High-throughput imaging-based nephrotoxicity prediction for xenobiotics with diverse chemical structures

Our phenotypic profiling strategy (Supplementary Material 1—Fig. S1) was to automatically measure a large number of quantitative phenotypic features of primary human PTCs under in vitro conditions, and then systematically screen for a subset of phenotypic features that were the most predictive of in vivo PTC toxicity. Our features were based on four fluorescent markers (Fig. 2a). We used 4′,6-diamidino-2-phenylindole (DAPI) and rhodamine phalloidin to label the DNA and actin cytoskeleton, respectively. Nuclear and chromatin structure alterations are involved in many fundamental cellular processes, such as transcription, mitosis, and cell death. Actin filaments play an important role in maintaining the cellular function of PTCs (Kellerman et al. 1990). We also labeled the cells with an antibody specific for a subunit of the nuclear factor (NF)-κB complex, RelA. This was motivated by our previous models based on IL-6/8 (Li et al. 2013, 2014; Su et al. 2014; Kandasamy et al. 2015), which are regulated by the NF-κB complex (Matsusaka et al. 1993). The final marker was a whole-cell stain (WCS) used to facilitate automated cell segmentation and measurements of cellular morphology features.

Open image in new window

After compound treatment, we stained PTCs with these four fluorescent markers and imaged them using a high-throughput imaging system. We automatically identified ~500–1000 cells from 36 microscopy images captured for each compound and treatment dosage (Supplementary Material 1—Fig. S2). Then, for each cell, we extracted 129 quantitative phenotypic features (Fig. 2b and Supplementary Material 2), which include 78 Haralick’s texture features (Haralick et al. 1973; measuring the statistics of the spatial co-occurrence patterns of the markers), 29 intensity features (measuring the staining levels of the markers at different subcellular regions), 9 intensity ratio features (measuring the ratios between intensity features), 6 correlation features (measuring the spatial correlations between two markers at the single-cell level), and 6 morphology features (measuring the shape properties of the nuclear and cellular regions). We also included cell count as a feature. Similar phenotypic features and profiling methods were previously used to classify large numbers of small molecules according to their targets/mechanisms (Loo et al. 2007). Therefore, we hypothesized that a subset of these features might also be discriminative enough to predict PTC toxicity.

For each phenotypic feature, we first computed the log₂-ratios (“\(\varDelta\)”) of its values at all the tested dosages with respect to the vehicle controls. Then, we estimated the feature’s dose response curve using a standard log-logistic model, and its maximum response value (“\(\varDelta_{\hbox{max} }\)”) from the curve (Fig. 2c). Finally, the median \(\varDelta_{\hbox{max} }\) values across three biological replicates were computed. The final dataset (“HPTC-A”) was a matrix of \(\varDelta_{\hbox{max} }\) values for 129 phenotypic features (\({\mathbf{F}}_{\text{all}} = \{ {\mathbf{f}}_{1} ,\;{\mathbf{f}}_{2} ,\; \ldots ,\;{\mathbf{f}}_{129} \}\), rows) and 44 xenobiotic compounds (columns; Fig. 2d and Supplementary Material 3). For brevity, all features that we mention in this article are referring to the \(\varDelta_{\hbox{max} }\) values of the respective features and not the raw measured feature values, unless otherwise indicated. Hierarchical clustering of the compounds based on the phenotypic feature values revealed two major clusters (Fig. 2d). One of them was significantly enriched with the PTC-toxic compounds (83 % of the cluster were PTC-toxic compounds; P = 1.59 × 10⁻³, hypergeometric test), and the other one was significantly enriched with the non-PTC-toxic compounds (65 % of the cluster were non-PTC-toxic compounds; P = 1.59 × 10⁻³, hypergeometric test). Most of the phenotypic features showed larger changes under the first cluster than the second cluster, suggesting that non-PTC-toxic compounds only induced small or no change in the primary human PTCs. We also performed similar clustering analysis on the phenotypic features and found two major clusters corresponding to either increased or decreased feature values after treatments with PTC-toxic compounds (Fig. 2d). Features from all markers were represented in both clusters. Therefore, most of our phenotypic features are diverse and capture both increasing and decreasing properties of the markers.

