable to assess the risk of chemicals with non-monotonic CRCs. The non-monotonic CRCs of mixtures may relate to the characteristics of some individual component. To reveal the cause of the mixtures resulting in the non-monotonic CRCs, we used the microplate toxicity analysis to determine the toxicity effects of six 1-alkyl-3-methyl-imidazolium ([amim]X) salts and their mixtures on Vibrio qinghaiensis sp.-Q67 (Q67). It was shown that the CRCs of six [amim]X salts are monotonic S-shaped while those of the senary mixtures designed by the uniform design ray (UD-ray) are all non-monotonic. The mixtures were further split into two ternary mixtures, one containing 1-ethyl-3-methyl-imidazolium ([emim]X) salts (noted as UTE) and the other one containing 1-butyl-3-methyl-imidazolium ([bmim]X) salts (noted as UTB). It was found that the CRCs of UTE mixtures are all non-monotonically J-shaped, while only one (UTB-R3) among UTB mixtures has a little stimulating effect and the CRCs of the other three mixtures (UTB-R1, UTB-R2 and UTB-R4) are monotonic. The CRCs of the binary mixtures designed by the direct equipartition ray design (EquRay) procedure were further examined. The CRCs of the mixtures containing [emim]Cl are non-monotonic J-shaped while those of the mixtures without [emim]Cl are still monotonic. Thus, it can conclude that it is [emim]Cl that causes the non-monotonic CRCs in [amim]X mixtures, even though the CRC of individual [emim]Cl is monotonic.

Two-stage prediction of the effects of imidazolium and pyridinium ionic liquid mixtures on luciferase.

The predicted toxicity of mixtures of imidazolium and pyridinium ionic liquids (ILs) in the ratios of their EC50, EC10, and NOEC (no observed effect concentration) were compared to the observed toxicity of these mixtures on luciferase. The toxicities of EC50 ratio mixture can be effectively predicted by two-stage prediction (TSP) method, but were overestimated by the concentration addition (CA) model and underestimated by the independent action (IA) model. The toxicities of EC10 ratio mixtures can be basically predicted by TSP and CA, but were underestimated by IA. The toxicities of NOEC ratio mixtures can be predicted by TSP and CA in a certain concentration range, but were underestimated by IA. Our results support the use of TSP as a default approach for predicting the combined effect of different types of ILs at the molecular level. In addition, mixtures of ILs mixed at NOEC and EC10 could cause significant effects of 64.1% and 97.7%, respectively. Therefore, we should pay high attention to the combined effects in mixture risk assessment.

Hybrid in silico models for drug-induced liver injury using chemical descriptors and in vitro cell-imaging information.

Drug-induced liver injury (DILI) is a major adverse drug reaction that accounts for one-third of post-marketing drug withdrawals. Several classifiers for human hepatotoxicity using chemical descriptors with limited prediction accuracies have been published. In this study, we developed predictive in silico models based on a set of 156 DILI positive and 136 DILI negative compounds for DILI prediction. First, models based on a chemical descriptor (CDK, Dragon and MOE) and in vitro cell-imaging endpoints [human hepatocyte imaging assay technology (HIAT) descriptors] were built using random forest (RF) and five-fold cross-validation procedure. Then three hybrid models were built using HIAT and a single type of chemical descriptors. Generally, the models based only on chemical descriptors were poor, with a correct classification rate (CCR) around 0.60 when the default threshold value (i.e. threshold = 0.50) was used. The hybrid models afforded a CCR of 0.73 with a specificity of 0.74 and a better true positive rate (sensitivity of 0.71), which is crucial in drug toxicity screening for the purpose of patient safety. The benefit of hybrid models was even more drastic when stricter classification thresholds were employed (e.g. CCR would be 0.83 when double thresholds (non-toxic < 0.40 and toxic > 0.60) were used for the hybrid model). We have developed rigorously validated hybrid models which can be used in virtual screening of lead compounds with potential hepatotoxicity. Our study also showed a chemical structure and in vitro biological data can be complementary in enhancing the prediction accuracy of human hepatotoxicity and can afford rational mechanistic interpretation.

