Machine learning on multiple epigenetic features reveals H3K27Ac as a driver of gene expression prediction across patients with glioblastoma.
Our study's two patient epigenetic marker files are provided in the data folder. Specifically, the latest versions of these GSC measurements are located in the latest_versions_of_all_raw subfolder. Additionally, the ind_shuffle.npy file is provided. This file was used to create consistent dataset splits for train, validation, and test sets.
We evaluated the cross-patient prediction methodology with all of the model scripts using this project's datasets. We performed extended evaluation and analysis with the XGBoost-based model using both this study's datasets and those created by adapting data from the study "Chromatin landscapes reveal developmentally encoded transcriptional states that define human glioblastoma" (https://doi.org/10.1084/jem.20190196). (1) Their study's data is avaliable from the Gene ExpressionOmnibus under accessions GSE119755 and GSE119834.
Clone this repository to the local filesystem using the link provided by the "Code" dropdown button above. For example:
git clone https://github.com/rsinghlab/ML_epigenetic_features_glioblastoma.git
Change the current working directory to the folder created by the clone process:
cd ./ML_epigenetic_features_glioblastoma
We recommend that a virtual environment be created to allow for the installation of the required packages and libraries in a without potential conflict with other packages already installed on the system. In the example here the virtual environment is given the same name as the project folder.
python3 -m venv ML_epigenetic_features_glioblastoma
Activate the new python environment.
source ./ML_epigenetic_features_glioblastoma/bin/activate
you can now install packages into the new environment using the included requirements.txt file.
pip3 install -r requirements.txt
The project's cross-patient prediction models are avaliable in the following locations:
XGBoost (XGBR) code/models/xgboost/xgboost_cross_patient_pred_regression_gsc_stem_standard_log2.py
"Branched" Multi-layered Perceptron ("Branched" MLP) code/models/mlp/mlp_cross_patient_regression_gsc_stem_sequence_standard_log2.py
Multi-layered Perceptron (MLP) code/models/mlp/mlp_cross_patient_regression_gsc_stem_standard_log2.py
Convolutional Neural Network (CNN) code/models/cnn/cnn_cross_patient_pred_regression_gsc_stem_standard_log2.py
Recurrent Neural Network (RNN) code/models/rnn/rnn_models.py
Gradient Boosting Regression (GBR) code/models/gbr/gbr_cross_patient_pred_regression_gsc_stem_standard_log2.py
Support Vector Machine (SVR) code/models/svm/svm_cross_patient_pred_regression_gsc_stem_standard_log2.py
Multiple Linear Regression (MLR) code/models/mlr/mlr_cross_patient_pred_regression_gsc_stem_standard_log2.py
XGBoost, Gradient Boosting Regression, Convolutional Neural Network, Multi-layered Perceptron, Support Vector Machine, and Multiple Linear Regression
A) The script's path and filename.
B) The first data file's path and filename. This script creates the model's training and validation (or training only) sets from this file.
NOTE: The creation of a validation set is controlled by the validation = True or False statement in the script's main function. The proportions given to each set specified in the get_data_patient_1 function under the comment #HYPERPARAMETER TUNING SPLITS and #TESTING SPLITS.
C) The second data file's path and filename.
D) The ind_shuffle.npy file (or equivalent) mentioned in the "Datasets" section above.
E) The integer to be used as the random seed.
F) The absolute or relative directory path where the various script functions will direct model output, predictions and visualizations. If no directory is specified, a directory name will be automatically generated and the directory created in the same directory where the script resides. The save directory's name will include the date and time the script was run.
A) The script's path and filename.
B) The first data file's path and filename. This script creates the model's training and validation (or training only) sets from this file.
NOTE: The creation of a validation set is controlled by the validation = True or False statement in the script's main() function. The proportions given to each set specified in the get_data_patient_1 function under the comment #HYPERPARAMETER TUNING SPLITS and #TESTING SPLITS.
C) The second data file's path and filename.
D) The script requires a file containing the genomic sequences for the corresponding genes in the first and second input files. Only one file is necessary since the script will use the sequences for the first and second inputs with the expectation that the sequence file includes data for all the genes contained in both.
The HG38_reference_genome_sequence_input_numerical_encoding.npy in the data/HG38_reference_genome_data_files folder was the primary version used for the study's experiments.
E) The ind_shuffle.npy file (or equivalent) mentioned in the "Datasets" section above.
F) The integer to be used as the random seed.
G) The absolute or relative directory path where the various script functions will direct model output, predictions and visualizations. If no directory is specified, a directory name will be automatically generated and the directory created in the same directory where the script resides. The save directory's name will include the date and time the script was run.
A) The script's path with the rnn_models.py filename.
B) The designation of the train and validation dataset. Either GSC1 or GSC2 PLEASE NOTE: This designation is not the path and filename for the dataset.
B) The designation of the test dataset. Either GSC1 or GSC2 PLEASE NOTE: This designation is not the path and filename for the dataset.
C) The designation of which model architecture to be used: LSTM, biLSTM, LSTM-concatenated, GRU, or biGRU.
D) A true or false for this argument will activate or de-activate the script's dataset preprocessing. If the script has been run previously with the chosen datasets, indicating false will allow the script to load the previously saved preprocessed files.
E) The script will either be run for hyperparameter-tuning, testing, or visualizing.
F) The random seed integer.
G) Indicating attention will be used. If this is left blank attention will not be used.
PLEASE NOTE: The path and filename for each dataset and the indices file should be entered in the under GSC1 and GSC2 in the script's main function.
- Mack SC, Singh I, Wang X, Hirsch R, Wu Q, Villagomez R, et al. Chromatin landscapes reveal developmentally encoded transcriptional states that define human glioblastoma. J Exp Med. 20190404th ed. 2019 May 6;216(5):1071–90