Bioinfor DanQ

DanQ is a hybrid convolutional and recurrent deep neural network for quantifying the function of DNA sequences. This is implemented by tensorflow-2.0 again.

DanQ

Model Architecture

CNN + BidLSTM + Dense

Loss Function

Binary Cross Entropy

Optimization Method

Adam

USAGE

Requirement

We run training on Ubuntu 18.04 LTS with a GTX 1080ti GPU.

Python (3.7.3) | Tensorflow (2.0.0) | CUDA (10.0) | cuDNN (7.6.0)

Data

You need to first download the training, validation, and testing sets from DeepSEA. You can download the datasets from here. After you have extracted the contents of the tar.gz file, move the 3 .mat files into the data/ folder.

Model

The model that trained by myself is available in BAIDU Net Disk here

Preprocess

Because of my RAM limited, I firstly transform the train.mat file to .tfrecord files.

python preprocess.py

Training

Then you can train the model initially.

CUDA_VISIBLE_DEVICES=0 python main_DanQ.py -e train

Test

When you have trained successfully, you can evaluate the model.

CUDA_VISIBLE_DEVICES=0 python main_DanQ.py -e test

RESULT

Yon can get the result in the ./result/ directory.

Loss Curve

For DanQ:

For DanQ-JASPAR:

Metric

We use two metrics to evaluate the model. (AUROC, AUPR)

For DanQ:

-	DNase	TFBinding	HistoneMark	All
AUROC	0.9022	0.9317	0.8303	0.9162
AUPR	0.4072	0.2984	0.3373	0.3176

For DanQ-JASPAR:

-	DNase	TFBinding	HistoneMark	All
AUROC	0.9124	0.9451	0.8395	0.9287
AUPR	0.4323	0.3271	0.3508	0.3441

REFERENCE

DanQ: a hybrid convolutional and recurrent deep neural network for quantifying the function of DNA sequences | Github