ic50 experiment

IC50’s: An Approach to High-Throughput Drug Discovery

Author: shouvik saleh, learning objectives.

Describe Drug-Drug Interaction studies and their role in clinical research.
Understand IC 50 ’s, how they are obtained, and its use in high-throughput settings.
List some limitations of IC 50 ’s.

Graphical Abstract

Images retrieved from: Boghog, https://commons.wikimedia.org/wiki/File:Drug_discovery_cycle.svg Tiia Monto, https://commons.wikimedia.org/wiki/File:High-throughput_screening.jpg Andrux, https://commons.wikimedia.org/wiki/File:Multipipet.JPG

Introduction/Background

The rates of biological processes can be affected by the presence of inhibitors or enzymes, but you may wonder – What happens when a drug is added to the equation? How does the quantity and duration of effect of the drug produce changes in these biological processes? This is where IC 50 ’s come in.

DDI Studies

The term IC 50 refers to the half maximal inhibitory concentration and is a highly relevant concept for drug discovery and analysis. A study with this might be trying to look at how concentrations of a drug change over time in a body and what interactions occur during that period; in clinical settings, these are called Drug-Drug Interaction (DDI) studies 1 . There are many different interactions that can occur between drugs and compounds or structures in the body; for example, two drugs may work in concert to produce an additive effect relating to a disease, two drugs may work antagonistically to change the intended effects for a disease, or a new drug and a well-characterized drug with established effects toward a disease or biological process may be used to learn about what kind of interactions the new drug may have in certain conditions. 1 A study might look at what maximum and minimum concentrations of a drug are observed as time passes following exposure to a tissue or administration to the body. In these studies, well characterized enzymes, such as cytochrome P450 enzymes, are often used as the target of inhibition, since these liver enzymes are often involved in drug metabolism pathways after administration.

IC 50 ’s in High-Throughput

IC 50 ’s are used in lab settings to estimate the quantities of these drugs that result in a 50% inhibition of a biological process/mechanism or drug interaction specifically in cultured cells. 1 These are especially useful for studies looking at novel situations of adding one drug to the body to blunt the effect of another drug, for example; either drug may be the drug of interest for the study here. On the other hand, EC 50 ’s, or the half maximal effective concentration, are used to look at excitatory drug interactions rather than inhibitory ones. 2 In essence, both of these measures are of drug potency.

To study these drug interactions in high-throughput, primary screens of chemical libraries consisting of 100,000 to up to 2 million compounds may be tested with cells. To model the interaction of the compounds in diseased humans, proteins observed to be a possible or known cause for the disease may be engineered into cells. These cells are then exposed to compounds from chemical libraries using liquid handlers and their activity is observed over time. The activity is measured before compound addition (at time=0) to get a reading for minimal inhibition from a compound and in time increments until all activity of the cell ceases (at time = infinity) to get a reading for maximal inhibition. A dose-response curve is then made from wells containing cells that show inhibitory effects above a certain threshold, as some compounds may fail at producing an inhibitory effect. The IC 50 value is then estimated from the curve using a logistic regression equation, called the 4-parameter logistic Hill equation, used in dose-response relationships. 2

Limitations

There can be limitations to this process however. There can be variability in concentrations between samples due to liquid handling and characteristics of the reagents being used; inconsistencies in reported data being used for parameters used for calculations due to interactions between reagents and the assay being used; and influences on data due to experimental design and quality. 3 Since a basic assumption of this high-throughput design is that the percent inhibition and IC50 values correlate reasonably, there can be room left for error when completing calculations with the Hill equation. 3 Additionally, inaccuracies have been identified in public collections of compound data due to including below-threshold data. 4 It is important that false negatives and false positives are identified to accurately identify compounds that may play potential roles in drug-disease interactions.

Accuracy could possibly be improved by observing compounds at different concentrations but this would also increase the high-throughput demands of the experiment. All in all, IC50’s are an important concept allowing for the responses of pre-made chemical libraries to be observed as they are being exposed to biological compounds or cells. This is important not only for finding potential treatments for diseases, but also for determining the possible side effects of these treatments. High throughput technique enables information about these compounds to be obtained in a time- and resource-efficient manner.

Audio Recording

Brody, T. (2018). “Drug–Drug Interactions: Part One (Small Molecule Drugs).” FDA’s Drug Review Process and the Package Label: Strategies for Writing Successful FDA Submissions, Academic Press, 7(1), 255-335. Accessed: https://doi.org/10.1016/B978-0-12-814647-7.00007-5.
Sebaugh, J. L. (2011) “Guidelines for Accurate EC50/IC50 Estimation.” Pharmaceutical Statistics, vol. 10, no. 2, pp. 128–134. doi:10.1002/pst.426.
Limitations Gubler, H., Schopfer, U., & Jacoby, E. (2013). “Theoretical and Experimental Relationships between Percent Inhibition and IC50 Data Observed in High-Throughput Screening.” Journal of Biomolecular Screening, 18(1), 1–13. Accessed: https://doi.org/10.1177/1087057112455219.
Kalliokoski, T., Kramer, C., Vulpetti, A., & Gedeck, P. (2013). “Comparability of mixed IC₅₀ data – a statistical analysis.” PloS one, 8(4), e61007. https://doi.org/10.1371/journal.pone.0061007
What are IC50’s used for? These are used to look at the inhibition of a drug or biological process as a result of interruption or slowing caused by drug interactions from another drug.
What is needed to calculate an IC50? Minimal and maximal activity levels of cells following exposure to a constant concentration of compound
What are some applications for IC50’s in high-throughput studies? They can be used alongside chemical libraries with cells for high-throughput drug discovery and identification.
What is a possible interaction between drugs that might be seen in a DDI study for which IC50’s would be inappropriate? They cannot be used to look at non-inhibitory interactions, since the measurements are looking at decreases in activity of cells.
What are some limitations to IC50’s in high-throughput applications? They require optimization based on many characteristics related to the assay and are susceptible to errors due to possible missed recordings of variations in data.

An official website of the United States government

The .gov means it’s official. Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

The site is secure. The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

Publications
Account settings
My Bibliography
Collections
Citation manager

Save citation to file

Email citation, add to collections.

Create a new collection
Add to an existing collection

Add to My Bibliography

Your saved search, create a file for external citation management software, your rss feed.

Search in PubMed
Search in NLM Catalog
Add to Search

An examination of IC50 and IC50-shift experiments in assessing time-dependent inhibition of CYP3A4, CYP2D6 and CYP2C9 in human liver microsomes

Affiliation.

1 Pharmacokinetics and Drug Metabolism, Amgen, Inc., 1 Kendal Sq., Cambridge, MA 02139, USA. [email protected]
PMID: 19356071
DOI: 10.2174/187231208783478407

The relationship between time-dependent inactivation (TDI) and IC50 is examined using a consolidated method for evaluating CYP450 inhibition during drug discovery. An IC50 fold-shift of >1.5 indicated significant TDI potency. Further, the "shifted IC50" could be used to estimate, the K(I) and TDI potency ratio k(inact)/K(I) to within 2-fold in most cases.

PubMed Disclaimer

LinkOut - more resources

Full text sources.

Ingenta plc

Other Literature Sources

The Lens - Patent Citations
Citation Manager

NCBI Literature Resources

MeSH PMC Bookshelf Disclaimer

The PubMed wordmark and PubMed logo are registered trademarks of the U.S. Department of Health and Human Services (HHS). Unauthorized use of these marks is strictly prohibited.

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

View all journals
Explore content
About the journal
Publish with us
Sign up for alerts
Open access
Published: 22 April 2022

Numerical learning of deep features from drug-exposed cell images to calculate IC50 without staining

Kookrae Cho 1 na1 ,
Eun-Sook Choi 1 na1 ,
Jung-Hee Kim 1 ,
Jong-Wuk Son 1 &
Eunjoo Kim 1

Scientific Reports volume 12 , Article number: 6610 ( 2022 ) Cite this article

3174 Accesses

5 Citations

Metrics details

To facilitate rapid determination of cellular viability caused by the inhibitory effect of drugs, numerical deep learning algorithms was used for unlabeled cell culture images captured by a light microscope as input. In this study, A549, HEK293, and NCI-H1975 cells were cultured, each of which have different molecular shapes and levels of drug responsiveness to doxorubicin (DOX). The microscopic images of these cells following exposure to various concentrations of DOX were trained with the measured value of cell viability using a colorimetric cell proliferation assay. Convolutional neural network (CNN) models for the study cells were constructed using augmented image data; the predicted cell viability using CNN models was compared to the cell viability measured by colorimetric assay. The linear relationship coefficient (r 2 ) between measured and predicted cell viability was determined as 0.94–0.95 for the three cell types. In addition, the measured and predicted IC50 values were not statistically different. When drug responsiveness was estimated using allogenic models that were trained with a different cell type, the correlation coefficient decreased to 0.004085–0.8643. Our models could be applied to label-free cells to conduct rapid and large-scale tests while minimizing cost and labor, such as high-throughput screening for drug responsiveness.

Automated spheroid generation, drug application and efficacy screening using a deep learning classification: a feasibility study

DeepInsight-3D architecture for anti-cancer drug response prediction with deep-learning on multi-omics

Cancer drug sensitivity prediction from routine histology images

Introduction.

In vitro high-throughput assays for screening drug responsiveness are increasingly needed not only for medical and pharmaceutical research, but also for clinical purposes using cell lines, patient-derived primary cells, or organoids. In this context, automatic analysis of drug responsiveness using cell images have facilitated advanced microscopic technology such as counting cell numbers and discriminating between live and dead cells to determine in vitro responses 1 , 2 .

Recently, artificial intelligence-involved instrumental analysis has improved the experimental process by using accumulated data from researchers as training data to predict analytical results, reducing time, cost, and labor. Deep-learning-based training has been applied to obtain optimized results for the counting of cells with fluorescence or colorimetric staining 3 , 4 . An image-based learning approach was also applied to the discrimination of live and dead cells following the fluorescence staining of cells after drug treatment 5 , 6 . The continuous values of labels assigned to cell images were predicted by performing numerical deep learning on label-free input images using a recent image analysis technique 7 , 8 . In these studies, the continuous values of fluorescent intensity from stained cells were assigned to non-stained light microscope images, and then the fluorescent intensity was predicted using the label-free cell images.

To provide numerical values of cell viability without staining, label-free cell images could be assigned with measured values obtained by colorimetric cell proliferation assay and used as input data for numerical learning. In this case, it would be possible to predict 50% inhibitory concentration (IC50) based on precisely predicted cell viability from the training of cell images. Using this model, high-throughput drug responsiveness screening could be achieved by rapid and automatic analysis saving time, cost, and labor.

In this study, we provide numerical deep learning results using label-free cell images, which were used to determine IC50 values based on the predicted cell viability in cell culture dishes. The ground truth label was the cell viability indicator, optical density at 450 nm (OD 450 nm ), which was obtained by cell proliferation assay following drug treatment of cells. Cell viability can be defined as the number of cells that are either alive or dead after undergoing drug treatment. The prediction result was compared to the measured values, and the correlation coefficient and statistical difference between the measured and predicted data were evaluated. The drug doxorubicin (DOX) was used to induce cytotoxicity in three types of cell lines: A549, HEK293, and NCI-H1975. These cell lines were used to prepare label-free cell images.

Because an efficient method for cell viability prediction is critical for various cellular culture-based experiments, including anti-cancer drug discovery, a web-based algorithm for image-based IC50 determination without staining of cells was introduced to support high-throughput image analysis with low cost and time requirements. Finally, the prediction result of cell viability and IC50 values was evaluated in aspect of possible alternative method to replace the conventional colorimetric cell proliferation assays.

Materials and methods

Cell lines and cultures.

Human epithelial lung carcinoma A549 cells (CCL-185), human kidney HEK293 cells (CRL-1573), and human non-small cell lung cancer NCI-H1975 cells (CRL-5908) were purchased from the American Type Culture Collection (ATCC, Manassas, VA, USA). The cells were cultured in Gibco RPMI-1640 culture medium (Thermo Fisher, Waltham, MA, USA) supplemented with 10% fetal bovine serum albumin (FBS, Thermo Fisher, Waltham, MA, USA). During the culture process, the cells were placed in a humidified incubator at 37 °C with 5% CO 2 .

Measurement of cell viability

To prepare images of the cultured cells corresponding to the molar concentration of a drug to inhibit normal cellular growth, DOX was added to A549, HEK293, and NCI-H1975 cells in concentrations ranging from 0.001 nM to 100 μM (total 11 concentrations). For each DOX concentration, eight repetitive samples were prepared, and three sets of experiments were performed separately for the A549, HEK293, and NCI-H1975 cell types. Briefly, cells were seeded into 96-well plates at a density of 1 × 10 3 cells/well, and DOX was added to each well at a selected concentration after 24 h of incubation. After 48 h of incubation, the cell images from each well were captured by light microscopy (Leica DMI3000B, Leica Microsystems, Wetzlar, Germany) equipped with a Zyla sCMOS camera (Andor Technology Ltd, Belfast, UK). An image for each well was captured, and every sample was in the center of the field of view.

Cell viability was then determined using Cell Counting Kit-8 (CCK kit, Dojindo Molecular Technology Inc., Rockville, MD, USA), according to the manufacturer’s instructions. The OD 450 nm for the CCK assay was measured using a spectrophotometric microplate reader (Promega, Madison, WI, USA).

Image pre-processing

We used bright-field images of the cells in 96-well plates as input variables. For the classification of learning data, each cell image was assigned quantitative labels of cell viability data (OD 450 nm ) following 48 h of DOX exposure, which was also taken as the expected output. Before predicting drug responsiveness using label-free cell images and further calculation of IC50, we performed an exploratory analysis on the actual dataset. For all the dataset images, we constructed 1028-dimensional feature vectors using MobilenetV2, which was pre-trained using the ImageNet dataset ( https://www.image-net.org/ ) without classification layers. Patterns such as clustering, abnormality, and outliers were obtained as the training results from the learned feature vectors.