To test each individual feature, we constructed a binary classifier based on the feature using a random forest algorithm (Breiman 2001; Su et al. 2014) and estimated the prediction accuracy using a tenfold cross-validation procedure (Fig. 3a, Methods, and Supplementary Material 1—Text S1). In theory, the balanced accuracy (average of sensitivity and specificity) of a binary classifier ranges from 50 % (performance of a trivial random classifier) to 100 % (maximum); but in practice, it may go slightly below 50 %. The training accuracy is the accuracy in classifying the training data used to build the classifier, and the test accuracy is the accuracy in classifying independent test data unused during the training process. We used all the compounds during cross-validation, but our evaluation procedure ensured that the training and test samples were coming from different compounds and statistically independent. We repeated the cross-validation procedure 10 times with different random fold divisions, and all the mean accuracy values presented in this report were obtained by averaging the accuracy values from every fold and trial.

Open image in new window

For the HPTC-A dataset, we found that all single features had ~97 % and above training accuracy and ~43–76 % test accuracy (Supplementary Material 2). The high training-accuracy values do not imply all the features had similar or indistinguishable training performance, because the training accuracies of the features was still strongly correlated with their test accuracies (r = 0.842, Supplementary Material 1—Fig. S3). In our unbiased approach for phenotypic profiling, we started with a large number of general phenotypic features, but only expected a small number of features to be discriminative. We aimed to find out which of the four fluorescent markers and which of the five feature types produced the most discriminative features. Therefore, we grouped all the features according to their fluorescent markers or feature types and compared the maximum achievable accuracies of these feature groups. The feature marker group with the highest maximum test accuracy was DNA (75.8 %), followed by actin (73.7 %) and RelA (72.6 %; Fig. 3b). Surprisingly, RelA features were not highly predictive of PTC toxicity. For example, the RelA nuclear-to-whole-cell intensity ratio, which is an indicator of NF-κB nuclear translocation and transcriptional activation of its downstream effectors (Deptala et al. 1998), only had 61.0 % test accuracy. The feature type groups with the highest maximum test accuracy was Haralick’s texture (Haralick et al. 1973; 75.8 %), followed by intensity (73.7 %) and intensity ratio (69.9 %; Fig. 3c). In fact, six of the ten best-performing single features were all Haralick’s texture features, which are based on the gray-level co-occurrence matrices (GLCM; Haralick et al. 1973) of the fluorescent markers. The GLCM of a marker summarizes the distribution of spatial transitions between different intensity levels of the marker in a cell image (Haralick et al. 1973). Haralick’s features, which describe various statistical properties of a GLCM, can be used to represent the textural patterns found in the image (Methods). The best single feature, \({\mathbf{f}}_{\text{best}}\), among all the 129 features was the “mean entropy” of the DNA GLCM (75.8 % test accuracy). The feature is a measure of the homogeneity of the DNA GLCM. Cell images with more “random” DNA staining patterns, where the transitions between all intensity levels are more equally probable, have more homogenous GLCMs and thus higher values of GLCM entropy (Supplementary Material 1—Fig. S4). Overall, changes in the texture of the DNA and actin cytoskeleton localization patterns were highly predictive of the in vivo PTC toxicity of xenobiotics with diverse chemical structures. The high accuracy underscores the importance and advantage of using image-based phenotypic features as in vitro toxicity endpoints.