Combined toxicity of pesticide mixtures on green algae and photobacteria.

Different organisms have diverse responses to the same chemicals or mixtures. In this paper, we selected the green algae Chlorella pyrenoidosa (C. pyrenoidosa) and photobacteria Vibrio qinghaiensis sp.-Q67 (V. qinghaiensis) as target organisms and determined the toxicities of six pesticides, including three herbicides (simetryn, bromacil and hexazinone), two fungicides (dodine and metalaxyl) and one insecticide (propoxur), and their mixtures by using the microplate toxicity analysis. The toxicities of three herbicides to C. pyrenoidosa are much higher than those to V. qinghaiensis, and the toxicities of metalaxyl and propoxur to V. qinghaiensis are higher than those to C. pyrenoidosa, while the toxicity of dodine to C. pyrenoidosa is similar to those to V. qinghaiensis. Using the concentration addition as an additive reference model, the binary pesticide mixtures exhibited different toxicity interactions, i.e., displayed antagonism to C. pyrenoidosa but synergism to V. qinghaiensis. However, the toxicities of the multi-component mixtures of more than two components are additive and can be predicted by the concentration addition model.

The use of pseudo-equilibrium constant affords improved QSAR models of human plasma protein binding.

PURPOSE: To develop accurate in silico predictors of Plasma Protein Binding (PPB). METHODS: Experimental PPB data were compiled for over 1,200 compounds. Two endpoints have been considered: (1) fraction bound (%PPB); and (2) the logarithm of a pseudo binding constant (lnKa) derived from %PPB. The latter metric was employed because it reflects the PPB thermodynamics and the distribution of the transformed data is closer to normal. Quantitative Structure-Activity Relationship (QSAR) models were built with Dragon descriptors and three statistical methods. RESULTS: Five-fold external validation procedure resulted in models with the prediction accuracy (R²) of 0.67 ± 0.04 and 0.66 ± 0.04, respectively, and the mean absolute error (MAE) of 15.3 ± 0.2% and 13.6 ± 0.2%, respectively. Models were validated with two external datasets: 173 compounds from DrugBank, and 236 chemicals from the US EPA ToxCast project. Models built with lnKa were significantly more accurate (MAE of 6.2-10.7 %) than those built with %PPB (MAE of 11.9-17.6 %) for highly bound compounds both for the training and the external sets. CONCLUSIONS: The pseudo binding constant (lnKa) is more appropriate for characterizing PPB binding than conventional %PPB. Validated QSAR models developed herein can be applied as reliable tools in early drug development and in chemical risk assessment.

Modeling non-monotonic dose-response relationships: model evaluation and hormetic quantities exploration.

Non-monotonic (biphasic) dose-response relationships, known as hormetic relationships, have been observed across multiple experimental systems. Several models were proposed to describe non-monotonic relationships. However, few studies provided comprehensive description of hermetic quantities and their potential application. In this study, five biphasic models were used to fit five hormetic datasets from three different experimental systems of our lab. The bisection algorithm based on individual monotone functions was proposed to calculate arbitrary hormetic quantities instead of traditional methods (e.g., model reparameterization) which need complex mathematical manipulation. Results showed that all the five biphasic models could describe those datasets fairly well with coefficient of determination ( R(2) adj) greater than 0.95 and root mean square error (RMSE) smaller than 0.10. The best-fit model could be selected based on EC(R10), RMSE, and a supplemental criterion of Akaike information criterion (AIC). Hormetic quantities that trigger 10% stimulation at the left (EC(L10)) and right (EC(R10)) side of stimulatory peak were calculated and emphasized for their implication in hormesis exploration for the first time. Furthermore, the EC(L10), proposed as an alarm threshold for hormesis, was expected to be useful in risk assessment of environmental chemicals. This study lays a foundation in the quantitative description of the low dose hormetic effect and the investigation of hormesis in environmental risk assessment.

[Detecting the cytotoxicities of five bisphenol A analogues to the MCF-7 human breast carcinoma cell line through different endpoints].