In addition, we performed data augmentation for the training images, artificially increasing the diversity and size of the training samples to reduce the required number of datasets 9 , 10 . This included using random 90-degree rotations, random cropping (to 70% of the original size), and vertical/horizontal flipping of each image. All the original images are in RGB, JPG format with a resolution of 2160 × 2560 pixels. The training images were cropped to 1512 × 1792 pixels which is 70% of the original image. In addition, the images were adjusted for brightness in the range [− 0.2, 0.2], saturation in the range [0.6, 1.6], contrast in the range [0.7, 1.3], and hue in the range [–0.08, 0.08].

Model construction for the prediction of cell viability

For model training, transfer learning was applied using MobileNetV2 architecture and feature weights, which were pre-trained for the ImageNet dataset as described earlier. The MobileNetV2 architecture without the top classification layer was used as the base model for the CNN image analysis. One fully connected layer with 10 hidden nodes was linked to it, followed by another fully connected layer that outputs the prediction result. The frozen base model was used as a feature extractor, and only the added layers were trained using the base model.

Among the 264 images labeled with OD 450 nm values determined by the cell proliferation assay, 198 images were used for training and validation and 66 images were reserved for testing. Original (unaugmented) images were used to predict cell viability, with the splitting ratio of the dataset being 6:4. For a regression analysis, the numerical output for each cell image was predicted as a drug responsiveness score and compared it to the measured values determined by the cell proliferation assay. To support the test results, we additionally performed fourfold cross validation. Each fold had 25% of the data set without repetition. In addition to MobileNetV2, Inception V3 11 and InceptionResNet V2 12 models were also trained and applied for fourfold cross validation.

Visualization of feature attributions

We applied two visualization techniques to investigate the behavior of the trained model. First, we extracted the feature vectors of the test images from the output of the trained model without the final prediction layer. These high-dimensional feature vectors were transformed to low-dimensional features using two-dimensional (2D) t-distributed stochastic neighbor embedding (t-SNE) 13 . Next, we adopted Google Explainable AI with XRAI option ( https://cloud.google.com/explainable-ai ) 14 to display the attributes of features in the CNN model.

Determination of IC50 by doxorubicin

Based on the changes in cell viability following treatment of cells with DOX, IC50 values were determined from non-linear regression curves fitted to DOX concentration and OD 450 nm values, according to the following Hill Eq. ( 1 ), where Max and Min are the maximum and minimum values of OD 450 nm in the sigmoidal regression, respectively.

The measured IC50 was calculated using the least square fit of four parameter-sigmoidal curves executed using Prism ver. 9 (GraphPad Software, San Diego, CA, USA).

Web application

We developed a web application that allows a user to predict drug responses from cell images and calculate the IC50 from the predicted responses. First, we used the Python-based TensorFlow to build and train the model according to the structure and method mentioned earlier. Then, the trained model was saved in HDF5 standard format and converted to TensorFlow.js Layers format for use in the browser (Fig. S1 ). The converted model files were loaded from a URL where the hosted model files were hosted. When a user uploads the images of the cells that had been treated with a specific concentration of drugs to a web browser, the drug responsiveness in cell viability can be predicted, allowing for the IC50 to be calculated at once. Three types of models were developed separately, predicting drug responsiveness by calculating the IC50 based on the A549, HEK293, and NCI-H1975 cell images.

Statistical analysis

Comparison between the measured and predicted values was performed using the correlation coefficient r 2 . Unpaired and paired t tests were also performed for groups of measured and predicted values using Prism ver. 9.3.0 (GraphPad Software, San Diego, CA, USA, https://www.graphpad.com/scientific-software/prism/ ). Differences between measured and predicted values were calculated as average (AVE) ± standard deviation (SD) from the paired t test.

Results and discussion

Images obtained for training of the deep learning model.

Three cell types, A549, HEK293, and NCI-H1975, were treated with DOX at concentrations of 0.0001, 0.001, 0.005, 0.01, 0.05, 0.1, 0.5, 1, 5, 10, and 100 μM in 96-well cell culture plates. For each of the 11 concentrations, eight repetitive samples were prepared, and three sets of 11 concentration × 8 repetitive samples were prepared for the numerical learning of bright-field cell images. The images of cells in 96-well plate wells under the given concentration of DOX were captured using a light microscope at × 100 magnification and used as input data for image classification to predict the cell viability and IC50 quantitatively. Representative cell images are shown in Fig. 1 .

Representative cell images used in this study. The images were captured by light microscope with a magnification ratio × 100.

Convolution of neural network model construction to predict drug responsiveness

In this study, we propose a CNN-based regressor that predicts the drug responsiveness of cells from their bright-field images. The CNN model was constructed using MobileNet2, which is very similar to the original MobileNet, except that it uses inverted residual blocks with bottlenecking features for analysis 15 . MobileNet is suitable for mobile vision applications because it dramatically reduces network complexity and model size 16 . The model was trained using unlabeled cell images with the assigned cell viability scores as inputs. These cell viability scores were determined by a cell proliferation assay using the colorimetric method, with continuous values ranging between 0 and 2. By setting the cell viability predictions as expected outputs, the responsiveness to DOX was estimated using the IC50 indicator for three types of cells: A549, HEK293, and NCI-H1975.

In this study, data augmentation and transfer learning were performed to induce high performance in the accuracy of the prediction, showing that even a relatively small dataset (264 images for each cell type) could be used in numerical learning. Data augmentation techniques can be used to overcome the limited number of available data by artificially increasing the diversity and size of the training samples 9 , 10 . In this study, training images were augmented by performing operations such as zoom, flip, and rotate. When data augmentation such as random crop or zoom operation were applied to train a model that predicts continuous values, the augmented images might not reflect the absolute response. Therefore, the size of the cropping or zooming of the original image should be predetermined to sufficiently reflect the cell distribution of the original cell images. In this study, the responsiveness of drugs represented by IC50 was determined by the relative 50% inhibition concentration between 0 and 100% inhibition, rather than being determined by the absolute number of dead or living cells. Therefore, it was required to determine the appropriate random crop size to accurately predict the IC50 value while increasing the amount of training data by data augmentation. We randomly cropped the original image to 70% size, computed the feature vectors, and performed 2D t-SNE visualization. Figure S2 presents the 2D t-SNE visualization of the augmented image embedding. Images augmented at the same concentration formed a cluster, with a distance maintained between different clusters. It was expected that a random crop of 70% size would be optimal for data augmentation in this study. As a result, the data augmentation process is suitable for pre-processing the model construction.

Among 264 cell proliferation images for each cell type, 198 images were used for training and validation, and 66 images were kept for testing. The splitting ratio of the dataset used in this analysis (6:4) was different from the commonly suggested ratio for training, validation, and testing (8:1:1) 17 . This was because a set of testing images were composed of eleven images corresponding to the eleven concentration of DOX: six sets of experimental results were used as testing sets.

Finally, three types of models, A549-M, HEK293-M, and NCI-H1975-M, were constructed, each of which used input images of A549, HEK293, and NCI-H1975, respectively. The trained model was constructed in a web-based application so that it could be used in a web browser to provide an easily accessible user interface ( https://bioanalysis-79545.web.app/main/images/a549 ). Based on the cell viability predicted from the microscopic cell image taken under a specific concentration of DOX treatment, IC50 was subsequently calculated using a custom algorithm that solved the Hill equation with four parameters. To obtain an accurate result of cell viability using the web-based application, the images should be captured using a sCMOS camera (Andor Technology) under × 100 magnification ratio.

Prediction of cell viability using a CNN image analysis model

After training, the model was validated to determine whether the predictions based on label-free cell images were comparable to those of the conventional cell proliferation assay. The predicted values from the three types of models were plotted against the measured values obtained from the CCK assay. Linear regression analysis was performed to determine the correlation coefficient r 2 . Figure 2 shows the relationship between the measured and predicted values. The deviation from linearity was lowest when the OD 450 nm was estimated by the autologous model of each cell type; for example, the measured OD 450 nm values of A549 treated with various DOX concentrations were predicted more straightforwardly by A549-M, which was the model trained using A549 cell images. HEK293 and NCI-H1975 cells were also best fit by each autologous prediction model.

Correlation of measured and predicted optical density measured at 450 nm following colorimetric cell proliferation assay (CCK assay). The models constructed for the three types of cells were A549-M, HEK293-M, and NCI-H1975-M. The predicted values were drawn up from allogenic models as well as autologous models for cross validation. The linearity and statistical analysis results are shown in Table 1 .

Table 1 shows the quantitative evaluation of the correlation between the measured and predicted values of OD 450 nm by statistical analysis. The correlation coefficient r 2 values between the measured OD 450 nm of A549, HEK293, and NCI-H1975 and the values predicted by autologous models A549-M, HEK293-M, and NCI-H1975-M, were determined to be 0.9513, 0.9404, and 0.9491, respectively. These values are relatively higher than those estimated by allosteric prediction models: for example, decreased r 2 (0.07962) was observed between measured OD 450 nm values for A549 cell viability and predicted OD 450 nm by the model trained using NCI-H1975 cell images (NCI-H1975-M). The slopes of the linear regression curves between the measured and estimated values using autologous prediction models were 1.138, 1.079, and 1.002 for A549, HEK293, and NCI-H1975, respectively. This result indicates that the relationship is almost comparable with the proportional ratio near ~ 1.0.

The difference ± SD between the measured and predicted values was also determined, as shown in Table 1 . The difference was calculated as 0.04874 ± 0.1714, 0.02182 ± 0.1107, and 0.01163 ± 0.1370 for A549, HEK293, and NCI-H1975, respectively, when each autologous model was used for estimation. The deviation of the prediction from the measured values increased over tenfold when the allosteric models were used to estimate cell viability. In the case of the difference between the measured and predicted values for HEK293 cell viability using the allogenic model of NCI-H1975, the average difference decreased to 0.0041, even though the autologous model prediction was 0.0218, which was attributed to the compensation effect of the summation of the difference in positive and negative values. The SD of the difference was still enhanced in the case of allogenic NCI-H1975-M.

Figure 3 shows the distribution of the measured and predicted values based on the autologous and allogenic models. The difference between the measured and predicted values based on the autologous models showed a narrower distribution compared to that of the allogenic models, and the predicted values of OD 450 nm for A549, HEK293, and NCI-H1975 were more precisely matched to the estimated values from the respective autologous models A549-M, HEK293-M, and NCI-H1975-M.

Correlation of measured and predicted optical density measured at 450 nm (OD 450 nm ) analyzed by paired t test. The corresponding measured and predicted values and difference of two values are plotted. The left panel for each cell type shows predicted values from autologous models; in the middle and right panels, these values are compared with the predicted values from allogenic models.

In a previous study on the numerical deep learning of CNN to predict cell viability using label-free 3D cell images, the correlation coefficient r 2 was determined to be 0.82–0.93 7 . In this case, cell images were assigned by fluorescence intensity following LIVE/DEAD cell staining of 1920 tumor sphere samples. In our study, the correlation coefficient r 2 improved to over 0.94, and moreover, this was achieved by almost one-seventh of the training samples, more efficient than the previous study.

To support the improved accuracy even using the limited number of samples in this study, we additionally performed fourfold validation analysis by three types of CNN models, MobileNetV2, InceptionV3, and InceptionResNetV2. Table S1 shows the r 2 coefficients between the measured and predicted values of OD 450 nm for A549 cells by fourfold cross validation. The results indicated that the r 2 values were 0.9237, 0.9325, and 0.9331 for MobileNetV2, InceptionV3, and InceptionResNetV2 models, respectively, which were close to each other. Table S2 shows the results of fourfold cross validation using MobileNetV2 for A549, HEK293, and NCI-H1975 cells. The r 2 values determined by the fourfold cross validation were 0.9237, 0.9218, and 0.9290 for A549, HEK293, and NCI-H1975, respectively, which were comparable to those determined by A549-M (0.9513), HEK293-M (0.9404), and NCI-H1975-M (0.9491), as shown in Table 1 . These results supported the prediction power of the autologous models prepared in this study, which was achieved with total 264 samples for each cell line.

Visualization of learned features

To better investigate the classifiers discovered by deep learning, a visualization of the features linked to their biological meaning was presented, allowing for interpretation of the CNN model decisions. The extracted feature vectors of the test images in high-dimensional features by our trained model were transformed to lower-dimensional features by the well-known t-SNE 13 , as shown in Fig. 4 A. Each dot corresponds to a test image, and the color is the treated DOX concentration. The plot of t-SNE shows the local structure of the high-dimensional input space. OD 450 nm measured by cell proliferation assay according to the DOX concentration in Fig. 4 B represents the concentration-dependent drug responsiveness, which was fitted by a sigmoidal curve generated by the Hill equation (Fig. 4 C). Based on the Hill curve in Fig. 4 C, the IC50 concentration was determined for the concentration at which 50% cell viability was observed compared to that of control cells. Features from cell images corresponding to each concentration formed three clusters after transformation by t-SNE (I, II, and III in Fig. 4 A), which were largely divided in the Hill curve into top (I), slope (II), and bottom (III) regions in Fig. 4 C. The results implied that numerical values of cell viability by visual inspection of cell images could be learned, and that the trained model predicted cell viability in continuous numbers linked to cell images.

Feature visualization learned by MobileNetV2 for the entire dataset. ( A ) 1028-dimensional features of A549 cell images were projected to a 2D surface using t-SNE, and colored according to drug concentrations. ( B ) Responses to drug concentrations (OD 450 nm ). ( C ) Hill curve drawn using the concentrations and the response in ( B ). The image features of the top (I), slope (II), and bottom regions (III) of the Hill curve ( C ) are clustered in different sections of ( A ). The response values from the images in the slope region (II) increase according to the concentration, which is clearly reflected by the feature visualization ( A ).