Xenobiotic compounds may induce different types of PTC injuries and responses. Therefore, classifiers based on multiple different phenotypic endpoints are more likely to give higher overall prediction accuracy. To preserve the dependency between features, we trained multi-dimensional classifiers based on multiple features simultaneously (Fig. 3d). Then, a recursive feature elimination algorithm (Loo et al. 2007) was used to automatically remove irrelevant and/or redundant features (Fig. 3e, Methods, and Supplementary Material 1—Text S1). The number of retained features was automatically determined based on the training data only. Therefore, the process was repeated for every cross-validation fold, which had different training data. The features were ranked according to their importance values, which were estimated by the random forest algorithm (Breiman 2001) and averaged across all the cross-validation folds. For the HPTC-A dataset, we found a set of four features (\({\mathbf{F}}_{s}\)) that had the highest average importance values (Supplementary Material 1—Fig. S5). These features were the “coefficient of variation (CV)” of DNA intensity at the nuclear region, “mean angular second moment (ASM)” of the DNA GLCM, “mean sum of entropy” and “mean entropy” of the actin GLCM (Fig. 3f and Supplementary Material 1—Table S2). Similar to the single-feature classification results, these top features were all based on the DNA and cytoskeleton markers, and three of them were texture features. We trained multi-feature classifiers using these four features and obtained 78.3 % test accuracy, which were higher than the performances of all single-feature classifiers (Fig. 3f). The individual test accuracies of these four features only ranged between 65.5 and 74.4 %. Therefore, combining the features together increased the prediction performance. We also found that the inclusion of \({\mathbf{f}}_{\text{best}}\) into \({\mathbf{F}}_{s}\) did not further increase the prediction accuracy of our models (Fig. 3f), indicating our recursive feature elimination algorithm was highly effective.

In the feature space of \({\mathbf{F}}_{s}\), we found that most of the non-toxic compounds formed a tight cluster, where they were closer to each other than to the toxic compounds (Fig. 3g). Six of the eight toxic industrial chemicals with simple chemical structures could now be clearly separated from the non-toxic compounds (Fig. 3g). This separation was not evident in the original chemical structure space (Fig. 1b). Among all the tested compounds, 22 compounds were consistently being correctly classified by both the single- and multi-feature classifiers (Supplementary Material 1—Table S3). Only three compounds, namely ciprofloxacin (antibiotic), levodopa (psychoactive drug), and copper(II) chloride (industrial chemical), had consistently <50 % test accuracy in both types of classifiers. These results show that our computational models were general and did not favor specific classes of compounds.

To further investigate the type of cell injury and damage response represented by our phenotypic features, we focused on the two DNA features in \({\mathbf{F}}_{s}\), namely 1) the mean ASM of the DNA GLCM, which had the highest single-feature test accuracy among the four selected features, and 2) the CV of DNA intensity at the nuclear region (Fig. 3f). ASM is a measure of the heterogeneity of a DNA GLCM (Methods and Fig. 4a). The feature gives high values when the transitions between certain intensity levels are dominant (for example, when the intensity values form certain regular shapes), or low values when all transitions are equally probable (for example, when the intensity values are diffused and randomly distributed). CV, which is equal to standard deviation divided by mean, is a standardized measure of the dispersion of a set of values, which in our case were the DNA staining intensity levels within the nuclear region. By examining the fluorescence microscopy images of the cells, we found that cells with high values of these two features had disconnected, highly irregular, and punctate DNA staining levels (Fig. 4a), indicating profound changes in chromatin structure and the formation of distinct chromatin domains. However, the overall nuclear and cellular morphologies of these cells remained largely intact and not fragmented, as in typical apoptotic cells. Therefore, we hypothesized that the features may indicate a DNA damage response, which is known to be associated with the formation of distinct chromatin domains in the megabase size range and large-scale chromatin reorganization (Rogakou et al. 1998; Jakob et al. 2011).

