Deep-B3

A multi-model framework for blood-brain barrier permeability discovery

Dependencies

This repository has been tested on CentOS 7.5. We strongly recommend you to have Anaconda3 installed, which contains most of required packages for running this model.

Must installed packages or softwares

• FastAI 1.0.61

• PyTorch 1.9.0

• Torchvision 0.10.0

• Pandas 1.3.2

• Numpy 1.19.2

• RDKit 2020.09.1

Get Started

Generate features from the SMILES file and train a new deep-b3 model

• SMILES file is a csv format, within head (id,smi, label)
• The train and test feature file was stored in ./train and ./test, respectively.
• The iamges for training and testing the model were stored in ./train_images and ./test_images, respectively.
• You can run the script make_feature.py to generate features for your train data. Sample code:

  python make_feature.py --smifile train.csv --outfile train_feature.csv [--is_balance]

• if is_balance, the data will be balanced by multiple SMILES-Based Augmentation method
• if is test data, then the test feature are generated by

  python make_feature.py --smifile test.csv --outfile test_feature.csv --is_test

train a new deep-b3 model on your data

• You can run the script deep-b3.py to train a new model for the data or on your new data, and the models will be stored in ./models. Sample code:
  1. train a model including image, tabular and text features, and the features output from the CNN is n and NLP is m
python train.py train --feature train.csv --epoch 50 --bs 64 --vis_out n --text_out m
2. train a model based on only image, and the features output from the CNN is m
python train.py train --feature train.csv --epoch 50 --bs 64 --vis_out m --has_tab False --has_text False
3. train a model based on only text, and the features output from the NLP is m
python train.py train --feature train.csv --epoch 50 --bs 64 --text_out m --has_tab False --has_img False
4. train a model based on only tabular
python train.py train --feature train.csv --epoch 50 --bs 64 --has_img False --has_text False
  and so on...

  test the pre-trian deep-b3 model on the test data

• You can run the script train.py to test the pre-trained model on the test data used in this study. Sample code:

  python train.py test --feature test_feature.csv