As the main synthetic raw materials of polycarbonate, bisphenol A (BPA) and its analogues have been important issues in environmental pollution. The current studies focus mainly on BPA's estrogen effects and little on their cytotoxic effects. To assess the cytotoxicities of the five BPA analogues, we employed the MTS assay to determine the inhibition toxicity to MCF-7 (ER-), 2,4-dinitrophenylhydrazine assay to determine the release rate of lactate dehydrogenase (LDH) escaping into cell culture medium, and single cell gel electrophoresis assay (SCGE) to detect DNA damage. The dose-response curves (DRC) between the observed inhibition toxicities and concentrations of the BPA compounds in MTS assay were fitted by using the nonlinear least squares (NLS) and the results showed that all the dose-response relationships were effectively described by the Weibull or Logit function. The toxicities expressed by--lgpEC50 were BPB > BPC > TDP > BPE > BPA. LDH assay and SCGE assay showed that when the concentrations of BPA analogues were EC20, the MCF-7 cell proliferation was slightly inhibited due to its little damaged DNA, and at EC40 the cell proliferations were significantly inhibited due to the seriously damaged DNA, leading to the damage of cell membrane and release of LDH.

A novel method dependent only on the mixture information (MIM) for evaluating the toxicity of mixture.

Compound contamination and toxicity interaction necessitate the development of models that have an insight into the combined toxicity of chemicals. In this paper, a novel and simple model dependent only on the mixture information (MIM), was developed. Firstly, the concentration-response data of seven groups of binary and multi-component (pseudo-binary) mixtures with different mixture ratios to Vibrio qinghaiensis sp.-Q67 were determined using the microplate toxicity analysis. Then, a desirable non-linear function was selected to fit the data. It was found that there are good linear correlations between the location parameter (α) and mixture ratio (p) of a component and between the steepness (ß) and p. Based on the correlations, a mixture toxicity model independent of pure component toxicity profiles was built. The model can be used to accurately estimate the toxicities of the seven groups of mixtures, which greatly simplified the predictive procedure of the combined toxicity.

Predicting hormetic effects of ionic liquid mixtures on luciferase activity using the concentration addition model.

The concept of hormesis has generated considerable interest within the environmental and toxicological communities over the past decades. However, toxicological evaluation and prediction of hormesis in mixtures are challenging and only just unfolding. The hormetic effects of ten ionic liquids (ILs), singly and in mixtures in the ratios of their individual EC50, EC10, EC0, and ECm (maximal stimulatory effect concentration), on luciferase luminescence were determined by using microplate toxicity analysis. There was good agreement between the effects observed and predicted by concentration addition (CA) for all four mixtures. This evidence supports the use of CA model as a default approach for assessing the combined effect of chemicals at the molecular level. Focusing on the selected points of the concentration-response curves (CRCs) of mixtures, the mixtures of IL chemicals mixed at concentrations that individually showed stimulatory effects could produce inhibitory or no effects, and the mixture of IL chemicals mixed at concentrations that individually showed no effects could produce significant inhibitory effect. The three interesting phenomena in mixture hormesis may have important implications for current risk assessment practices.

Comparison between the short-term and the long-term toxicity of six triazine herbicides on photobacteria Q67.

The bioluminescence inhibition of six triazine herbicides including desmetryne (DES), simetryn (SIM), velpar (VEL), prometon (PRO), metribuzin (MET), and aminotriazine (AMI) on Vibrio qinghaiensis sp.-Q67 (Q67) was determined to investigate the effects of exposure duration on the ecotoxicological relevance of triazine herbicides. Based on the short-term microplate toxicity analysis (MTA), a long-term MTA was established to assess the impact of exposure time on the toxicities of the herbicides. The results show that the long-term toxicities of DES and SIM are similar to their short-term toxicities, and the long-term toxicities of VEL, PRO, and MET are higher than their short-term toxicities, while AMI without short-term toxicity has a high long-term toxicity. In addition, a parabolic relationship was found between the pEC(50) (the negative logarithm of the EC(50), log 1/EC(50)) and the logarithm of octanol-water partition coefficient (logK(ow)). To better understand their toxicity process, the time-dependent toxicities of the six herbicides on Q67 were determined over a period of 12 h during which measurements were taken every 30 min to generate an integral effect surface related to both concentration and duration.