AI4AVP_predictor

AI4AVP is a sequence-based antiviral peptides (AVP) predictor based on PC6 protein encoding method [link] and deep learning.

AI4AVP (web-server) is freely accessible at https://axp.iis.sinica.edu.tw/AI4AVP/

requirements

Bio==1.3.9
matplotlib==3.5.1
modlamp==4.3.0
numpy==1.22.1
pandas==1.4.1
scikit_learn==1.1.1
tensorflow==2.7.0

Here we give a quick demo and command usage of our AI4AMP model.

1. quick demo of our PC6 model

For quick demo our model, run the command below

bash AI4AVP_predictor/test/example.sh

The input of this demo is 10 peptides (test/example.fasta) in FASTA format.

The prediction result (test/example_output.csv) below shows prediction scores and whether the peptide is an AMP in table.

2. command usage

Please make sure your working directory access to AI4AVP_predictor/predictor.py and execute command like the example below

python3 predictor.py -f [input.fasta] -o [output.csv] -m [model_type]

-f : input peptide data in FASTA format

-o : output prediction result in CSV

-m : model type (optional) "aug"( AVP+GAN model, defult), "n" (AVP model)

3. Model training source code

Our AI4AVP model traing source code is shown in AI4AVP_predictor/model_training

The model architecture was based on three layers of CNN (filters: (64, 32, 16), kernel_size: (8,8,8)) with rectified linear activation function (ReLU). Every output from the CNN layer was conducted to batch normalization and dropout (rate: (0.5,0.5,0.5)). Finally, there was a fully connected layer (units: 1) with a sigmoid activation function making output values between 0 to 1.

For model training, we randomly split 10% of training data as a validation dataset and set the batch size to 1000. We focused on validation loss every epoch during model training, then stopped training when the training process was stable, and the validation loss was no longer decreasing. Meanwhile, the model at the epoch with the lowest validation loss was saved as the final best model.

Reference

If you find AI4AVP useful, please consider citing: Developing an Antiviral Peptides Predictor with Generative Adversarial Network Data Augmentation