Another visualization technique is to reveal the localization of discriminative regions for decisions in an input image. Google explainable AI was utilized to understand how our trained model predicted drug responsiveness from cell images. The XRAI option of explainable AI was used to display attributions, which was highlighted in accordance with the prominent image features that were impactful in the model rather than the individual pixels. XRAI highlights the most influential regions in yellow and the least influential in blue, based on the viridian color palette, as shown in Fig. 5 . The localization of the attributions in the input images shows that the prominent features of the decision overlapped with the area of viable cells. This result complemented the t-SNE visualization by localizing important features in the final convolutional layer of the CNN.

Heat map of important features in the input cell images for the decision of cell viability generated using the prediction model A549-M.

Prediction of IC50 using label-free cell images

To calculate the IC50, images corresponding to various drug concentrations were required as input data. We used 11 drug concentrations to treat three types of cells, and captured the images of cells in each well. The histogram of the response values (assigned by cell viability) for the 198 training data samples are shown in Fig. S3 . The responses were not uniformly distributed but covered a sufficiently large area.

Figure 6 shows the measured and predicted IC50 values, which were estimated using autologous and allogenic prediction models. For the three cell types, IC50 pred estimated by the autologous model was closest to IC50 meas for each cell line. The accuracy of IC50 pred reflected the precise cell viability estimated by the CNN models. This result indicated that prediction of drug responsiveness as IC50 was available, using the method of label-free imaging analysis through the numerical deep learning developed in this study.

Distribution of IC50 values from 6-repeated tests of doxorubicin effects. Measured and predicted IC50 for three cell types compared to each other. The results of the statistical analysis are provided in Table 2 .

Table 2 shows the IC50 concentration for DOX exposure obtained from the CCK assay (IC50 meas ) and prediction by CNN models (IC50 pred ) for A549, HEK293, and NCI-H1975 cells. The average difference between IC50 meas and IC50 pred for the autologous model was 0.0935, 0.0015, and 0.0785 for A549, HEK293, and NCI-H1975 cells, respectively. However, when the estimation was performed using the allogenic training model, the difference between IC50 meas and IC50 pred increased from several to hundreds of times over. In addition, unpaired t tests showed that IC50 meas and IC50 pred were not significantly different if autologous models were used to estimate IC50 pred ( p = 0.094, 0.8205, 0.4334 for A549, HEK293, and NCI-H1975, respectively). In other cases, such as IC50 pred of A549 cells using HEK293-M and NCI-H1975-M, the unpaired t test showed that there was a significant difference compared to IC50 meas ( p = 0.001 and p < 0.0001, respectively). For HEK293, IC50 meas , and IC50 pred estimated from allogenic NCI-H1975-M showed no statistically significant difference ( p = 0.1597). This result was attributed to the estimated values of HEK293 IC50 pred using NCI-H1975-M showing wide variation: the difference between IC50 meas and IC50 pred was 3.266 ± 5.266, which was almost a hundred times higher than that of IC50 pred using the autologous model.

In the paired t test, the estimated values of IC50 pred of DOX for HEK293 ( p = 0.7765) and NCI-H1975 ( p = 0.4645) from autologous models were not significantly different compared to IC50 meas values of HEK293 and NCI-H1975 cells, respectively. For A549 cells, p-values for measured and autologous model prediction of IC50 were different in the paired t test ( p = 0.012); however, the difference of IC50 pred to IC50 meas for A549 from allogenic models, HEK293-M and NCI-H1975-M, was dramatically enhanced, which was reflected by statistical parameter, p = 0.0015 and < 0.0001, respectively.

As a result, the predicted IC50 values for DOX exposure were comparable to that of the IC50 values measured by colorimetric assay, because there was no statistical difference between the distribution of IC50 meas and IC50 pred by the paired and unpaired t test (the only exception was the paired t test for A549 cells). The results indicated that the CNN model predicted the cell viability sufficiently precisely to determine drug responsiveness statistically, but the prediction capability could degrade when the numerical learning was performed using cell images from different types of cells.

Table S3 shows IC50 pred of A549 cells determined by three CNN models (MobileNetV2, InceptionV3 and InceptionResNetV2) by fourfold cross validation. When IC50 meas was 0.3747, IC50 pred values were determined as 0.3939, 0.3753, and 0.3163, and the average difference (IC50 pred − IC50 meas ) was 0.01928, 0.00067 and − 0.0535, by MobileNetV2, InceptionV3 and InceptionResNetV2, respectively. In this case, three types of CNN models estimated IC50 pred similar to IC50 meas , without statistical difference ( p > 0.05). Table S4 shows the IC50 pred for A549, HEK293, NCI-H1975, determined by fourfold cross validation using MobileNetV2. The average difference between IC50 meas and IC50 pred (IC50 pred − IC50 meas ) for A549, HEK293, NCI-H1979 were 0.01928, 0.00202, and 0.03274, respectively. The paired and unpaired t test indicated that the distribution of IC50 meas and IC50 pred values was not statistically different ( p > 0.05). These results supported the IC50 prediction capability using A549-M, HEK293-M, and NCI-H1975-M provided in Table 2 .

There are several reports to predict IC50 of anti-cancer drugs using genomic profiles and drug fingerprints 18 , 19 . In these studies, genomic profiles of various cancer cell lines and fingerprints of drugs were trained to categorize IC50 levels in three steps, high responsiveness (class 0), intermediate responsiveness (class 1), and low responsiveness (class 2). The genomic profiles could be an excellent input data to predict IC50 values, because response to drugs in cancers are known to be closely related to genetic variations by pharmacogenomic mechanisms. However, the genomic profiles require time and cost to obtain, compared to the captured images of cells following an exposure to drugs. The present models developed in this study involved simple, cheap, and prompt methods to prepare input datasets, and furthermore, could be exploited in well-established assay methods to determine cell viability and IC50.

We constructed cell viability-based IC50 prediction models for three cell types: A549, HEK293, and NCI-H1975. These cells originated from different tissues and diseases, which have different morphologies and IC50 values. In this context, three models were trained with different input images, and the prediction capabilities were cross-validated to determine the specificity of the models according to the input cell images. The results showed that the model accuracy was dependent on the input data: by using autologous input data with the corresponding model, more accurate prediction results were achieved (r 2 > 0.94).

Recently, a CNN model was shown to predict drug responsiveness in IC50 by examining 3D tumor spheroids to determine responsivity of anticancer drugs such as doxorubicin, oxaliplatin, and irinotecan 7 . The ground truth value of the cell viability was determined by fluorescence intensity followed by LIVE/DEAD cell staining, which was used as the input image layer. In another study, fluorescent LIVE/DEAD cell classification was performed in image processing, using deep learning and taking bright-field images as input 5 . The ground truth label for each cell was set by LIVE/DEAD cell fluorescence staining and paired to the bright-field images, which were used to train the prediction model. In this case, the quantitative assessment of cell growth by drug treatment was not exploited. Until now, non-labeled microscopic images and numerical values determined by conventional cell proliferation assays has not been applied to determine IC50 values by deep-learning approaches. In this study, numerical deep learning models constructed with those input datasets could be proposed as an alternative method to replace colorimetric cell proliferation assays.

In this study, we used over 200 image data samples, which is a relatively low number of images for deep learning. However, the augmentation of data during pre-processing was effective in improving the prediction accuracy of the models. In addition, the visualization of important image features for model decisions supported the classification results by t-SNE and the heat map of the images.

Our model does not require any labeling of probes; thus, using cell images from bright field microscopy of cell culture dishes, drug responsiveness could be determined rapidly and with low cost and labor. The cultured cells used for the estimation of drug responsiveness are still available for further study.

Bykov, Y. S. et al. High-throughput ultrastructure screening using electron microscopy and fluorescent barcoding. J. Cell. Biol. 218 , 2797–2811 (2019).

Article CAS Google Scholar

Collins, T. J. ImageJ for microscopy. Biotechniques 43 , 25–30 (2007).

Article MathSciNet Google Scholar

Liu, Q., Junker, A., Murakami, K. & Hu, P. Automated counting of cancer cells by ensembling deep features. Cells 8 , 1019 (2019).

Rahman, S. et al. Automatic identification of abnormal blood smear images using color and morphology variation of RBCS and central pallor. Comput. Med. Imaging Graph. 87 , 101813 (2021).

Article Google Scholar

Pattarone, G., Acion, L., Simian, M. & Iarussi, E. Learning deep features for dead and living breast cancer cell classification without staining. Sci. Rep. 11 , 10304 (2021).

Article CAS ADS Google Scholar

Akın, Ö., Sultan Belgin, İ, Gökhan, Ş & Yasemin Gülgün, İ. Benchmarking classification models for cell viability on novel cancer image datasets. Curr. Bioinform. 14 , 108–114 (2019).

Zhang, Z. et al. Label-free estimation of therapeutic efficacy on 3D cancer spheres using convolutional neural network image analysis. Anal. Chem. 91 , 14093–14100 (2019).

Christiansen, E. M. et al. In silico labeling: Predicting fluorescent labels in unlabeled images. Cell 173 , 792-803.e19 (2018).

Hussain, Z., Gimenez, F., Yi, D. & Rubin, D. Differential data augmentation techniques for medical imaging classification tasks. AMIA Annu. Symp. Proc. 2017 , 979–984 (2017).

PubMed Google Scholar

Shorten, C. & Khoshgoftaar, T. M. A survey on image data augmentation for deep learning. J. Big Data 6 , 60 (2019).

Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J. & Wojna, Z. Rethinking the inception architecture for computer vision. Preprint at https://arxiv.org/abs/1512.00567 (2015).

Szegedy, C., Ioffe, S., Vanhoucke, V. & Alemi, A. A. Inception-v4, inception-ResNet and the impact of residual connections on learning. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 31 (2017).

Maaten, L. V. D. & Hinton, G. Visualizing data using t-SNE. J. Mach. Learn. Res. 9 , 2579–2605 (2008).

MATH Google Scholar

Kapishnikov, A., Bolukbasi, T., Vi´egas, F. & Terry, M. Xrai: Better attributions through regions. Preprint at https://arxiv.org/abs/1906.02825 (2019)

Sandler, M., Howard, A., Zhu, M., Zhmoginov, A. & Chen, L.-C. Mobilenetv2: Inverted residuals and linear bottlenecks. Preprinted at https://arxiv.org/abs/1801.04381 (2018).

Howard, A. G. et al . Mobilenets: Efficient convolutional neural networks for mobile vision applications. Preprinted at https://arxiv.org/abs/1704.04861 (2017).

Goodfellow, I., Bengio, Y. & Courville, A. DeepLearning (MIT Press, 2016).

Google Scholar

Joo, M. et al. A deep learning model for cell growth inhibition IC50 prediction and its application for gastric cancer patients. Int. J. Mol. Sci. 20 , 6276 (2019).

Lee, Y. & Nam, S. Performance comparisons of AlexNet and GoogLeNet in cell growth inhibition IC50 prediction. Int. J. Mol. Sci. 22 , 7721 (2021).

Download references

Acknowledgements

This study was supported by the Ministry of Science and ICT, Republic of Korea (DGIST Basic Research BT-21-05).

Author information

These authors contributed equally: Kookrae Cho and Eun-Sook Choi.

Authors and Affiliations

Division of Electronics and Information System Research, Daegu Gyeongbuk Institute of Science and Technology (DGIST), Techno-Jungangdaero 333, Daegu, 42988, Republic of Korea

Kookrae Cho, Eun-Sook Choi, Jung-Hee Kim, Jong-Wuk Son & Eunjoo Kim

You can also search for this author in PubMed Google Scholar

Contributions

K.C. and J.W.S. constructed the model. E.S.C., J.H.K., and E.K. performed the cell viability assay and IC50 determination. J.W.S. and E.K. designed the study and wrote the manuscript. All authors reviewed the manuscript.

Corresponding authors

Correspondence to Jong-Wuk Son or Eunjoo Kim .

Ethics declarations

Competing interests.

The authors declare no competing interests.

Additional information

Publisher's note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary Information

Supplementary information., rights and permissions.

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ .

Reprints and permissions

About this article

Cite this article.

Cho, K., Choi, ES., Kim, JH. et al. Numerical learning of deep features from drug-exposed cell images to calculate IC50 without staining. Sci Rep 12 , 6610 (2022). https://doi.org/10.1038/s41598-022-10643-9

Download citation

Received : 26 October 2021

Accepted : 07 April 2022

Published : 22 April 2022

DOI : https://doi.org/10.1038/s41598-022-10643-9

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

This article is cited by

Front-end deep learning web apps development and deployment: a review.

Hock-Ann Goh
Chin-Kuan Ho
Fazly Salleh Abas

Applied Intelligence (2023)

By submitting a comment you agree to abide by our Terms and Community Guidelines . If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.

Quick links

Explore articles by subject
Guide to authors
Editorial policies

Sign up for the Nature Briefing: AI and Robotics newsletter — what matters in AI and robotics research, free to your inbox weekly.

An official website of the United States government

The .gov means it’s official. Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

The site is secure. The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

Publications
Account settings

Preview improvements coming to the PMC website in October 2024. Learn More or Try it out now .

Advanced Search
Journal List
ACS AuthorChoice

When Does the IC 50 Accurately Assess the Blocking Potency of a Drug?

Julio gomis-tena.

† Centro de Investigación e Innovación en Bioingeniería (Ci2B), Universitat Politècnica de València, Camino de Vera, s/n, 46022 Valencia, Spain

Brandon M. Brown

‡ Department of Pharmacology, University of California, Davis, One Shields Avenue, Davis, California 95616-8636, United States

Beatriz Trenor

Pei-chi yang, javier saiz, colleen e. clancy, lucia romero, associated data.

