Graph Convolutions For Tox21

In this notebook, we will explore the use of TensorGraph to create graph convolutional models with DeepChem. In particular, we will build a graph convolutional network on the Tox21 dataset.

Let’s start with some basic imports.

from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals

import numpy as np
import tensorflow as tf
import deepchem as dc
from deepchem.models.tensorgraph.models.graph_models import GraphConvModel

Now, let’s use MoleculeNet to load the Tox21 dataset. We need to make sure to process the data in a way that graph convolutional networks can use For that, we make sure to set the featurizer option to ‘GraphConv’. The MoleculeNet call will return a training set, an validation set, and a test set for us to use. The call also returns transformers, a list of data transformations that were applied to preprocess the dataset. (Most deep networks are quite finicky and require a set of data transformations to ensure that training proceeds stably.)

# Load Tox21 dataset
tox21_tasks, tox21_datasets, transformers = dc.molnet.load_tox21(featurizer='GraphConv')
train_dataset, valid_dataset, test_dataset = tox21_datasets
Loading dataset from disk.
Loading dataset from disk.
Loading dataset from disk.

Let’s now train a graph convolutional network on this dataset. DeepChem has the class GraphConvModel that wraps a standard graph convolutional architecture underneath the hood for user convenience. Let’s instantiate an object of this class and train it on our dataset.

model = GraphConvModel(
    len(tox21_tasks), batch_size=50, mode='classification')
# Set nb_epoch=10 for better results., nb_epoch=1)
/home/leswing/miniconda3/envs/deepchem/lib/python3.5/site-packages/tensorflow/python/ops/ UserWarning: Converting sparse IndexedSlices to a dense Tensor of unknown shape. This may consume a large amount of memory.
  "Converting sparse IndexedSlices to a dense Tensor of unknown shape. "
Starting epoch 0
Ending global_step 126: Average loss 590.739
TIMING: model fitting took 7.317 s

Let’s try to evaluate the performance of the model we’ve trained. For this, we need to define a metric, a measure of model performance. dc.metrics holds a collection of metrics already. For this dataset, it is standard to use the ROC-AUC score, the area under the receiver operating characteristic curve (which measures the tradeoff between precision and recall). Luckily, the ROC-AUC score is already available in DeepChem.

To measure the performance of the model under this metric, we can use the convenience function model.evaluate().

metric = dc.metrics.Metric(
    dc.metrics.roc_auc_score, np.mean, mode="classification")

print("Evaluating model")
train_scores = model.evaluate(train_dataset, [metric], transformers)
print("Training ROC-AUC Score: %f" % train_scores["mean-roc_auc_score"])
valid_scores = model.evaluate(valid_dataset, [metric], transformers)
print("Validation ROC-AUC Score: %f" % valid_scores["mean-roc_auc_score"])
Evaluating model
computed_metrics: [0.80045699830862893, 0.83618637604367374, 0.83908539936708681, 0.77873855933094183, 0.67692252993044244, 0.75578036941489168, 0.75895796821704797, 0.70234314980793855, 0.76387081283102387, 0.65924917162534913, 0.78448201448364341, 0.76675448900822296]
Training ROC-AUC Score: 0.760236
computed_metrics: [0.71533171721169553, 0.74090608465608465, 0.81106357802757933, 0.70627859684799188, 0.63177272727272715, 0.6326016835811501, 0.61491865697473169, 0.71286314850043442, 0.67676006592889104, 0.51656328658755846, 0.75414979999520937, 0.6603359173126615]
Validation ROC-AUC Score: 0.681129

What’s going on under the hood? