This repo comprises a method to identify atypical dose-response curves using a Convolutional Neural Network trained from simulation data. Most of the implementation is aimed at 7-dose response curves and aim to predict hermetic transition points.
-
Follow the instructions provided by the synthetic_doseresponse_generator github repo to produce your training data or use the data provided here.
-
Run the following scripts (assumes you've cloned the repo)
# This will preprocess as many chained datasets (produced by the generator above) together and split training/test dataset
python atyp_gen_preprocessing_classifier.py ../../data/synth_set1.csv ../../data/typical_noisy_set.csv
# This will train a model, use tensorboard to choose the optimal model
python atyp_train_classifier.py
# This step will take a while to package the results into a dataframe, expect 30+ minutes
#path/to/model #number of plots to display
python atyp_test_classifier.py ./models/best_model.06-0.09.h5 10This will produce the csv file: classifier_test_results.csv in ./python/classifier/. Using this data, open the Rmd file ./R/atyp_classifier_performance.Rmd to review performance and choose an optimal cutoff. Be wary of class imbalances.
Data for training and testing the CNN can be found here. This is a synthetic dataset, prduced using the methods described in synthetic_doseresponse_generator
This is the model we suggest using.
This model attempts to predict the transition point, it was trained with a class imbalance that renders it non-generalizable.
Details on how these metrics were caculated can be found in hermetic_cnn_test.py.
average observation dose point classification accuracy: 89.7%
dose point classification specificity: 73.9%
dose point classfication sensitivity: 99.8%
A more in-depth performance analysis can be found in here. As a general overview, the parameter space regression is shown below.