An external file that holds a picture, illustration, etc.
Object name is ci9b01085_0010.jpg

Preclinical assessment of drug-induced proarrhythmicity is typically evaluated by the potency of the drug to block the potassium human ether-à-go-go-related gene (hERG) channels, which is currently quantified by the IC 50 . However, channel block depends on the experimental conditions. Our aim is to improve the evaluation of the blocking potency of drugs by designing experimental stimulation protocols to measure the IC 50 that will help to decide whether the IC 50 is representative enough. We used the state-of-the-art mathematical models of the cardiac electrophysiological activity to design three stimulation protocols that enhance the differences in the probabilities to occupy a certain conformational state of the channel and, therefore, the potential differences in the blocking effects of a compound. We simulated an extensive set of 144 in silico I Kr blockers with different kinetics and affinities to conformational states of the channel and we also experimentally validated our key predictions. Our results show that the IC 50 protocol dependency relied on the tested compounds. Some of them showed no differences or small differences on the IC 50 value, which suggests that the IC 50 could be a good indicator of the blocking potency in these cases. However, others provided highly protocol dependent IC 50 values, which could differ by even 2 orders of magnitude. Moreover, the protocols yielding the maximum IC 50 and minimum IC 50 depended on the drug, which complicates the definition of a “standard” protocol to minimize the influence of the stimulation protocol on the IC 50 measurement in safety pharmacology. As a conclusion, we propose the adoption of our three-protocol IC 50 assay to estimate the potency to block hERG in vitro. If the IC 50 values obtained for a compound are similar, then the IC 50 could be used as an indicator of its blocking potency, otherwise kinetics and state-dependent binding properties should be accounted.

1. Introduction

The rapid component of delayed rectifier current ( I Kr ), which is encoded by the human ether-à-go-go-related gene (hERG), plays an important role on the cardiac action potential (AP) duration. This current is a well-known promiscuous drug target, and many drugs associated with torsade de pointes inhibit the I Kr and hERG channels. 1 Therefore, a key test of the current cardiac safety assessment of pharmacological compounds consists of the observed in vitro block of these channels. 2 This is typically quantified by the IC 50 , which is the drug concentration that blocks 50% of the current. There is experimental evidence of the IC 50 dependency on the experimental conditions, such as voltage stimulus protocol, temperature, and expression system. 3 − 7 Indeed, hERG channel blockers can inhibit the channel by means of different mechanisms, which may exhibit time, voltage, and state dependence. 5 , 8 , 9 However, there is no standardization of these assays at present, which favors the existence of a high variability of the IC 50 values reported in the literature and databases, such as FDA drug labels, PubChem, 10 and DrugBank. 11 A few experimental works have compared the IC 50 values using different voltage protocols and have reported variations in the IC 50 values up to 10-fold when only changing the voltage protocol. 4 − 6 , 12 However, the number of drugs used in these studies was reduced. A very recent investigation of the factors that contribute to the IC 50 differences has been performed using a in silico drug binding and unbinding to the open and inactivated states not allowing drug-bound channels to change their conformational state. 13 With these simple drug–channel interactions, the authors have elegantly shown that state dependence of drug binding is a major determinant of the protocol dependence of I Kr IC 50 . However, that study only considered in silico drug binding and unbinding in the open and/or inactivated states, not in the closed state, despite the existence of compounds, such as ketoconazole and BeKm-1, that preferentially block the channel in the closed state. 5 , 8 , 9 In addition, drug-bound channels in that study were not allowed to change their conformational state, which avoids simulation of drug trapping, a very well-known phenomenon that takes place in the presence of certain drugs. 14 , 15

Here, we attempt to shed light on the relevance of the IC 50 as an indicator of the I Kr blocking potency of a compound and to improve the characterization of its blocking effects using a highly detailed Markov model considering a wide range of drug–channel interactions. We hypothesize that, as the drug–channel interaction may depend on the conformational state of the channel, stimulation at certain voltages where the probability of these states is very different will provide more information about the blocking potency than a unique voltage clamp protocol. In this work, we designed voltage protocols that could unmask distinct state-dependent potencies of block. Then, we systematically carried out “in silico drug genesis” by creating a wide range of virtual drugs with different kinetics and affinities to the conformational states of the I Kr channel. In silico drugs are able to bind and unbind to any conformational state of the channel: closed, open, and/or inactivated. Moreover, two kinds of drug-bound channels were simulated: those that do not change their conformational state and those that do it, which allows the simulation of drug trapping. Next, we obtained the Hill-plots for each virtual drug using our new protocols as well as other existing protocols and calculated the IC 50 s. Finally, we performed some experiments to support our simulation results.

2. Materials and Methods

2.1. drug models.

The human ventricular I Kr was simulated using the five-state Markov chain proposed by Fink et al. 16 This model has five states: three closed states (C3, C2, and C1), an open state (O), and an inactivated state (I). In order to simulate drug interactions with I Kr , we included the new states the channel can occupy in the presence of the drug, namely, C 3d , C 2d , C 1d , O d , and I d . Figure Figure1 1 shows the simulated I Kr Markov model for multiple drug-bound configurations together with the corresponding type drug–channel interaction label. As the ion channel targeting drugs display complex properties determined by preferential binding to distinct conformational states and/or distinct affinity to discrete states, we simulated a wide variety of likely drug–channel interactions: drugs that exclusively interact in the closed ( Figure Figure1 1 A,B), open ( Figure Figure1 1 C,D), or inactivated ( Figure Figure1 1 E,F) states, drugs binding simultaneously to both the closed and open states ( Figure Figure1 1 G,H), or to both the open and inactivated states ( Figure Figure1 1 I,J) and drugs binding simultaneously to all states ( Figure Figure1 1 K,L). We allowed drug-bound channels to change their conformational state ( Figures Figures1 1 A,C,E,G,I,K) as in our previous work, 17 and we labeled them unstuck, but we also considered the possibility that the drug-bound channels do not change their conformational state unless unbinding occurs, and we labeled them stuck ( Figures Figures1 1 B,D,F,H,J,L). Microscopic reversibility was ensured by equaling the product of the rates going clockwise to the product going anticlockwise in closed loops. 18 As drug-bound channels are electrically silent, which precludes the assessment of the transition rates between states, we modified the transition rates from I d to O d and from O d to C 1d when appropriate. Drug kinetics were also analyzed in detail by testing a range of diffusion ( k ) and dissociation rates ( r ) for the various drug configurations. Dissociation rates ranged from 0.001 to 1000 s –1 using logarithmic or half-logarithmic increments, in line with other simulation works, 19 , 20 and the diffusion was the same in all the states, where the drug binds. A total of 144 prototypical drugs were simulated, and their names were generated depending on the states the drug binds and unbinds to and the speed of the dissociation rates. We called Closed, Open, and Inactivated drugs to those binding exclusively to the closed, open, or inactivated states, respectively. We labeled ClosedO, OpenC, and CO the drug binding simultaneously to both the open and closed states with higher affinity to the closed state, to the open state, and with the same affinity, respectively. We labeled OpenI, InactivO, and IO the drugs binding simultaneously to both the open and inactivated states with higher affinity to the open state, to the inactivated state, and with the same affinity, respectively. Finally, we labeled COI, ClosedOI, OpenCI, and InactivOC the drug binding simultaneously all states with the same affinity, with higher affinity to the closed, to the open, and to the inactivated state, respectively. We added the suffixes sss, ss, s, m, f, and ff, depending on the slowest dissociation rate of the drug, which corresponded to 0.001, 0.003, 0.01, 0.1, 1 and 10 s –1 , respectively. Diffusion ( k ) and dissociation ( r ) rate constants for each drug– I Kr interaction as tested in the model are included in the Supporting Information (Tables S1 and S2). Drug doses ranging from 10 –11.7 to 10 –2.7 mol/L (M) with 10 0.1 M steps were simulated for each virtual drug in order to build their respective Hill plots. The temperature was set to 22 or 37 °C and intracellular and extracellular potassium concentrations were fixed to 130 and 4 mM, respectively.

An external file that holds a picture, illustration, etc.
Object name is ci9b01085_0001.jpg

Simulated Markov drug– I Kr interaction models with nondrug-bound (C 3 , C 2 , C 1 , O, and I) and drug-bound (C 3d , C 2d , C 1d , O d , and I d ) states considering unstuck (A,C,E,G,I, and K) and stuck (B,D,F,H,J, and L) drug-bound channels. D is the drug concentration, and its product with k C , k O , and k I corresponds to the association rates constants in the closed, open, and inactivated states, respectively, and r C , r O , and r I are the dissociation rate constants in the closed, open, and inactivated states, respectively. Binding states are red colored. First column indicates the corresponding type of the drug–channel interaction and first row specifies the state of the channel when the drug is bound.

2.2. Simulation of the Pseudo-ECG

Pseudo-ECGs were computed using a one-dimensional (1D) tissue model of a transmural wedge preparation, as in our previous work. 21 The 1D model was composed by 60 endocardial cells, 45 midmyocardial cells, and 60 epicardial cells, each cell being 100 μm long, as defined in the O’Hara et al. model 22 and it was paced at 1 Hz. The propagation of the AP was described by the following nonlinear reaction diffusion equation

where C m stands for the membrane capacitance, a is the radius of the fiber, ∑ I ion is the sum of all the ionic currents flowing through the cellular membrane, and R i represents the intracellular resistivity. Drug blocking effect on I Kr was formulated using the standard sigmoid dose–response curve, parameterized using the half-maximal response dose (IC 50 ), and considering a Hill coefficient of 1 as in previous studies 21 , 23 − 26

where D is the drug concentration and “1 – b ” is the fraction of unblocked channels.

2.3. Experimental Methods

All experiments were conducted manually with an EPC-10 amplifier (HEKA, Lambrecht/Pfalz, Germany) at room temperature in the whole-cell mode of the patch-clamp technique. HEK-293 cells stably expressing hKv11.1 (hERG) under G418 selection were a generous gift from Craig January (University of Wisconsin, Madison). Cells were cultured in Dulbecco’s modified Eagle’s medium containing fetal bovine serum 10%, glutamine 2 mM, Na + pyruvate 1 mM, penicillin 100 U/L, streptomycin 171.94 μM (100 μg/mL), and G418 1 M (500 mg/mL). Before experiments, cells were lifted using TrypLE and plated onto poly- l -lysine-coated coverslips, patch pipettes were pulled from soda lime glass (micro-hematocrit tubes) and had resistances of 2–4 MΩ. We used normal sodium Ringer for the external solution (in mM: NaCl 160, KCl 4.5, CaCl 2 2, MgCl 2 1, HEPES 10 (adjusted to pH 7.4, using HCl and NaOH, and 290–310 mOsm). The internal solution contained (in mM) CaCl 2 5.375, MgCl 2 1.75, EGTA 10, HEPES 10, KCl 120, and NaATP 4 (adjusted to pH 7.2, using HCl and NaOH, and 300–320 mOsm). For all experiments, solutions of dofetilide and moxifloxacin were always freshly prepared from 1, 10, or 100 mM stock solutions in dimethyl sulfoxide (DMSO) during the experiment. The final DMSO concentration never exceeded 1%.

2.4. Stimulation Protocols

Three different sets of voltage clamp protocols were used. The first and third sets were designed in this work while the second was adopted from the literature. The first set was composed of our new stimulation voltage clamp protocols, which consisted of a 5 s variable voltage conditioning step (at −80, 0, and 40 mV) followed by a 0.2 s test pulse at −60 mV repeated at 5.4 s intervals, from a holding potential of −80 mV ( Figure Figure2 2 , top). When the 5 s variable voltage was fixed at −80 mV, a 0.5 ms prepulse at 20 mV was included and the 0.2 s test pulse was applied at −50 mV. These protocols were called P-80, P0, and P40, respectively. The second set was composed of Protocol-O, Protocol-C, and the standard protocol (SP) defined by Yao et al. 2005. 5 Protocol-O consisted of a 4.8 s conditioning step at 20 mV followed by a 0.5 s test pulse at −50 mV repeated at 6 s intervals, from a holding potential of −80 mV. Protocol-C consisted of a 1 s conditioning step at 20 mV followed by a 5 s test pulse at −50 mV repeated at 60 s intervals, from a holding potential of −80 mV. The SP consisted of a 4.8 s conditioning step at 20 mV followed by a 5 s test pulse at −50 mV repeated at 15 s intervals, from a holding potential of −80 mV. The third set of protocols consisted of two AP clamp protocols, P_AP1 and P_AP2, which were generated using a version of the mid-myocardial O’Hara et al. AP model 22 whose I Kr is reduced to 40% at 0.5 and 2 Hz, respectively.

An external file that holds a picture, illustration, etc.
Object name is ci9b01085_0002.jpg

Simulated influence of the voltage of the stimulation protocol on the probabilities of the states of the I Kr channel at 22 °C. Stimulation protocol (top), averages (A) of the simulated probabilities of the closed states ( C AVG , solid line), the open state ( O AVG , dashed line) and the inactivated state ( I AVG , dotted line) for the whole protocol duration as a function of the voltage of the conditioning step ( V m ) and the difference between the average of the simulated probabilities of the closed states and the open state [ C AVG – O AVG , (B)] and the difference between the average of the probabilities of the inactivated state and the open state [ I AVG – O AVG , (C)].

