Human iPSC-derived neural network drug response
Culturing of a human iPSC-derived neural network seeded on an MEA was possible without cell aggregation even on the 12th week of culturing. Network burst firing was observed from the 6th week of culture onward. Figure1A (a) shows a cultured 81days in vitro (DIV) phase contrast image, and Fig.1A (b) shows a typical network burst signal. Concentration-dependent data were obtained for 13 seizure-causing compounds and two seizure-free compounds after the 14th week of culturing, when the neural networks were considered mature28. Whenever the signal that was obtained passed a threshold, the spikes detected were used to create a raster plot. Figure1A (c) shows the threshold used to detect spikes in the single electrode signal (top portion) and a raster plot of the detected spikes (bottom portion). Figure1B shows raster plots of compounds with different mechanisms of action: (a) 4-aminopyridine (4-AP), (b) pentylenetetrazol (PTZ), (c) carbamazepine, (d) N-methyl-D-aspartic acid (NMDA), (e) acetaminophen, and (f) dimethyl sulfoxide (DMSO). Seizure-causing compounds caused different changes depending on their mechanism of action (Fig.1B). Figure1C shows a schematic of five analytic parameters calculated from raster plots (total spikes (TS), number of network bursts (NoB), inter network burst interval (IBI), duration of a network burst (DoB), and spikes in a network burst (SiB)). Figure2 shows the drug response of each parameter when the vehicle response is set to 100%. The numerical data are listed in supplementary Tables S1S5. The maximum increases in the NoB of 4-AP and PTZ were 321.0%15.4% (30M) and 147.3%2.7% (10M), respectively. The IBI, DoB, and SiB decreased starting at a concentration of 1M for 4-AP and PTZ (Fig.2a,b). The DoB decreased starting at 0.3M of picrotoxin (Fig.2c). For carbamazepine, the TS and NoB decreased at 30M, and the DoB decreased and the IBI increased at 100M (Fig.2d). For pilocarpine, the IBI increased starting at 10M, the DoB decreased starting at 30M, and the TS decreased at 100M (Fig.2e). For kainic acid, the TS decreased at 0.3M and the NoB went to 0 starting at 1M (Fig.2f). For NMDA, the TS increased at 1M whereas the TS, DoB, and SiB decreased and the NoB increased at 10M (Fig.2g). For tramadol, the NoB decreased and the SiB increased starting at 3M, the TS, DoB, and SiB decreased at 30M, and the IBI increased at 100M (Fig.2h). For theophylline, the IBI increased starting at 10M, and the SiB increased whereas the NoB decreased starting at 30M (Fig.2i). For paroxetine, the DoB decreased starting at 0.3M, and the TS decreased starting at 1M (Fig.2j). For varenicline, the IBI increased and the DoB decreased at 30M (Fig.2k). For venlafaxine, the DoB decreased at 10M, and the TS and SiB decreased at 30M (Fig.2l). For acetaminophen, the DoB decreased starting at 3M (Fig.2m). For DMSO and amoxapine, no changes in any parameters were observed (Fig.2n,o).
MEA data from a cultured human iPSC-derived neural network. (A) (a) Phase-contrast image of neurons on an MEA chip at 81days in vitro (DIV). (b) Typical action potential waveform in a spontaneous recording. (c) Upper graph shows the action potential waveform acquired with a single electrode and the voltage threshold for spike detection (red line). Raster plots of detected spikes (black circles) are shown under the graph. (B) Concentration-dependent Raster plot images of typical mechanisms of action (a) 4-AP, (b) carbamazepine, (c) NMDA, (d) PTZ, (e) acetaminophen, (f) DMSO. (C) Schematic diagram of analysis parameters.
Concentration-dependent changes of 15 compounds in five parameters: TS (pink), NoB (black), IBI (green), DoB (blue), SiB (cyan). Parameters were depicted as the average % change of control (vehicle control set to 100%)SEM from n=34 wells. Data were analyzed using one-way ANOVA followed by post hoc Dunnett's test (*p<0.05, **p<0.01 vs. vehicle).
Based on the preceding results, we found that the changes in the parameters studied were not similar among all seizure-causing compounds; changes differed based on the mechanism of action of the drug. At the same time, a significant difference in the DoB was detected for acetaminophen, which is a seizure-free compound. Changes in DoB may be observed for certain seizure-free compounds. Consequently, we found that there are difficulties in using a single parameter to distinguish between seizure-causing compounds with different mechanisms of action and seizure-free compounds.
