DeepMir

This repository contains the source code required to replicate the experiments described in the pre-print submited to bioRxiv Detection of pre-microRNA with Convolutional Neural Networks. Additionally, it contains a ready to use implementation of the pre-miRNA detection framework proposed in the pre-print.

The content of this project is organized as follows:

• the model_selection directory contains the scripts used to train and evaluate several CNNs based on VGG, Inception, and ResNet architectures.
• the hairpin_image_generator directory contains the JAR file used for hairpin image generation.
• The notebooks directory contains Jupyter notebooks used to analize CNN models and generate images included in the pre-print.
• The datasets directory contains the datasets used to train our best CNN model. Additionally, the data inside benchmark was used to evaluate the performance of the CNN model.
• The models directory contains the best CNN model obtained by using a fine-tuning procedure during training. The base CNN model (without fine-tuning) is also included here.
• The examples directory contains examples of fasta files that the script predictor.py can process.
• The predictor.py script is an implementation of the pre-miRNA detection framework described in the pre-print.

Dependencies

This project requires:

• python 3.5 or later
• pyfaidx 0.5.7
• imageio 2.5.0
• xlrd
• Keras 2.2.4
• tensorflow 1.14.0
• ipython 7.9.0
• scikit-learn 0.21.3
• matplotlib 3.0.3
• pandas 0.24.2
• jupyter 5.7 or later

Usage

To use the pre-miRNA detection framework run

python predictor [filename.fasta]

If the input file is successfully processed, the prediction results will be available in the file user_data/filename/results.csv. Additionally, the hairpin images corresponding to the sequences defined in filename.fasta will be available in the directory user_data/filename/images. Numpy files containing the hairpin images encoded as arrays and their corresponding names encoded will also be available as user_data/filename/images.npz and user_data/filename/names.npz.

To train new CNN models run the trainer scripts inside the directory model_selection. For example,

python vgg_model_trainer_fine_tuning.py

This script will train several models based on the VGG architecture using fine-tuning. The resulting models will be available in the models directory. To modify the resulting CNN models, modify the constants defined in each trainer script: TRAIN_EPOCHS, PRETRAIN_EPOCHS, BATCH_SIZE, N_FILTERS, and DENSE_UNITS.

The script model_selection/benchmark_fine_tuned_model.py computes the performance of fine-tuning-CNN using several benchmark datasets.

Note: this repository contains some relatively large datasets such as nonhsa_modmirbase_images.npz, modmirbase_train_images.npz, mirbase_images.npz, and neghsa_images.npz. Take this into consideration when working with this project.