I Kr and hERG channels were stimulated repeatedly until reaching the steady state at pretreatment control and under drug application. Peak tail current amplitudes were measured at steady state and Hill plots were constructed by plotting the steady-state tail peak current normalized to control for each concentration versus the decimal logarithm of the drug concentration, as in previous studies. 5 , 6 , 25 , 27

3.1. Design of Voltage Protocols

As a drug–channel interaction may depend on the conformational state of the channel, and it depends on the membrane voltage, we studied the influence of the voltage of the conditioning step of the stimulation protocol on the probability of the I Kr channel to occupy a specific conformational state using computer simulations. For this purpose, we considered a stimulation voltage clamp that consisted of a 5 s variable voltage ( V m ) conditioning step followed by a 0.2 s test pulse at −60 mV repeated at 5.4 s intervals from a holding potential of −80 mV ( Figure Figure2 2 , top). This protocol was applied in control (absence of drug) at different conditioning step voltages. Then, the average of the probabilities of the three closed states ( C AVG , solid line), the open state ( O AVG , dashed line), and the inactivated state ( I AVG , dotted line) for the whole protocol duration were computed as a function of the conditioning step voltage ( Figure Figure2 2 A). Moreover, the differences C AVG – O AVG ( Figure Figure2 2 B) and I AVG – O AVG ( Figure Figure2 2 C) were also calculated, as these differences will be key to select the conditioning step voltages that will provide more information about the blocking potency of the drug. Indeed, unstuck OpenC drugs are expected to produce the highest block when the stimulation protocol is such that it maximizes the probability of the open state (close to 0 mV, Figure Figure2 2 A, long dashed line) while the probability of the closed state is low. It would occur when the C AVG – O AVG is small and O AVG is relatively high, which would correspond to a conditioning pulse close to 0 mV ( Figure Figure2 2 B). In addition, the lowest inhibition of the channels would occur when the C AVG – O AVG is maximum, which takes place for conditioning pulses at low voltages ( Figure Figure2 2 B). Therefore, the maximum and minimum IC 50 of unstuck OpenC will be expected when applying this protocol with conditioning pulses close to −80 and 0 mV, respectively. For conditioning pulses at higher voltages, such as 40 mV the IC 50 would be expected to be closer to the value obtained with the conditioning pulse at 0 mV. In the case of unstuck ClosedO drugs, the opposite behavior is expected. Regarding drugs with different affinities to the open and inactivated states, as I AVG – O AVG is maximum at 40 mV ( Figure Figure2 2 C), adoption of this voltage for the conditioning pulse would yield high inhibition for unstuck InactivO drugs. Therefore, the application of this protocol with conditioning steps at −80, 0, and 40 mV would highlight the differences in the potency of the block with the voltage. As conditioning steps at −80 mV raised very small currents to be measured in the experiments, we modified this protocol to include a prepulse at 20 mV for 0.5 s to open the channels. These protocols were labeled P-80, P0, and P40, respectively, as indicated in the Materials and Methods section. Figure Figure3 3 A shows a representation of each protocol.

An external file that holds a picture, illustration, etc.
Object name is ci9b01085_0003.jpg

Simulated effects of voltage clamp protocols on IC 50 . Voltage clamp protocols (A) and the corresponding steady-state current traces before and after the application of selected virtual drugs: unstuck Inactivated_s (B), stuck Inactivated_s (C), and stuck ClosedO_s (D) at 22 °C. First column represents the Markovian schemes of the simulated drug– I Kr interactions. Second, third, and fourth columns correspond to the steady-state currents traces elicited for each protocol and arrows indicate peak tail current amplitudes at marked concentrations. Last column illustrates the corresponding Hill plots.

3.2. Simulated Effects of the Voltage Protocol on the IC 50

Once the stimulation protocols were designed, I Kr inhibition produced by all the prototypical drugs was examined using P-80, P0, and P40.

Figure Figure3 3 summarizes the results obtained for three selected drugs: unstuck Inactivated_s ( Figure Figure3 3 B), stuck Inactivated_s ( Figure Figure3 3 C), and stuck ClosedO_s ( Figure Figure3 3 D). The voltage clamp protocols are represented at the top panel ( Figure Figure3 3 A). The Markovian schemes of the simulated drug– I Kr interactions are illustrated in the first column, the steady-state currents traces elicited for each protocol, namely, P-80, P0, and P40, are depicted in the second, third, and fourth column, respectively, and the corresponding Hill plots are constructed in the last column. Unstuck Inactivated_s ( Figure Figure3 3 B) produced similar inhibition of I Kr tail currents with P-80, P0, and P40, so the resulting Hill plot curves are superimposed and the IC 50 s values are the same. Indeed, in the case of unstuck drugs that only bind and unbind to one state, the IC 50 values do not depend on the stimulation protocol, as it is determined by the ratio between the diffusion ( k ) and the “off” rate ( r ). Although the steady-state block is the same for each protocol, the time needed to reach it depends on the voltage protocol as it determines the mean probabilities of the channel of being on each state, and, therefore the average of the time during the cycle to be on the state where the drug can bind and unbind. However, stuck Inactivated_s ( Figure Figure3 3 C) had higher inhibitory effects with protocols P0 (second column) and P40 (third column) than with P-80 (first column), which is consistent with the fact that I AVG is high for P0 and P40 and almost zero for P-80 ( Figure Figure2 2 A, V m = 0, 40 and −80 mV, respectively). For example, 10 nM stuck Inactivated_s inhibited tail currents by approximately 50% with P0 and P40, whereas it only reached approximately 20% with P-80. Subsequently, the Hill plot curves and the IC 50 values corresponding to P0 (red) and P40 (green) are similar while the one corresponding to P-80 (blue) is shifted to the left. Therefore, Hill plots of drugs binding just to one state of the channel were highly dependent on the state of the drug-bound channel. Unstuck variants had the same IC 50 with the three protocols while the stuck ones exhibited the smallest IC 50 with the protocol that enhanced the probability of the state where the drug binds and unbinds; P40, P0, and P-80 for Inactivated ( Figure Figure3 3 C), Open, and Closed drugs, respectively (not shown). Finally, stuck ClosedO_s ( Figure Figure3 3 D) revealed higher potency to block I Kr with P-80, followed by P0, than with P40, so the Hill plot curves as well as the IC 50 values are different. It is in close agreement with the inverse dependency of C AVG and C AVG – O AVG with V m ( Figure Figure2 2 A,B). These results indicate that unstuck Inactivated_s ( Figure Figure3 3 B) produces voltage independent I Kr steady-state blocks. On the contrary, stuck Inactivated_s ( Figure Figure3 3 C) produces smaller I Kr inhibition at low voltages, as it binds and unbinds to the inactivated state, and stuck ClosedO_s ( Figure Figure3 3 D) at high voltages, as it has a preferential affinity to the closed states. Therefore, the dissimilar effects produced by the drugs when applying our set of voltage clamp protocols manifest the differences in drug–channel interactions.

Figure Figure4 4 illustrates the simulated Hill plots for each type of the prototypical drug binding to two states with state-dependent affinities using the proposed protocols: P-80 (blue), P0 (red), and P40 (green) at 22 °C. Both variants of ClosedO_s ( Figure Figure4 4 A) have the minimum IC 50 with P-80, as expected, as more channels are closed at −80 mV, while the maximum IC 50 is registered with P0 or P40. In the case of OpenC_s drugs, the maximum IC 50 is registered with P-80, which maximizes the time the channels are closed and tends to reveal the drug’s affinity to this state. OpenI_s drugs only showed small differences of IC 50 , P0 being the protocol showing the smallest IC 50 , as it is the one that enhances the most the probability of the open state. Finally, the maximum IC 50 of InactivO_s ( Figure Figure4 4 D) is registered with P-80 as this protocol minimizes the probability of the inactivated state, when the affinity of the drug is higher. Drugs with similar state preferences and drug-bound states exhibited similar Hill plot patterns although the maximum IC 50 ratio depended on the value of the slowest dissociation rate of the drug. For example, the maximum IC 50 of InactivO_m also corresponded to P-80 and the IC 50 s obtained with P0 and P40 were very similar, like InactivO_s. However, the maximum IC 50 ratio was 13.0 instead of 23.2, which was the corresponding to InactivO_s ( Figure Figure4 4 D). These results suggest that the influence of the voltage clamp protocol on the estimation of the inhibitory effects of a compound depends on the specific interaction with the channel.

An external file that holds a picture, illustration, etc.
Object name is ci9b01085_0004.jpg

Simulated Hill plots for each type of the prototypical drugs binding to two states with state-dependent affinities using the proposed protocols: P-80 (blue), P0 (red) and P40 (green) at 22 °C. Unstuck (top) and stuck (bottom) variants of ClosedO_s (A), OpenC_s (B), OpenI_s (C), and InactivO_s (D). The maximum IC 50 ratio for each drug is also indicated in each panel.

As this study was extended to the 144 in silico drugs, Hill plots for every prototypical drug were constructed using our proposed protocols (P0, P40, and P-80) and IC 50 values were extracted. Figure Figure5 5 summarizes the maximum IC 50 ratios for each drug–channel interaction. Unstuck and stuck variants are represented with nonfilled and filled bars, respectively. The highest IC 50 ratios were observed for the stuck variants of Closed, ClosedO, and ClosedOI drugs, and some unstuck variants of InactivOC, InactivO, ClosedO, ClosedOI, and OpenC drugs. The highest, mean, and median values of the maximum IC 50 ratio were 51.2, 8.7, and 2.7, respectively. Moreover, 13.9% of the prototypical drugs exhibited a ratio above 20-fold and the 34% yielded a ratio above 10-fold. On the contrary, unstuck drug binding and unbinding to one state (Closed, Open, and Inactivated), two states (CO and IO) or all states with the same affinity (COI) exhibited voltage independent IC 50 s. IC 50 s of stuck drugs whose preferential state for binding and unbinding are the open state (Open, OpenC, OpenI, and OpenCI) showed a very small dependence on the voltage protocol, followed by the unstuck variant of Open_I and both unstuck and stuck variants of InactivO_f, InactivO_ff, and OpenC_ff. Stuck drugs tended to register higher IC 50 ratios than unstuck drugs, the mean maximum IC 50 ratio for stuck drugs being 11.2 while for unstuck drugs being 6.2. However, most unstuck variants of OpenC, OpenI, and InactivO displayed higher IC 50 ratios than the corresponding stuck variants. Finally, the speed of the association and dissociation rates played a relevant role, although their effects were highly drug-dependent. For example, fast rates tended to increase the maximum IC 50 ratio in stuck drugs binding and unbinding to the closed state. By contrast, fast dynamics decreased this ratio in drugs binding simultaneously to both the inactivated and open states with higher affinity to the inactivated state.

An external file that holds a picture, illustration, etc.
Object name is ci9b01085_0005.jpg

Maximum IC 50 ratios obtained with our proposed protocols (P0, P40, and P-80) at 22 °C. Filled (blue and green) and nonfilled (black and red) bars for stuck and unstuck drugs, respectively.

I Kr inhibition produced by all the prototypical drugs was also simulated using the Protocol-O, Protocol-C, and the SP experimentally used by Yao, et al. 2005 5 ( Figure Figure6 6 C). Figure Figure6 6 shows the simulated Hill plots of stuck ClosedO_f obtained with ours (A) and Yao and colleagues’ ones (B). In this case, our protocols provided a maximum IC 50 ratio of 51.5 while the Yao and co-workers’ ones yielded 37.6. Maximum IC 50 ratios obtained with both sets of protocols for all prototypical drugs are provided in the Supporting Information (Figure S1). Maximum, mean, and median values of the maximum IC 50 ratios obtained with Yao and co-workers’ protocols were 37.7, 6.5, and 3.1, respectively, which are smaller than those registered with ours (51.2, 8.7, and 2.7, respectively). Therefore, our new protocols could be more useful than those currently available in the literature to detect those compounds that obstruct the channel to a different extent depending on the stimulation voltage.

An external file that holds a picture, illustration, etc.
Object name is ci9b01085_0006.jpg

Simulated Hill plots for stuck ClosedO_f using our proposed protocols (A) and with Yao et al. (2005) protocols (B) at 22 °C. (A) P0 (red), P40 (green), and P-80 (blue). (B) Protocol-O (P-O, red), Protocol-C (P-C, blue), and SP (green). (C) Yao et al. voltage clamp protocols. 5 The maximum IC 50 ratio for each drug is also indicated in each panel.

Our protocols were also used to simulate Hill plots for every prototypical drug at 35 °C. Although the effects of temperature on binding and unbinding rates of the virtual drugs were not included, our results were temperature-dependent as the formulation of the transition rates between the channel states was temperature-dependent. Absolute and relative to 22 °C maximum IC 50 ratios at 35 °C are provided in the Supporting Information (Figure S2). Maximum IC 50 ratios at 35 °C exhibited a similar tendency to those at 22 °C although important differences were observed. The highest IC 50 ratio at 35 °C was 105.1. The maximum IC 50 ratio that increased the most with the temperature belonged to unstuck ClosedO_s ( Figure Figure7 7 A) while the one that decreased the most corresponded to stuck InactivO_sss ( Figure Figure7 7 B). Temperature-related differences for the other virtual drugs were smaller than two-fold. Therefore, the impact of the voltage protocol on the IC 50 is influenced by temperature, although to a small extent.

An external file that holds a picture, illustration, etc.
Object name is ci9b01085_0007.jpg

Simulated Hill plots for unstuck ClosedO_s (A) and stuck InactivO_sss (B) using the new protocols at 22 °C (top) and 35 °C (bottom). The maximum IC 50 ratio for each drug is also indicated in each panel.

3.3. Experimental Validation

In order to provide an experimental validation to our results, our protocols were applied to construct the experimental Hill plots of two well-known I Kr blockers, moxifloxacin and dofetilide, at 22 °C. The moxifloxacin IC 50 corresponding to P0, P40, and P-80 was 373, 196, and 143 μM, respectively ( Figure Figure8 8 A, left panel), which gives rise to a maximum ratio of 2.6. This ratio is in accordance to the experiments of Alexandrou et al. 2006 28 performed at 22 °C, that provide a maximum ratio of 1.9. A much more dilated influence of the stimulation protocol on dofetilide IC 50 was registered. Hill plots look completely different ( Figure Figure8 8 B, left panel) and disparate IC 50 values are obtained: 57, 193, and 695 nM, which correspond to P0, P40, and P-80, respectively. It yields a maximum ratio of 12.2, which is approximately 3-fold the one calculated from studies, where the only factor that changed was the voltage protocol. 12 Moreover, our experiments support our finding that no stimulation protocol can provide the maximum IC 50 for every drug. Indeed, the P-80 protocol raised the maximum moxifloxacin IC 50 value while P0 provided the minimum, contrarily to dofetilide.

