This is the GitHub repository for the manuscript

"An Integrated Machine Learning Approach Delineates Entropic Modulation of alpha-Synuclein by Small Molecule"

doi: https://doi.org/10.1101/2024.01.15.575555

This GitHub repository contains two directories : (1) variational autoencoder (VAE) and (2) denoising convolutional VAE (DCVAE). This contains programs for VAE and DCVAE models used in our work.

#----------REQUIRED PACKAGES----------#

1. TensorFlow
2. NumPy
3. Scikit-learn

#----------PROGRAM INPUTS----------#

For both VAE and DCVAE models, one requires a file that contains the training and testing data in a single file in *.npy format. For VAE, the input shape is (9453,) and for DCVAE, the input shape is (144,144,1).

#----------PROGRAM OUTPUTS----------#

A directory containing the training and validation loss and the model parameters.

#----------HOW TO RUN THE VAE PROGRAM----------#

1. Modify the following variables for a different protein in the file "train_dvae.py" :

   a. "input_shape"

   b. "path" # Path to data

   c. filename as in "ca_contact_scaled = np.load(path+f"data.npy")". Change "data.npy" to "your_filename.npy".

   d. Change the number of encoder, latent and decoder neurons as per requirement.

   e. Change the optimizer, loss, learning_rate, batch_size, epochs, alpha and beta as per requirement.

2. Type "./train_dvae.py" in the terminal with the environment activated with the specified packages and hit Enter.

3. To create the latent data and reconstruct the input open a jupyter notebook and type the following statements :

      ate=dense_variational_autoencoder.load(save_dir_name) # Name of the directory where the model parameters are saved. 
      ca_contact_scaled = np.load(path+f"data.npy") # Load the data
      latent_data=ate.reconstruct_latent(ca_contact_scaled) # Construct the latent data
      reconstructed_data = ate.reconstruct_input(ca_contact_scaled) # Reconstruct the input data
   

#----------HOW TO RUN THE DCVAE PROGRAM----------#

1. Modify the following variables for a different protein in the file "train_denoising_vae.py" :

   a. "path" # Path to data

   b. filename as in "ca_contact_scaled_asyn_fas = np.load(path+f"data_asyn_fas.npy")". Change "data_asyn_fas.npy" to "your_filename.npy".

   c. Change the noise_factor, input_shape, filters, kernels, strides, padding, number of encoder, latent and decoder neurons as per requirement.

   d. Change the optimizer, loss, learning_rate, batch_size, epochs and beta as per requirement.

2. Type "./train_denoising_vae.py" in the terminal with the environment activated with the specified packages and hit Enter.

3. To create the latent data and reconstruct the input open a jupyter notebook and type the following statements :

    ate=variational_convolutional_autoencoder.load(save_dir_name) # Name of the directory where the model parameters are saved. 
    ca_contact_scaled = np.load(path+f"data_asyn_fas.npy") # Load the data
    latent_data = ate.reconstruct_latent(ca_contact_scaled) # Construct the latent data
    reconstructed_data = ate.reconstruct_input(ca_contact_scaled) # Reconstruct the input data