We created an artificial intelligence (AI) that was trained on raster plots so that it could classify the responses of seizure-causing compounds with different mechanisms of action as well as the responses of seizure-free compounds. Raster plots were created from the time-series data of the detected spikes, and then images were created by segmenting the data into time windows four times that of the inter-maximum frequency of a network burst interval (IMFI) in the pre-drug administration. The network burst frequency differed depending on the well, so the number of segmented raster plot images also differed depending on the well. The reason for choosing four times the IMFI is that it is suitable for capturing both the regularity of network burst activity and fine firing patterns, and reduces variability between wells. Next, the segmented raster plot images were input into AlexNet36, an object recognition model, and the 4096-dimensional parameters which were output from the fully connected layer (the 21st layer) were extracted as image feature quantities. Lastly, we corrected for differences between the wells due to differing initial states by normalizing the feature quantities of each drug around the mean value of the feature quantities when the vehicle was administered to each well. The 13 seizure-causing compound and two seizure-free compound datasets, which is the number of split raster plots per concentration, were created as shown in Table 1. We used a pattern recognition neural network composed of 4096 neuron input layers, nine sigmoid neuron hidden layers, and an output layer with two classes, which made up a toxicity prediction model to predict whether a compound was a seizure-causing compound or a seizure-free compound (Fig.3A). We used four seizure-causing compounds with different mechanisms and burst frequency responses (4-AP [30 and 60M, n=3 wells, respectively], carbamazepine [100M, n=3 wells], NMDA [3 and 10M, n=3 wells], and PTZ [1000M, n=3 wells]) and two seizure-free compounds (all concentrations of acetaminophen [n=3 wells] and all concentrations of DMSO [n=3 wells]) to train and validate the effectiveness of this model; 75% of the dataset was used for training, and the remaining 25% was used for validation after training (Table 1). The reason for selecting the four seizure-causing compounds is that, in order to cover the firing pattern of the seizure-causing compound, compounds having different mechanisms of action were selected as training data, one in which the firings increased and the other in which the firings decreased. The accuracy was evaluated using the raster plots of unlearned wells after training, i.e., using the holdout scheme. The training data used contained 330 4-AP plots, 822 carbamazepine plots, 1323 NMDA plots, 198 PTZ plots, and 3546 acetaminophen plots and DMSO 2286 plots. The test data used contained 111 4-AP plots, 294 carbamazepine plots, 441 NMDA plots, 54 PTZ plots, and 1182 acetaminophen plots and 702 DMSO plots. We created a confusion matrix of the seizure-causing and seizure-free classification results from the training data and test data (Fig.3B). Next, a receiver operating characteristic curve and the area under the curve (AUC) were calculated for all training data and all test data, and the optimal operating point was determined (Fig.3C(a)). The accuracy, positive predictive value, sensitivity, specificity, and F-measure of the prediction results of the model at the optimal operating point were calculated (Table 2). The model trained on raster plot feature quantities had an AUC in the training data of 0.9998 and an AUC in the unlearned data of 0.9967; the optimal operating point was 0.158. The classification precision in the training data for each drug at the optimal operating point was as follows: 100% for 4-AP, 97.8% for carbamazepine, 99.6% for NMDA, and 96.0% for PTZ. The classification precision in the unlearned data was 100% for 4-AP, 91.5% for carbamazepine, 100% for NMDA, and 94.4% for PTZ. The prediction accuracy for all compounds was 98.4%.
Creation of seizure risk prediction AI using raster plot images and evaluation of the prediction model. (A) Data flow and architecture of seizure risk prediction model. w1 is the weight between the input layer and the hidden layer, w2 is the weight between the hidden layer and the output layer. (B) (a) Confusion matrix for each compound used for training, (b) confusion matrix for the entire training dataset, (c) confusion matrix for each compound used for the test, (d) confusion matrix for the entire test dataset. The test dataset used the data of the wells not used for training dataset. Vehicle in the confusion matrix indicates vehicle data in four seizure-causing compounds. (C) (a) Receiver operating characteristic (ROC) curve after classification of training and testing data in a neural network model (black line: training data; red line: testing data; red dot: optimum operating point). (b) Comparison of ROC curves after classification of the same testing data in NN and SVM models (black line: SVM model; red line: NN model).
Figure3C(b) shows the ROC curve using a support vector machine (SVM) model trained with the same 4096-dimensional feature dataset as the neural network (NN) model. Comparing the AUC in the test data of SVM and NN revealed that the NN model had an AUC of 0.9967 and the SVM model had an AUC of 0.9841; thus, the NN model was superior to the SVM model [Fig.3C(b)]. Therefore, in this study, we used the NN model.
The seizure-causing/seizure-free classification AI trained on the raster plots that we created accurately classified the responses of seizure-causing compounds with differing mechanisms and seizure-free compounds.
If we are able to establish a ranked development priority for compounds based on their seizure liability, it will lead to more efficient drug discovery and development. Determining the concentration dependence is necessary in order to assign priority to drugs. Thus, using the AI we created, we investigated the concentration dependence of seizure-causing/seizure-free judgments. The concentration data toxicity probabilities predicted by the AI are shown in Fig.4. The proportions of the images classified as seizure-causing and as seizure-free used the time-series data from each well, and then the mean probability for each well was calculated and used to represent the toxicity risk at each concentration. For unlearned sample, which includes data of the wells that were not used for training dataset, the following concentrations were determined to have a seizure liability probability of 50% or higher4-AP: 1M (62.2%), 10M (94.6%), 30M (100%), and 60M (100%); carbamazepine: 30M (76.9%) and 100M (85.0%); NMDA: 1M (63.3%), 3M (100%), and 10M (100%); and PTZ: 1M (51.9%), 10M (81.5%), 100M (88.9%), and 1000M (88.9%) (Fig.4 (a), (b), (d), (e)). The seizure liability at concentrations lower than the concentrations the AI was trained on was shown, and then the concentration dependence was calculated. Acetaminophen, which is a seizure-free compound, was determined to be seizure-free with a probability of 97.9% or higher, regardless of the concentration. DMSO was also determined to be seizure-free with a probability of 99.1% or higher, regardless of the concentration (Fig.4c,f). The seizure liability prediction AI we created determined the concentration dependence of seizure-causing compounds and identified seizure-free compounds as seizure-free regardless of the concentration.
Concentration-dependent prediction of seizure risk in learning drugs by AI. AI predicted the negative probabilities (blue bar) and seizure risk (red bar) at each concentration of training data (left) and test data (right). (a) 4-AP, (b) NMDA, (c) acetaminophen, (d) carbamazepine, (e) PTZ, (f) DMSO.
It is important for the AI that we created to detect the toxicity of drugs that it has not been trained on. Thus, we used the AI we created to determine the toxicity of nine unlearned seizure-causing compounds based on data collected on them. In order to verify AI, nine unlearned seizure-causing compounds were regarded as unknown compounds and were not trained. Figure5 shows the seizure toxicity determination results for each concentration of the unlearned drugs. The concentrations that showed a 50% or higher probability of seizure liability were as followskainic acid: 1M (81.8%), 3M (100%), and 10M (100%); paroxetine: 3M (73.7%), 10M (100%), and 30M (100%); picrotoxin: 0.1M (91.4%), 0.3M (93.7%), 1M (91.8%), 3M (97.8%), and 10M (91.5%); varenicline: 10M (52.6%) and 30M (77.1%); pilocarpine: 1M (62.3%), 3M (75.8%), 10M (86.8%), 30M (89.4%), and 100M (97.0%); tramadol: 3M (61.9%), 10M (88.6%), 30M (98.9%), and 100M (100%); and venlafaxine: 10M (90.5%), 30M (100%), and 100M (100%). Seven of the unlearned drugs were determined to have concentration-dependent seizure liability (Fig.5ad, fh). On the other hand, amoxapine and theophylline were determined to be seizure-free at all concentrations (Fig.5e,i). This showed that the AI was able to detect seizure toxicity in a concentration-dependent manner, even for unlearned drugs.
Concentration-dependent prediction of seizure risk in non-training drugs by AI. AI predicted the negative probabilities (blue bar) and seizure risk (red bar) at each concentration. (a) Kainic acid, (b) paroxetine, (c) picrotoxin, (d) varenicline, (e) amoxapine, (f) pilocarpine, (g) tramadol, (h) venlafaxine, (i) theophylline.