An external file that holds a picture, illustration, etc.
Object name is ci9b01085_0008.jpg

Experimental Hill plots (left column) and simulated steady-state AP of isolated endocardial cells (right columns) for moxifloxacin (top row) and dofetilide (bottom row). Hill plots were obtained using the proposed protocols: P-80 (blue), P0 (red), and P40 (green). Symbols and vertical bars are presented as mean ± standard error of the mean ( n = 4 for all data points). An extra sum-of-squares F test with α set to 0.05 (GraphPad Prism 5; GraphPad Software, La Jolla, CA) was performed to compare the curves to each other (moxifloxacin: P40 vs P0 p = 0.0013, P40 vs P-80 p = 0.1646, P0 vs P-80 p < 0.0001 and dofetilide: P40 vs P0 p = 0.0003, P40 vs P-80 p < 0.0001, P0 vs P-80 p < 0.0001). Simulated steady-state pseudo-ECG in control (black) and in the presence of 196 μM of moxifloxacin and 193 nM of dofetilide considering the IC 50 obtained using the P-80 (blue), P0 (red), and P40 (green).

Therefore, our experiments support the potential use of our protocols to discriminate drugs with a small protocol dependence of the drug block, such as moxifloxacin, from drugs with an enormous dependence, such as dofetilide. Our experiments also corroborate that the maximum IC 50 ratios obtained with our protocols are higher than with previous protocols, and the difficulty to define a unique protocol to assess the I Kr IC 50 for all I Kr blockers.

3.4. Simulated Effects of IC 50 Differences on the QT Interval

In order to show how dissimilar estimates for the IC 50 would affect the prediction of drug-induced QT interval prolongation, pseudo-ECGs were computed in the presence of moxifloxacin and dofetilide. Concentrations of both drugs were fixed to the IC 50 values obtained with P40, as this protocol provided an intermediate IC 50 value for both drugs. Then, the drug block was simulated using the simple pore equation without considering the kinetics and conformational state preference, as done in many previous works. 21 , 23 − 25 Figure Figure8 8 shows that when the estimate of the IC 50 used in the simulations was the one obtained with P40, a 106 ms QT prolongation—from 310 ms in control (black) to 416 ms (green)—was predicted in both cases, as 50% of the channels are closed. However, different QT prolongations were observed when considering the IC 50 estimates obtained with P-80 (blue) and P0 (red). The discrepancies were higher for dofetilide (242 vs 34 ms, bottom row) than for moxifloxacin (134 vs 60 ms, top row), as estimates of IC 50 were more disparate. We also simulated the pseudo-ECGs in the presence of the following therapeutic concentrations: 6.23 μM moxifloxacin and 2 nM dofetilide (see Figure S3 in the Supporting Information ). The predicted QT intervals for moxifloxacin were 318, 319, and 323 ms when using the IC 50 s corresponding to P40, P0, and P-80, respectively, and for dofetilide they were 326, 322, and 317 ms, respectively. Again, the discrepancies were higher for dofetilide (9 ms) than for moxifloxacin (5 ms). Therefore, differences in estimates for the IC 50 involve variances in the prediction of the QT interval.

3.5. Clinical Relevance of the IC 50 s Obtained with the Proposed Stimulation Protocols

The ultimate objective of studying the blocking potency of drugs is to know the effects of the drugs in vivo. As our proposed stimulation protocols are far from the time courses of the membrane potentials in vivo, we also aimed to investigate the drug effects when stimulating the channels with AP waveforms to study whether the blocking effects observed with our three proposed protocols are close to those estimated with more realistic voltage waveforms. For this purpose, we simulated the Hill plots for every prototypical drug with P_AP1 and P_AP2, which correspond to the steady-state APs obtained using a version of the mid-myocardial O’Hara et al. AP model 22 whose I Kr is reduced to 40% at 0.5 and 2 Hz, respectively. Figure Figure9 9 illustrates these AP clamp protocols (A and B) and shows a comparison of the simulated Hill plots with these AP clamps (dotted) and with our three proposed protocols (solid) for each type of the prototypical drugs binding to two states with state-dependent affinities. Our results showed that the curves obtained with P_AP1 were similar to the ones corresponding to P80 while those registered with P_AP2 looked like those obtained with P0. This observation seems reasonable as in P_AP1 the membrane voltage is −80 mV most of the time with short intervals of positive potential and in P_AP2 the membrane voltage is close to 0 mV for a long proportion of the time. These results may lead to the conclusion that P-80 and P0 would be enough to characterize the I Kr block under realistic conditions, P40 being less relevant. However, the IC 50 obtained with P40 could be useful to study the I Kr block in situations that promote channel inactivation. Our results suggest that the blocking potencies observed with our three proposed protocols are in line with the ones that will be exerted under realistic voltage waveforms.

An external file that holds a picture, illustration, etc.
Object name is ci9b01085_0009.jpg

Comparison of the simulated Hill plots obtained with two AP clamp protocols, P_AP1 (dotted blue) and P_AP2 (dotted red), which are illustrated in (A,B), and with our proposed protocols: P-80 (solid blue), P0 (solid red), and P40 (solid green), at 22 °C. Each type of the prototypical drugs binding to two states with state-dependent affinities are represented: unstuck (top) and stuck (bottom) variants of ClosedO_s (C), OpenC_s (D), OpenI_s (E), and InactivO_s (F).

4. Discussion

4.1. main findings.

We developed a computational approach to investigate whether the IC 50 values obtained for a certain drug could be good estimators of the inhibitory effects in vivo and to propose improvements in the assessment of the blocking potency. First, we designed new experimental stimulation protocols to detect different inhibitory potencies depending on the voltage. Second, we simulated a wide variety of I Kr –drug interactions with increasing drug concentrations using the new stimulation protocols. Third, we extracted the IC 50 values for each drug with the new protocols and with others from the literature and calculated the maximum ratio of IC 50 for each drug–protocol combination. Fourth, we performed experiments to support our theoretical observations. Finally, we investigated the drug effects when stimulating the channels with realistic AP waveforms at different frequencies and they were in line with the effects observed with our three new protocols.

Our results revealed that our proposed three-protocol IC 50 assay improves the assessment of the blocking potency of drugs and can be very useful to decide whether the IC 50 values accurately assess the inhibitory effects of the drug in vivo. Our results suggest that when the IC 50 values resulting from applying our three protocols to a compound are similar, then, the IC 50 could be a good indicator, otherwise kinetics and preferential state biding properties should be taken into account to predict the blocking potency of the drug in vivo. Our results revealed a much more pronounced impact of the stimulation protocol on the IC 50 than previous experimental studies. Indeed, the mean and the highest value of the maximum ratio of IC 50 were 8.9 and 105.1, respectively, much higher than 4.3 and 10.3, the corresponding values calculated from experimental studies, where the voltage protocol was the only factor that changed. 5 Our experiments also support that our protocols may yield higher IC 50 differences than other protocols available in the literature. This can be due to two important aspects. First, our protocols were specifically designed to unmask the potential differences in the blocking effects of a compound because of the existence of dissimilarities in the affinities to each conformational state of the hERG channel. In addition, the generation and simulation of a wide variety of dynamic models of the I Kr –drug interaction with very diverse kinetics and affinities to the conformational states of the channel, which is to date hardly possible to achieve experimentally. Importantly, our experiments confirmed that the protocol providing the maximum IC 50 value was drug-specific. This suggests that the adoption of a standard stimulation protocol would dramatically underestimate or overestimate the blocking potency of certain drugs. In our opinion, the use of our three proposed protocols is crucial to build a better picture of the inhibitory effects and the possible clinical outcomes of a compound.

4.2. Impact of the Stimulation Protocol on Blocking Potency Estimation

Some experimental studies have evidenced that the blocking potencies of drugs may vary with the stimulus pattern. Kirsch, et al. 2004 4 used several patch-clamp voltage protocols to study hERG inhibition of 15 drugs. They found differences in the IC 50 for some drugs, the maximum IC 50 ratio being 3.2. Later, Yao et al. in 2005 5 designed two voltage protocols, Protocol-O and Protocol-C, and compared their results with the SP. BeKm-1, a compound that preferentially blocks the channel in the closed, showed the biggest differences in the concentration–response curves. This is in agreement with our simulations, as most of the highest IC 50 ratios correspond to virtual drugs that exclusively or preferentially bind in the closed state (see Figure Figure5 5 ). However, the IC 50 ratios obtained for these drugs in our simulations are higher than 20 (up to 105.1) while the ratio registered for BeKm-1 is 10.3. It corresponds to the ratio between the IC 50 obtained with a SP over the IC 50 obtained with Protocol-O. Protocol-C revealed a smaller block but, unfortunately, the concentration–response curve was incomplete and no IC 50 was provided. Obtaining the full curve could have provided a higher IC 50 ratio.

More recently, Milnes et al. in 2010 12 studied the effects of the stimulation protocol on hERG inhibition for cisapride and dofetilide at 37 °C. They provided maximum IC 50 ratios of 10.3 and 3.75, respectively, when only changing the voltage protocol. The maximum ratio in our experiments with dofetilide is 12.8, which is higher than 3.75. This can be because of the differences on the stimulation protocols and temperature.

Our results also reveal that protocols yielding the maximum IC 50 and minimum IC 50 depend on the drug. Our experiments provided the lowest IC 50 value with P-80 in the case of moxifloxacin and with P0 for dofetilide. Our observation that the protocol revealing the maximum potency of the block is drug-dependent is also supported by Yao et al. 2005. 5

Therefore, our study of the impact of the stimulation protocol on the estimation of current inhibition is in accordance with previous experiments, but it reveals a more critical role of the voltage protocol. A very recent investigation has studied protocol-dependent differences in IC 50 and observed that state preferential binding, drug-binding kinetics and trapping are key factors. 13 Their Markov models included a state-dependent block, but they did not reproduce other important characteristics, such as closed-state trapping. 13 Contrarily, our Markovian models are very comprehensive as they reproduce a state-dependent block, trapping as well as drug binding and unbinding to any state of the channels. Moreover, our models can mimic drug-bound channels changing its conformational state or remaining unchanged.

In order to know if our main results were highly dependent on the ionic channel model, we repeated some key simulations using two additional formulations of the hERG channel: Lee et al. 19 and Li et al. 29 models. These two Markovian models have distinct structures and transition rates, which are also different from the Fink et al. model. Figures S4 and S5 of the Supporting Information represent the Markovian schemes (left column) and the simulated Hill plots for each type of the prototypical drugs binding to two states with state-dependent affinities using the proposed protocols: P-80 (blue), P0 (red), and P40 (green) at 22 °C using Lee et al. and Li et al. hERG models, respectively. These figures show the simulated Hill plots for each type of the prototypical drug binding to the two states with state-dependent affinities using the proposed protocols, as in Figure Figure4 4 , where they were simulated using the Fink et al. model. 16 The patterns of the Hill plots obtained with the three models were very similar, although there are quantitative differences that affect the values of the maximum IC 50 ratios. In the three cases, the IC 50 protocol dependency relied on the tested compounds and the protocols yielding the maximum IC 50 and minimum IC 50 depended on the drug with the three ionic models. Indeed, both variants of ClosedO_s ( Figures Figures4 4 A, S4B and S5B ) have the minimum IC 50 with P-80 and the IC 50 registered with P0 and P40 are substantially higher. Also, in the case of OpenC_s drugs ( Figures Figures4 4 B, S4C and S5C ), the maximum IC 50 was registered with P-80 and the minimum with P0, which is similar to the one obtained with P40. In addition, OpenI_s drugs showed small differences of IC 50 with the three models ( Figures Figures4 4 C, S4E and S5E ), the minimum IC 50 being obtained with P0, although it could be very similar to the ones registered with the other protocols. Moreover, the maximum IC 50 of InactivO_s was registered with P-80 with the three models ( Figures Figures4 4 D, S4F and S5F ), and the IC 50 values obtained with P0 and P40 were similar. We have also obtained the Hill plots of unstuck and stuck Inactivated_s with Lee et al. and Li et al. models (see middle and bottom rows of Figure S6 of the Supporting Information ) and they clearly resemble the ones obtained with the Fink et al. model (top row of Figure S6 and right panels of Figure Figure3 3 B,C), the unstuck variant having the same IC 50 for the three protocols. Therefore, there are also drugs that showed no differences or small differences on the IC 50 value when simulated with Lee et al. and Li et al. models. Overall, the main conclusions of this work hold when the ionic channel model is simulated with Lee et al. or the Li et al. models, which have different structures and transition rates from the Fink et al. model, which suggests that the main conclusions of this work are not dependent of the ionic model used.

4.3. Implications for Drug Safety Assessment

Our work has important implications for drug safety assessment. Indeed, one of the most relevant cardiac safety tests of pharmacological compounds consists of the measurement of hERG IC 50 in vitro. 2 As previously explained, other authors have shown differences on IC 50 values, but they were smaller than in our work, and some of these authors considered that the use of a certain protocol could be enough for safety studies. 4 , 5 However, different protocols and temperatures are proposed. Kirsch and co-workers propose a step-ramp protocol at near-physiological temperatures, 4 while Yao and colleagues propose the long-pulse step protocol at room temperature. 5 More recently, the comprehensive in vitro proarrhythmia assay initiative, led by the FDA, has raised the need of a standardization of the experiments used to obtain the IC 50 values. 30 However, our results suggest that the existence of a wide variety of drug–channel interactions impairs the definition of a “standard” protocol to minimize the influence of the stimulation protocol on the IC 50 measurement. In order to improve the assessment of drug safety, we suggest the adoption of a three-protocol IC 50 assay. Provided that the differences in IC 50 for a compound are small enough, the IC 50 could be used for the assessment of the inhibitory effects of the compound. On the contrary, supposing the IC 50 s resulted in very different values, the IC 50 would be a poor indicator. Then, other characteristics, like kinetics and state-dependent binding properties should be investigated to have a better picture of the blocking effects of the compound.