Open image in new window

To test our hypothesis, we repeated the treatment experiments for 42 reference compounds, but replacing the RelA marker with an antibody specific for histone H2AX phosphorylated on serine 139 (γH2AX), which is a DNA damage response marker (Rogakou et al. 1998; Fig. 4b). Under endogenous or exogenous DNA damage conditions, γH2AX is induced and recruits repair factors to the sites of double-strand breaks (Paull et al. 2000). We repeated the experiments in both primary human PTCs (the “HPTC-B” dataset) and an immortalized human PT cell line, human kidney 2 (the “HK-2” dataset, Supplementary Material 1—Fig. S6). At the single-cell level, we found that cells with higher raw DNA CV levels induced by xenobiotic compounds tended to have higher raw mean γH2AX nuclear levels, but the responses might be highly heterogeneous (Fig. 4b). For example, 500 μg/mL cyclosporin A caused ~40-fold increases in the raw mean γH2AX nuclear levels, but only in ~13 % of the cells (Fig. 4c). Nevertheless, due to the large increases in magnitude, the effects could still be detected at the population-averaged level. Similar increases in γH2AX nuclear levels were also observed in cells treated with other PTC-toxic compounds (Supplementary Material 1—Fig. S6). Across all the tested compounds, the maximum increases (i.e., \(\varDelta_{\hbox{max} }\)) in DNA CV and mean nuclear γH2AX levels were strongly and positively correlated with each other in both primary and HK-2 cells (r = 0.639 and 0.667, respectively; Fig. 4d, e). Furthermore, both features were significantly higher in cells treated with the PTC-toxic compounds than those with the non-PTC-toxic compounds (all P < 0.01, one-tailed t test; Fig. 4d, e). These results suggest that most of the PTC-toxic compounds induce a DNA damage response, even though many of them are not known to bind to DNA directly.

To what extent can the γH2AX marker improve the prediction performance of our computational models? We repeated the same phenotypic profiling procedure but using 129 phenotypic features based on the DNA, γH2AX and actin markers (Fig. 4f, g). For the HPTC-B dataset, we found that the best single feature \({\mathbf{f}}_{\text{best}}\) was the ratio of total γH2AX levels at the nuclear to the whole-cell regions (77.6 % test accuracy), which indicates the generation of γH2AX at the nuclear region (Supplementary Material 1—Fig. S4). The best multi-feature set \({\mathbf{F}}_{s}\) were four nuclear and actin cytoskeletal features (81.6 % test accuracy, see Supplementary Material 1—Fig. S5 and Table S2 for the complete listing of features). For the HK-2 dataset, we found that its \({\mathbf{f}}_{\text{best}}\) was the mean correlation of DNA GLCM (83.9 % test accuracy), which is a measure of the linear dependency of intensity levels of neighboring pixels (Supplementary Material 1—Fig. S4). The best multi-feature set \({\mathbf{F}}_{s}\) were five chromatin and cytoskeleton features (88.9 % test accuracy, Supplementary Material 1—Fig. S5 and Table S2). For both datasets (HPTC-B and HK-2), we found a consistent trend that multi-feature classifiers had significantly higher test accuracies than single-feature classifiers (Fig. 4f, g, P = 0.075 and 0.024, one-sided Wilcoxon signed-rank test). However, single-feature classifiers had higher test specificities, while multi-feature classifiers had higher sensitivities (Fig. 4h, i). Furthermore, the number of compounds that could be predicted with 100 % average test accuracy in both single- and multi-feature classifiers had increased from 22 (HPTC-A) to 25 (HPTC-B) or 28 (HK-2) (Supplementary Material 1—Table S3). Together, these results show that the inclusion of the γH2AX marker allowed us to obtain higher prediction accuracies.

We also compared the optimum phenotypic features selected for all three datasets (Supplementary Material 1—Table S2) and found several interesting and consistent trends. First, the mean ASM of DNA GLCM was automatically selected in the \({\mathbf{F}}_{s}\)’s for both HPTC-A and HPTC-B datasets. Second, the \({\mathbf{f}}_{\text{best}}\) for the HPTC-B dataset (i.e., ratio of total γH2AX levels at the nuclear to the whole-cell regions) and one of the features in the \({\mathbf{F}}_{s}\) for the HK-2 dataset (i.e., ratio of total γH2AX and DNA intensity levels at the whole-cell region) are two closely related features that indicate nuclear increase in γH2AX levels (Supplementary Material 1—Fig. S4). Third, although the best single features were based on DNA or γH2AX markers, actin features were still retained by the multi-feature classifiers, suggesting that the actin marker was needed to correctly classify some compounds that induced actin remodeling. Together, these results show that our predictive models are highly reproducible, and a xenobiotic-induced DNA damage response is a general phenomenon that occurs in both human primary and immortalized PTCs.