In order to verify whether AI can determine the safety of unlearned negative compounds, the negative compounds Aspirin (1, 3, 10, 30, 100M) and Amoxicillin (1, 3, 10, 30, 100) M) and Felbinac (1, 3, 10, 30, 100M) data were judged (Fig.6). The negative probabilities of Aspirin were 76.3% (1M), 82.0% (3M), 79.0% (10M), 80.8% (30M), and 81.7% (100M). Amoxicillin were 91.3% (1M), 86.3% (3M), 86.4% (10M), 81.1% (30M), and 77.6% (100M). Felbinac were 83.8% (1M), 80.9% (3M), 76.1% (10M), 71.8% (30M), and 77.7% (100M) (Fig.6b). Although there were some significant differences in the conventional analysis parameters (Fig.6a), AI judged negative at all three concentrations. From these results, it was confirmed that AI can be judged to be negative even for negative compounds that are unlearned drugs.
Prediction of seizure risk in non-training negative compounds by AI. (A) Concentration-dependent changes of 3 negative compounds in five parameters: TS (pink), NoB (black), IBI (green), DoB (blue), SiB (cyan). (a) Aspirin, (b) amoxicillin, (c) felbinac, (B) AI predicted the negative probabilities (blue bar) and seizure risk (red bar) at each concentration.
Because seizure-causing compounds with differing mechanisms elicit different responses, if the AI is able to classify the compounds despite this, it can also predict the mechanism of seizure liability of unlearned drugs. Thus, we trained the AI on drug names and raster plots in order to classify compounds as seizure-causing compounds with differing mechanisms or seizure-free compounds.
We used a pattern recognition neural network composed of 4096 neuron input layers, 120 hidden layers containing sigmoid neurons, and an output layer with 14 classes (Fig.7), which made up a drug identification model to predict the name of seizure-causing compounds and seizure-free compounds. The model was trained on a dataset composed of 4-AP (30 and 60M), amoxapine (100M), carbamazepine (30 and 100M), kainic acid (1, 3, and 10M), NMDA (3 and 10M), PTZ (1000M), paroxetine (3, 10, and 30M), picrotoxin (1, 3, and 10M), pilocarpine (10, 30, and 100M), theophylline (100M), tramadol (30 and 100M), varenicline (30M), and venlafaxine (10, 30, and 100M) as well as all concentrations of acetaminophen as well as all concentrations of DMSO as seizure-free compounds (Table 3). The all compounds dataset that was used was made up of 56 wells. Training was conducted by excluding one of the 56 wells and training the AI on the names of the drugs in the other 55 wells; 75% of the 55 well datasets were used for training, and 25% were used for validation after training. The excluded well was used for obtaining test data. The prediction accuracy was calculated using the leave-one-sample (well)-out scheme. We created five AIs for each excluded well. In other words, we created 565=280 AIs. For the data from the single well (the data from the single well that was not used to train the AI), the name of the drug was identified based on the five models we created, and the mean value was calculated. The deviation of the five models prediction accuracy was 0.11% at the trained concentrations of all drugs. The deviation of the prediction accuracy at all concentrations of all drugs was 1.6%. The predictive probabilities at different drug concentrations are shown in Table 4. DMSO and acetaminophen, which are seizure-free compounds, were judged to be seizure-free at all concentrations for every drug vehicle, with a mean probability of 99.9%0.3%. 4-AP (1M), amoxapine (3M), NMDA (1M), picrotoxin (0.1M), pilocarpine (1M), PTZ (10M), theophylline (3M), varenicline (10M), venlafaxine (3M), and tramadol (10M) were correctly identified at concentrations lower than those in the training data. Carbamazepine (30M), kainic acid (1M), and paroxetine (3M) were correctly identified at the concentrations used to train the AI. The drugs that could not be identified at certain concentrations were all seizure-free compounds, and no drugs were misidentified as different drugs. The mean predictive accuracy for all drugs at the concentrations used to train the AI was 99.9%0.1%. The drug identification AI we created correctly identified the responses of 13 seizure-causing compounds and two seizure-free compounds.
Creation of drug name prediction AI using raster plot images. Data flow and architecture of drug name prediction model. w1 is the weight between the input layer and the hidden layer; w2 is the weight between the hidden layer and the output layer.
View original post here:
Raster plots machine learning to predict the seizure liability of drugs and to identify drugs | Scientific Reports - Nature.com
Read More..