Although the proposed protocols do not correspond to electrophysiological conditions, our simulations with the AP clamp protocols have shown that the Hill plots obtained with P-80 are close to those obtained with P_AP1 and P0 with P_AP2 which come from voltage membrane time courses of cells with a reduced repolarization reserve at slow and fast pacing, respectively. Therefore, the IC 50 s obtained from our protocols would be related to the blocking potencies of the drugs in vivo. However, considering only these two IC 50 values would be an oversimplification, as electrical activity is very different during arrhythmic episodes or in the presence of pathologies, like hypo or hyperkalemia, ischemia, or heart failure. In addition, the AP waveform is not uniform in the heart. There are apico-basal and transmural differences. Purkinje AP time courses are also different from ventricular AP time courses and there is a natural intersubject variability. These reasons led us to try to design protocols to infer the drug potency in each conformational state of the channel. We designed P-80, P0, and P40 to investigate the drug block in the closed, open, and inactivated states, respectively. Although at 0 mV not all channels are open, the open probability is relatively high at that voltage. If the IC 50 s obtained with the three protocols are similar, we can assume that the channel block that can occur in any real situation will be similar. On the contrary, if the values are disparate, the channel block produced by the drug may be extremely dependent on the situation.

Recent works propose alternative methods to assess the proarrhythmic risk of drugs by using the modeling and simulation of drug–channel interactions and considering the kinetics of block. 31 − 33 Some authors have even attempted to implement a standardized protocol for the measurement of kinetics and potency of the hERG block. Unfortunately, their results highlight the challenges in identifying it over a range of kinetics. 34 We also agree that drug safety assessment would improve by considering the kinetics of the block, but, to the best of our knowledge, most pharmaceutical companies are not constructing mathematical models of drug–hERG interactions based on the current block measured using a dynamic voltage protocol, which seems to require a substantial time. Formulation of mathematical models describing drug–channel interactions is not an easy work. Even the authors proposing this method obtain different models depending on the seed used to fit the model, 32 which may lead to different predictions. In addition, drugs may bind and unbind the channel by many mechanisms and, as far as we are concerned, only a few possible types of drug–channel interactions are being accounted for in these attempts. Indeed, their Markovian models do not consider the possibility of the drug binding and unbinding to any channel state and their simulated drug-bound channels have less conformational states than the unbound channels. Therefore, only a few types of drug–channel interactions are considered in these attempts. The above-mentioned restrictions reduce the number of parameters to be fitted in the process of drug model development and simplify it. However, it can also lead to a misunderstanding of the mechanism of drug–channel interaction, which can result in unrealistic predictions of the effects of the compounds. Therefore, we suggest the application in the industry of the protocols designed here. If the three IC 50 values are similar, then IC 50 is a good indicator of the blocking effects of the compound and it can be used to predict its proarrhythmic risk, by using the Tx index 21 for example. Otherwise, the study of the kinetics and state-dependent binding would be needed to better characterize it, and the formulation of mathematical models describing drug–channel interactions would be worthy.

4.4. Limitations

Our work suggests the use of three voltage protocols instead of one when assessing the blocking potential of drugs. We have applied them to a wide range of virtual drugs and to two off-the-shelf drugs. Although it is not possible to experimentally reproduce our simulations, our work would also benefit from experiments with more types of drugs.

We have accounted for the effect of the temperature on the transition rates between the channel states. However, the influence of temperature on binding and unbinding rates of the virtual drugs has not been included as there is not a universal dependence followed by all compounds.

It is to mention that there are factors affecting data interpretation in ligand binding assays under equilibrium conditions that must be considered when designing and performing experiments to obtain Hill plot curves, such as ligand depletion, nonattainment of equilibrium, buffer composition, and the temperature at which the assay is conducted. 35

All in all, we believe that our three-protocol hERG-IC 50 assay would improve the evaluation of the proarrhythmic risk of drugs in the early stages of drug development.

5. Conclusions

Our work reveals that the evaluation of the blocking potency of drugs in the early stages of drug development could be improved by the use of our three-protocol hERG-IC 50 assay, which was designed to reveal the dissimilarities in the affinity of the drug to the different conformational states of the channel. Our results show that the influence of the stimulation protocol on IC 50 evaluation depends on the specific I Kr –drug interaction. In some cases, the three IC 50 values registered for a compound are the same or very similar, then, the IC 50 could be used as an estimator of the inhibitory potency. However, in other cases, the IC 50 estimated by two different protocols could vary as much as 2 orders of magnitude. Then, kinetics and state-dependent properties would also be necessary to predict drug effects. Importantly, as the protocol that provided the maximum IC 50 was specific to the drug, the design of a “standard” protocol that provides a representative IC 50 value for any compound becomes pointless. To sum up, adoption of our hERG-IC 50 assay on the methods of routinely evaluating the effects of a drug on hERG channels on safety pharmacology would ultimately result in more accurate clinical predictions.

Acknowledgments

This work was partially supported by the Dirección General de Política Científica de la Generalitat Valenciana (PROMETEU2016/088), Ministerio de Economía y Competitividad, and Fondo Europeo de Desarrollo Regional (FEDER) DPI2015-69125-R (MINECO/FEDER, UE) as well as Ayuda a Primeros Proyectos de Investigación (PAID-06-18), Vicerrectorado de Investigación. Innovación y Transferencia de la Universitat Politècnica de València (UPV), València, Spain. The research was also supported by the National Institutes of Health 1R01HL128537-01A1.

Abbreviations

	the average of the probabilities of the three-closed states
CiPA	comprehensive in vitro pro-arrhythmia assay
ClosedO	drug binding simultaneously to both the open and closed states with higher affinity to the closed state
CO	drug binding simultaneously to both the open and closed states with the same affinity to both states
COI	drug binding simultaneously all states with the same affinity
ClosedOI	drug binding simultaneously with higher affinity to the closed state
hERG	human ether-à-go-go-related gene
	the average of the probability of the inactivated state
InactivO	drug binding simultaneously to both the inactivated and open states with higher affinity to the inactivated state
InactivOC	drug binding simultaneously all states with higher affinity to the inactivated state
IO	drug binding simultaneously to both the inactivated and open states with the same affinity to both states
IC	drug concentration that obstructs the 50% of the channels
	rapid component of the delayed rectifier current
	the average of the probability of the open state
OpenC	drug binding simultaneously to both the open and closed states with higher affinity to the open state
OpenCI	drug binding simultaneously all states with higher affinity to the inactivated state
OpenI	drug binding simultaneously to both the open and inactivated states with higher affinity to the open state
stuck drug	drug that does not allow bound channels to change their conformational state unless unbinding occurs
TdP	torsade de pointes
unstuck drug	drug that allows bound channels to change their conformational

Supporting Information Available

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.jcim.9b01085 .

Maximum IC 50 ratios for each simulated drug obtained with our protocols and with Yao, et al. 2005 protocols at 22 °C; maximum IC 50 ratios obtained with our proposed protocols at 35 °C and comparison with 22 °C; simulated steady state pseudo-ECGs for moxifloxacin and dofetilide; simulated Hill plots for each type of the prototypical drugs binding to two states with state-dependent affinities using the proposed protocols at 22 °C using the Lee et al. hERG model; 19 simulated Hill plots for each type of the prototypical drugs binding to two states with state-dependent affinities using the proposed protocols at 22 °C using the Li et al. hERG model; 29 and simulated Hill plots for unstuck and stuck Inactivated_s using the three ionic channel models: Fink et al., 16 Lee et al. model, 19 and Li et al. 29 ( PDF )

Author Contributions

The manuscript was written through contributions of all authors. All authors have given approval to the final version of the manuscript.

The authors declare no competing financial interest.

