desmondyuan / deepd

Developed and maintained by Xu lab.

DeepD

This is the main script of DeepD v0.1

Developed and maintained by Xu lab at https://github.com/DesmondYuan/deepD

For quick start, try this with

python scripts/main.py -config=configs/example_GENT_NC.json

If you want to discuss the usage or to report a bug, please use the 'Issues' function here on GitHub.

If you find DeepD useful for your research, please consider citing the corresponding publication.

Installation using pip

The following command will install pertbio from a particular branch using the '@' notation:

git clone https://github.com/DesmondYuan/deepD
cd deepD
pip install -r requirements.txt

Note that only python=3.7 is currectly supported. Anaconda or pipenv is recommended to create a python environment.

Quick start

The experiment type configuration file is specified by --experiment_config_path or -config and a random seed can also be assigned by using argument -seed

python scripts/main.py -config=configs/example_GENT_NC.json -seed=1234

Repo structure

./DeepD/ - the core DeepD package

• The module DeepD.data is designed for normalization and data clipping as we described in Methods section.
• The module DeepD.model designs all the DeepD models we discussed about in the paper.
• The module DeepD.train contains core optimization methods used for training.
• The module DeepD.util contains a few helper functions for evaluation and training monitoring.

./scripts/ folder

This folder contains the main scripts that is ready to use.

• The main.py is the main script of DeepD v0.1
• The preprocess_data.py is the main script for data preprocessing. For quick start, preprocessed data files can be found in the ./data/ directory and can be directly used for training.

./data/ folder

This folder contains preprocessed dataset for testing runs.

• Dataset1_GENT_L1000_U133Plus2 is the dataset from this paper. We used this dataset for training normal vs cancer classification.

  Shin G, Kang TW, Yang S, Baek SJ, Jeong YS, Kim SY. GENT: gene expression database of normal and tumor tissues. Cancer Inform 2011; 10:149-157.Shin G, Kang TW, Yang S, Baek SJ, Jeong YS, Kim SY. GENT: gene expression database of normal and tumor tissues. Cancer Inform 2011; 10:149-157.

  Update 2020: The GENT2 is release at http://gent2.appex.kr/gent2/ and the MySQL dataset can be accessed at http://www.appex.kr/web_download/GENT2/GENT2_dump.sql.gz

• Dataset2_GDC_L1000 is the dataset from the NCI Genome Data Commons available at https://gdc.cancer.gov.

• L1000_reference.csv is the L1000 landmark genes we used for dataset preprocessing and model training. It is defined by the NIH LINCS project.

./configs/ folder

The folder contains the configuration files in json format.

Example configs are provided for results reproductivity and a debug.json is also included for testing compilation.

An example json looks like this

{
    "expr_name": "Example_GENT",
    "train_dataset": "data/Dataset1_GENT_L1000_U133Plus2.experiment.csv",
    "test_dataset": "data/Dataset1_GENT_L1000_U133Plus2.withheld.csv",
    "annotation_col": "nc_label",
    "validation_ratio": 0.3,
    "n_genes": 978,
    "unsupervised_layers": [
        [978, 1000],
        [1000, 500],
        [500, 200],
        [200, 100],
        [100, 30]
    ],
   "supervised_layers": [
        [30, 30],
        [30, 30],
        [30, 30],
        [30, 2]
    ],
    "pretrain_tp2vec": true,
    "plot_pretrain_results": true,
    "train_disconnected_classifier": true,
    "train_connected_classifier": true,
    "max_iteration": 100000,
    "max_iteration_pretrain": 3000,
    "n_iter_patience": 1000,
    "n_iter_patience_pretrain": 100,
    "n_iter_buffer": 5,
    "activation": "tf.nn.relu",
    "learning_rate": 1e-3,
    "l1": 1e-4,
    "l2": 1e-2,
    "optimizer": "tf.compat.v1.train.AdamOptimizer({}, beta1=0.9, beta2=0.9)",
    "verbose": 4,
    "listen_freq": 10,
    "pretrain_batch_size": 1024,
    "batch_size": 1024
}

Each configuration needs the following information

"expr_name": (str) Label used as folder name under results
"train_dataset": (str) Location of the dataset that would be further split into training set and validation set with
                 the "validation_ratio.
"test_dataset": (str) Location of the withheld/test dataset.
"annotation_col": (str) On which column of the input data frame would a supervised model be trained to classify.
"validation_ratio": (float) The training/validation ratio for data partition used for "train_dataset".
"n_genes": (int) Number of genes from the input data.
"unsupervised_layers": (list) A list of layer sizes used for encoders and decoders.
"supervised_layers": (list) A list of layer sizes used for supervised classifier DeepDCancer.
"pretrain_tp2vec": (bool) Whether to perform unsupervised pretraining.
"plot_pretrain_results": (bool) Whether to plot the results after pretraining.
"train_disconnected_classifier": (bool) Whether to perform the disconnected supervised classification (DeepDCancer).
"train_connected_classifier": (bool) Whether to perform the connected supervised classification (DeepDcCancer).
"max_iteration": (int) Maximum number of iterations used for training.
"max_iteration_pretrain": (int) Maximum number of iterations used for pretraining.
n_iter_buffer (int): The moving window for eval losses during training.
n_iter_patience (int): How many iterations without buffered loss on validation dataset decreases would result in
                       an earlystop in training.
"n_iter_patience_pretrain":How many iterations without buffered loss on validation dataset decreases would result in
                       an earlystop in pretraining (for each layer).
learning_rate (float): The learning rate for Adam Optimizer. Note that we used the default beta1 and beta2 for Adam.
l1 (float): l1 regularization strength.
l2 (float): l2 regularization strength.
"activation": (tensorflow) Activation function for dense layers.
"optimizer": (tensorflow) Which optimizer would be used for training.
"verbose": (int) export verbose
"listen_freq": (int) Printing training loss for each # of iterations.
"pretrain_batch_size": Batch size for each iteration in pretraining.
"batch_size": Batch size for each iteration in training.