deepdrug3D/DeepDrug.py

#!/usr/bin/env python
# coding: utf-8

# # DeepDrug3D

# ## Importing library

# In[ ]:


import numpy as np
import tensorflow as tf
from sklearn.preprocessing import LabelEncoder
from keras.models import Sequential
from keras import optimizers, callbacks
from keras.layers import Dense, Flatten, TimeDistributed, Dropout
from keras import Input, Model
from keras.layers import add, Activation
#from keras.utils import plot_model  # Needs pydot.
from keras.layers import Conv3D, MaxPooling3D


# ### used to store model prediction in order to plot roc curve

# In[ ]:


class prediction_history(callbacks.Callback):
    def __init__(self):
        self.predhis = []
    def on_epoch_end(self, epoch, logs={}):
        self.predhis.append(model.predict(predictor_train))


# ### Creating input and ouputs

# In[ ]:


def in_out_lists(size=1000):
    """
    returns a tuple of array used as input and output for the model
    Arguments:
        - size, int: default 1000, size of the lists to be created
        
    Returns:
        - tuple (data_onehot, output):
            -data_onehot, ndarray: containing one-hot encoded pockets
            -output, ndarray: containing size-3 vectors for classification
    """
    with open("control.list", "r") as filin:
        control = filin.read()
        control = control.split("\n")
        control.pop()

    with open("steroid.list", "r") as filin:
        steroid = filin.read()
        steroid = steroid.split("\n")
        steroid.pop()

    with open("heme.list", "r") as filin:
        heme = filin.read()
        heme = heme.split("\n")
        heme.pop()

    with open("nucleotide.list", "r") as filin:
        nucleotide = filin.read()
        nucleotide = nucleotide.split("\n")
        nucleotide.pop()
    
    lmin = len(heme)
    lmid = len(nucleotide)
    lmax = len(control)
    tot_size = lmin + lmid + lmax
    data_onehot = np.ndarray(shape=(size, 14, 32, 32, 32)) # initializing empty array

    np.random.seed(9001)
    indices = np.random.permutation(tot_size)
    indices = indices[:size]
    output = np.ndarray(shape=(size, 3)) # softmax 3, {steroid=1, heme=1, nucleotide=1}

    n = -1
    for i in indices:
        n += 1
        if i < lmin:
            data_onehot[n,] = np.load("deepdrug3d_voxel_data/"+heme[i]+".npy")
            output[n,] = [1,0,0]
        elif i > lmin and i < (lmin + lmid):
            data_onehot[n,] = np.load("deepdrug3d_voxel_data/"+nucleotide[i - lmin]+".npy")
            output[n,] = [0,1,0]
        else:
            data_onehot[n,] = np.load("deepdrug3d_voxel_data/"+control[i - (lmin+lmid) - 1]+".npy")
            output[n,] = [0,0,1]
    
    return (data_onehot, output)


# ### Defining different model to test and compare

# In[ ]:


def model_heavy(): # créer un objet modèle
    """
    Return a simple sequentiel model
    
    Returns :
        - model : keras.Model
    """
    inputs = Input(shape=(14,32,32,32))
    conv_1 = Conv3D(64, (28, 28, 28), padding="same", activation="relu", kernel_initializer="he_normal")(inputs)
    conv_2 = Conv3D(64, (26, 26, 26), padding="same", activation="relu", kernel_initializer="he_normal")(conv_1)
    drop_1 = Dropout(0.2)(conv_2)
    maxpool = MaxPooling3D()(drop_1)
    drop_2 = Dropout(0.4)(maxpool)
    dense = Dense(512)(drop_2)
    drop_3 = Dropout(0.4)(dense)
    flatters = Flatten()(drop_3)
    #output = TimeDistributed(Dense(3, activation='softmax'))(drop_3)
    output = Dense(3, activation='softmax')(flatters)
    model = Model(inputs=inputs, outputs=output)
    my_opt = optimizers.Adam(learning_rate=0.000001, beta_1=0.9, beta_2=0.999, amsgrad=False)
    print(model.summary)
    model.compile(optimizer=my_opt, loss="categorical_crossentropy",
                  metrics=["accuracy"])
    return model


# In[ ]:


def model_light(): # créer un objet modèle
    """
    Return a simple sequentiel model
    
    Returns :
        - model : keras.Model
    """
    inputs = Input(shape=(14,32,32,32))
    conv_1 = Conv3D(32, (28, 28, 28), padding="same", activation="relu", kernel_initializer="he_normal")(inputs)
    conv_2 = Conv3D(64, (26, 26, 26), padding="same", activation="relu", kernel_initializer="he_normal")(conv_1)
    drop_1 = Dropout(0.2)(conv_2)
    maxpool = MaxPooling3D()(drop_1)
    drop_2 = Dropout(0.3)(maxpool)
    maxpool_2 = MaxPooling3D()(drop_2)
    drop_3 = Dropout(0.3)(maxpool_2)
    dense = Dense(256)(drop_3)
    drop_4 = Dropout(0.4)(dense)
    flatters = Flatten()(drop_4)
    output = Dense(3, activation='softmax')(flatters)
    model = Model(inputs=inputs, outputs=output)
    my_opt = optimizers.Adam(learning_rate=0.000001, beta_1=0.9, beta_2=0.999, amsgrad=False)
    print(model.summary)
    model.compile(optimizer=my_opt, loss="categorical_crossentropy",
                  metrics=["accuracy"])
    return model


# ## Create pocket lists
# 4 lists are created :
#   + control
#   + steroid
#   + heme
#   + nucleotide

# In[ ]:


data = in_out_lists(1400)
pockets = np.cumsum(data[1], axis=0)[-1]


# In[ ]:


print("with random seed=9001 and a 1400 pockets dataset the rates are:\n      {} heme, {} nucleotide, {} control\n      Total avaible dataset are composed of the following proportions:\n      {} heme, {} nucleotide, {} control".format(pockets[0]/1400, pockets[1]/1400,pockets[2]/1400,
                                                0.145, 0.380, 0.475))


# In[ ]:


data_onehot = data[0]
output = data[1]
X_train = data_onehot[0:1000,]
Y_train = output[0:1000,]
X_test = data_onehot[1000:,]
Y_test = output[1000:,]


# In[ ]:


my_model = model_light()


# In[ ]:


tf.test.is_gpu_available()
#my_model.fit(X_train, Y_train, epochs=50, batch_size=30)


# In[ ]:


history_mild_2mp = my_model.fit(X_train, Y_train, validation_data=(X_test, Y_test), epochs=30, batch_size=32)
my_model.save('light_model_2mp_e30_b32.h5')


# In[ ]:


#predictions=prediction_history()