Supplementary Material

Gintant G. A. Preclinical Torsades-de-Pointes Screens: Advantages and Limitations of Surrogate and Direct Approaches in Evaluating Proarrhythmic Risk . Pharmacol. Ther. 2008, 119 , 199–209. 10.1016/j.pharmthera.2008.04.010. [ PubMed ] [ CrossRef ] [ Google Scholar ]
International Conference on Harmonisation; Guidance on S7B Nonclinical Evaluation of the Potential for Delayed Ventricular Repolarization (QT Interval Prolongation) by Human Pharmaceuticals . Fed. Regist. 2005, 70 , 61133–61134. [ PubMed ] [ Google Scholar ]
Witchel H. J.; Milnes J. T.; Mitcheson J. S.; Hancox J. C. Troubleshooting Problems with in Vitro Screening of Drugs for QT Interval Prolongation Using HERG K+ Channels Expressed in Mammalian Cell Lines and Xenopus Oocytes . J. Pharmacol. Toxicol. Methods 2002, 48 , 65–80. 10.1016/s1056-8719(03)00041-8. [ PubMed ] [ CrossRef ] [ Google Scholar ]
Kirsch G. E.; Trepakova E. S.; Brimecombe J. C.; Sidach S. S.; Erickson H. D.; Kochan M. C.; Shyjka L. M.; Lacerda A. E.; Brown A. M. Variability in the Measurement of HERG Potassium Channel Inhibition: Effects of Temperature and Stimulus Pattern . J. Pharmacol. Toxicol. Methods 2004, 50 , 93–101. 10.1016/j.vascn.2004.06.003. [ PubMed ] [ CrossRef ] [ Google Scholar ]
Yao J.-A.; Du X.; Lu D.; Baker R. L.; Daharsh E.; Atterson P. Estimation of Potency of HERG Channel Blockers: Impact of Voltage Protocol and Temperature . J. Pharmacol. Toxicol. Methods 2005, 52 , 146–153. 10.1016/j.vascn.2005.04.008. [ PubMed ] [ CrossRef ] [ Google Scholar ]
Stork D.; Timin E. N.; Berjukow S.; Huber C.; Hohaus A.; Auer M.; Hering S. State Dependent Dissociation of HERG Channel Inhibitors . Br. J. Pharmacol. 2009, 151 , 1368–1376. 10.1038/sj.bjp.0707356. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]
Hancox J. C.; McPate M. J.; El Harchi A.; Zhang Y. H. The HERG Potassium Channel and HERG Screening for Drug-Induced Torsades de Pointes . Pharmacol. Ther. 2008, 119 , 118–132. 10.1016/j.pharmthera.2008.05.009. [ PubMed ] [ CrossRef ] [ Google Scholar ]
Dumaine R.; Roy M.-L.; Brown A. M. Blockade of HERG and Kv1.5 by Ketoconazole . J. Pharmacol. Exp. Ther. 1998, 286 , 727–735. [ PubMed ] [ Google Scholar ]
Milnes J. T.; Dempsey C. E.; Ridley J. M.; Crociani O.; Arcangeli A.; Hancox J. C.; Witchel H. J. Preferential Closed Channel Blockade of HERG Potassium Currents by Chemically Synthesised BeKm-1 Scorpion Toxin . FEBS Lett. 2003, 547 , 20–26. 10.1016/s0014-5793(03)00662-8. [ PubMed ] [ CrossRef ] [ Google Scholar ]
Kim S.; Thiessen P. A.; Bolton E. E.; Chen J.; Fu G.; Gindulyte A.; Han L.; He J.; He S.; Shoemaker B. A.; Wang J.; Yu B.; Zhang J.; Bryant S. H. PubChem Substance and Compound databases . Nucleic Acids Res. 2016, 44 , D1202–D1213. 10.1093/nar/gkv951. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]
Wishart D. S.; Feunang Y. D.; Guo A. C.; Lo E. J.; Marcu A.; Grant J. R.; Sajed T.; Johnson D.; Li C.; Sayeeda Z.; Assempour N.; Iynkkaran I.; Liu Y.; Maciejewski A.; Gale N.; Wilson A.; Chin L.; Cummings R.; Le D.; Pon A.; Knox C.; Wilson M. DrugBank 5.0: A Major Update to the DrugBank Database for 2018 . Nucleic Acids Res. 2018, 46 , D1074–D1082. 10.1093/nar/gkx1037. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]
Milnes J. T.; Witchel H. J.; Leaney J. L.; Leishman D. J.; Hancox J. C. Investigating dynamic protocol-dependence of hERG potassium channel inhibition at 37 °C: Cisapride versus dofetilide . J. Pharmacol. Toxicol. Methods 2010, 61 , 178–191. 10.1016/j.vascn.2010.02.007. [ PubMed ] [ CrossRef ] [ Google Scholar ]
Lee W.; Windley M. J.; Perry M. D.; Vandenberg J. I.; Hill A. P. Protocol-Dependent Differences in IC50 Values Measured in Human Ether-Á-Go-Go-Related Gene Assays Occur in a Predictable Way and Can Be Used to Quantify State Preference of Drug Binding . Mol. Pharmacol. 2019, 95 , 537–550. 10.1124/mol.118.115220. [ PubMed ] [ CrossRef ] [ Google Scholar ]
Carmeliet E. Voltage- and Time-Dependent Block of the Delayed K+ Current in Cardiac Myocytes by Dofetilide . J. Pharmacol. Exp. Ther. 1992, 262 , 809–817. [ PubMed ] [ Google Scholar ]
Mitcheson J. S.; Chen J.; Sanguinetti M. C. Trapping of a Methanesulfonanilide by Closure of the HERG Potassium Channel Activation Gate . J. Gen. Physiol. 2000, 115 , 229–240. 10.1085/jgp.115.3.229. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]
Fink M.; Noble D.; Virag L.; Varro A.; Giles W. R. Contributions of HERG K+ Current to Repolarization of the Human Ventricular Action Potential . Prog. Biophys. Mol. Biol. 2008, 96 , 357–376. 10.1016/j.pbiomolbio.2007.07.011. [ PubMed ] [ CrossRef ] [ Google Scholar ]
Romero L.; Trenor B.; Yang P.-C.; Saiz J.; Clancy C. E. In silico screening of the impact of hERG channel kinetic abnormalities on channel block and susceptibility to acquired long QT syndrome . J. Mol. Cell. Cardiol. 2015, 87 , 271–282. 10.1016/j.yjmcc.2015.08.015. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]
Colquhoun D.; Dowsland K. A.; Beato M.; Plested A. J. R. How to Impose Microscopic Reversibility in Complex Reaction Mechanisms . Biophys. J. 2004, 86 , 3510–3518. 10.1529/biophysj.103.038679. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]
Lee W.; Mann S. A.; Windley M. J.; Imtiaz M. S.; Vandenberg J. I.; Hill A. P. In silico assessment of kinetics and state dependent binding properties of drugs causing acquired LQTS . Prog. Biophys. Mol. Biol. 2016, 120 , 89–99. 10.1016/j.pbiomolbio.2015.12.005. [ PubMed ] [ CrossRef ] [ Google Scholar ]
Ellinwood N.; Dobrev D.; Morotti S.; Grandi E. Revealing Kinetics and State-Dependent Binding Properties of IKur-Targeting Drugs That Maximize Atrial Fibrillation Selectivity . Chaos 2017, 27 , 093918. 10.1063/1.5000226. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]
Romero L.; Cano J.; Gomis-Tena J.; Trenor B.; Sanz F.; Pastor M.; Saiz J. In Silico QT and APD Prolongation Assay for Early Screening of Drug-Induced Proarrhythmic Risk . J. Chem. Inf. Model. 2018, 58 , 867–878. 10.1021/acs.jcim.7b00440. [ PubMed ] [ CrossRef ] [ Google Scholar ]
O’Hara T.; Virág L.; Varró A.; Rudy Y. Simulation of the Undiseased Human Cardiac Ventricular Action Potential: Model Formulation and Experimental Validation . PLoS Comput. Biol. 2011, 7 , e1002061 10.1371/journal.pcbi.1002061. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]
Mirams G. R.; Cui Y.; Sher A.; Fink M.; Cooper J.; Heath B. M.; McMahon N. C.; Gavaghan D. J.; Noble D. Simulation of Multiple Ion Channel Block Provides Improved Early Prediction of Compounds’ Clinical Torsadogenic Risk . Cardiovasc. Res. 2011, 91 , 53–61. 10.1093/cvr/cvr044. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]
Obiol-Pardo C.; Gomis-Tena J.; Sanz F.; Saiz J.; Pastor M. A Multiscale Simulation System for the Prediction of Drug-Induced Cardiotoxicity . J. Chem. Inf. Model. 2011, 51 , 483–492. 10.1021/ci100423z. [ PubMed ] [ CrossRef ] [ Google Scholar ]
Mirams G. R.; Davies M. R.; Brough S. J.; Bridgland-Taylor M. H.; Cui Y.; Gavaghan D. J.; Abi-Gerges N. Prediction of Thorough QT Study Results Using Action Potential Simulations Based on Ion Channel Screens . J. Pharmacol. Toxicol. Methods 2014, 70 , 246–254. 10.1016/j.vascn.2014.07.002. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]
Lancaster M. C.; Sobie E. Improved Prediction of Drug-Induced Torsades de Pointes Through Simulations of Dynamics and Machine Learning Algorithms . Clin. Pharmacol. Ther. 2016, 100 , 371–379. 10.1002/cpt.367. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]
Elkins R. C.; Davies M. R.; Brough S. J.; Gavaghan D. J.; Cui Y.; Abi-Gerges N.; Mirams G. R. Variability in High-Throughput Ion-Channel Screening Data and Consequences for Cardiac Safety Assessment . J. Pharmacol. Toxicol. Methods 2013, 68 , 112–122. 10.1016/j.vascn.2013.04.007. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]
Alexandrou A. J.; Duncan R. S.; Sullivan A.; Hancox J. C.; Leishman D. J.; Witchel H. J.; Leaney J. L. Mechanism of hERG K+channel blockade by the fluoroquinolone antibiotic moxifloxacin . Br. J. Pharmacol. 2006, 147 , 905–916. 10.1038/sj.bjp.0706678. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]
Li Z.; Dutta S.; Sheng J.; Tran P. N.; Wu W.; Colatsky T. A Temperature-Dependent in Silico Model of the Human Ether-à-Go-Go-Related (HERG) Gene Channel . J. Pharmacol. Toxicol. Methods 2016, 81 , 233–239. 10.1016/j.vascn.2016.05.005. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]
Sager P. T.; Gintant G.; Turner J. R.; Pettit S.; Stockbridge N. Rechanneling the Cardiac Proarrhythmia Safety Paradigm: A Meeting Report from the Cardiac Safety Research Consortium . Am. Heart J. 2014, 167 , 292–300. 10.1016/j.ahj.2013.11.004. [ PubMed ] [ CrossRef ] [ Google Scholar ]
Windley M. J.; Mann S. A.; Vandenberg J. I.; Hill A. P. Temperature Effects on Kinetics of Kv11.1 Drug Block Have Important Consequences for in Silico Proarrhythmic Risk Prediction . Mol. Pharmacol. 2016, 90 , 1–11. 10.1124/mol.115.103127. [ PubMed ] [ CrossRef ] [ Google Scholar ]
Li Z.; Dutta S.; Sheng J.; Tran P. N.; Wu W.; Chang K.; Mdluli T.; Strauss D. G.; Colatsky T. Improving the In Silico Assessment of Proarrhythmia Risk by Combining HERG (Human Ether-à-Go-Go-Related Gene) Channel-Drug Binding Kinetics and Multichannel Pharmacology . Circ.: Arrhythmia Electrophysiol. 2017, 10 , e004628 10.1161/circep.116.004628. [ PubMed ] [ CrossRef ] [ Google Scholar ]
Dutta S.; Chang K. C.; Beattie K. A.; Sheng J.; Tran P. N.; Wu W. W.; Wu M.; Strauss D. G.; Colatsky T.; Li Z. Optimization of an In Silico Cardiac Cell Model for Proarrhythmia Risk Assessment . Front. Physiol. 2017, 8 , 616. 10.3389/fphys.2017.00616. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]
Windley M. J.; Abi-Gerges N.; Fermini B.; Hancox J. C.; Vandenberg J. I.; Hill A. P. Measuring Kinetics and Potency of HERG Block for CiPA . J. Pharmacol. Toxicol. Methods 2017, 87 , 99–107. 10.1016/j.vascn.2017.02.017. [ PubMed ] [ CrossRef ] [ Google Scholar ]
Hulme E. C.; Trevethick M. A. Ligand Binding Assays at Equilibrium: Validation and Interpretation . Br. J. Pharmacol. 2010, 161 , 1219–1237. 10.1111/j.1476-5381.2009.00604.x. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]

IMAGES

IC50 Values for Paclitaxel and Analogs in Cytotoxicity Assays with
In vitro cytotoxicity assay. A) IC50 value of Pyrolipid and
IC50 determination using JC-1 and/or pLDH in parallel on drug treatedP
the best way to calculate the IC50 using graphpad prism 8
Half-maximal concentration of peptide (IC50) required to inhibit the
Example: IC50 Curve Calculation

COMMENTS

IC50
Half maximal inhibitory concentration ( IC50) is a measure of the potency of a substance in inhibiting a specific biological or biochemical function. IC 50 is a quantitative measure that indicates how much of a particular inhibitory substance (e.g. drug) is needed to inhibit, in vitro, a given biological process or biological component by 50%. [1]
Guidelines for accurate EC50/IC50 estimation
The absolute EC50/IC50 is the response corresponding to the 50% control (the mean of the 0% and 100% assay controls). The guidelines first describe how to decide whether to use the relative EC50/IC50 or the absolute EC50/IC50. Assays for which there is no stable 100% control must use the relative EC50/IC50.
Determination of half-maximal inhibitory concentration using ...
Half-maximal inhibitory concentration (IC50) is the most widely used and informative measure of a drug's efficacy. It indicates how much drug is needed to inhibit a biological process by half, thus providing a measure of potency of an antagonist drug in pharmacological research. Most approaches to d …
A simple and inexpensive quantitative technique for ...
As an alternative approach IC50 values can be determined in Excel. To determine IC50, first a dose dependent curve is generated by plotting ercent of growth versus concentration in log 10 scale ...
Comparison of Drug Inhibitory Effects (IC50) in Monolayer and Spheroid
Traditionally, the monolayer (two-dimensional) cell cultures are used for initial evaluation of the effectiveness of anticancer drugs. In particular, these experiments provide the IC 50 curves that determine drug concentration that can inhibit growth of a tumor colony by half when compared to the cells grown with no exposure to the drug. Low IC 50 value means that the drug is effective at low ...
The changing 50% inhibitory concentration (IC50) of cisplatin: a pilot
INTRODUCTION. Determination of the half-maximal (50%) inhibitory concentration (IC 50) is essential for understanding the pharmacological and biological characteristics of a chemotherapeutic agent [1, 2].Since the invention of a colorimetric technique - the 3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide (MTT) assay - the process used for IC 50 determination has become easier ...
The IC(50) concept revisited
The ability to accurately measure the concentration of the inhibitor which is required to inhibit a given biological or biochemical function by half is extremely important in ranking compounds. Since the concept of the half maximal inhibitory concentration (IC (50)) is used extensively for studying reversible inhibition enzymatic reactions, it ...
Multilevel models improve precision and speed of IC50 estimates
Here we presented a multi-level model that estimates the IC50 values for all cell lines and compounds of a drug screening simultaneously. This model outperform the Bayesian model used in Garnett et al.*. More specifically, it outputs fewer extreme IC50 values, more precise IC50 estimates at a computational speed that is 1,000 times faster.
Guidelines for accurate EC50/IC50 estimation
The relative EC50/IC50 is the parameter cin the 4-parameter logistic model. and is the concentration corresponding to a response midwa y between the estimates of the lower and upper plateaus. The ...
When Does the IC50 Accurately Assess the Blocking Potency of a Drug?
ACCESS. ABSTRACT: Preclinical assessment of drug-induced proarrhyth-micity is typically evaluated by the potency of the drug to block the potassium human ether-à -go-go-related gene (hERG) channels, which is currently quantified by the IC50. However, channel block depends on the experimental conditions.
PDF IC50 Determination
IC50 values are determined through a series of experiments. For all experiments, a high, constant concentration of substrate is present so that the enzyme can react at an appreciable rate. In each experiment, the amount of inhibitor (log [I]) is steadily increased, and the observed rate of the reaction (V) decreases accordingly. The various ...
Gene-signature-derived IC50s/EC50s reflect the potency of causative
Before compound treatment and for all experiments, the THP1 cells were differentiated with 100 nM Vitamin D 3 (Biotrend Chemicals AG, Switzerland, Cat. No. BG0684) for 3 days at 37 °C/CO 2. cAMP ...
Guidelines for accurate EC50/IC50 estimation
The absolute EC50/IC50 is the response corresponding to the 50% control (the mean of the 0% and 100% assay controls). The guidelines first describe how to decide whether to use the relative EC50/IC50 or the absolute EC50/IC50. Assays for which there is no stable 100% control must use the relative EC50/IC50.
Optimization of cell viability assays to improve replicability and
Cell viability was negatively affected by evaporation and DMSO solvent. To evaluate the effect suboptimal experimental design has on cell viability (IC50) and the area under the dose-response ...
Theoretical and experimental relationships between percent ...
The four-parameter logistic Hill equation models the theoretical relationship between inhibitor concentration and response and is used to derive IC(50) values as a measure of compound potency. This relationship is the basis for screening strategies that first measure percent inhibition at a single, …
IC50 or cell viability experiment
This video shows you the basic steps required to perform a IC50 or cell viability experiment for your new treatment (e.g. new drug, nanomaterial) and also ho...
IC50's: An Approach to High-Throughput Drug Discovery
IC50's in High-Throughput. IC 50 's are used in lab settings to estimate the quantities of these drugs that result in a 50% inhibition of a biological process/mechanism or drug interaction specifically in cultured cells. 1 These are especially useful for studies looking at novel situations of adding one drug to the body to blunt the effect ...
The Relationship between the IC50 Values and the Apparent Inhibition
An experimental design is proposed together with data analysis that allows several aspects to be determined with a few experiments—the IC 50 value, the type of inhibition and the K I app value. This design is described and carried out with data obtained by simulating the different mechanisms under study.
An examination of IC50 and IC50-shift experiments in assessing time
An IC50 fold-shift of >1.5 indicated significant TDI potency. Further, the "shifted IC50" could be used to estimate, the K(I) and TDI … The relationship between time-dependent inactivation (TDI) and IC50 is examined using a consolidated method for evaluating CYP450 inhibition during drug discovery.
Guidelines for accurate EC50/IC50 estimation
The absolute EC50/IC50 is the response corresponding to the 50% control (the mean of the 0% and 100% assay controls). The guidelines first describe how to decide whether to use the relative EC50/IC50 or the absolute EC50/IC50. Assays for which there is no stable 100% control must use the relative EC50/IC50.
Use of Different Parameters and Equations for Calculation of IC50
The plates were returned to the incubator on a plate shaker for agitation during the transport experiment. At the time points of 30, 60, 90, and 120 min, 0.5 mL samples were collected from the AP chamber and replenished with an equal volume of HBSS/MES buffer, with or without inhibitor. A sample was taken from the donor chamber at 120 min, and ...
Numerical learning of deep features from drug-exposed cell images to
This result was attributed to the estimated values of HEK293 IC50 pred using NCI-H1975-M showing wide variation: the difference between IC50 meas and IC50 pred was 3.266 ± 5.266, which was almost ...
When Does the IC50 Accurately Assess the Blocking Potency of a Drug?
1. Introduction. The rapid component of delayed rectifier current (I Kr), which is encoded by the human ether-à-go-go-related gene (hERG), plays an important role on the cardiac action potential (AP) duration.This current is a well-known promiscuous drug target, and many drugs associated with torsade de pointes inhibit the I Kr and hERG channels. 1 Therefore, a key test of the current cardiac ...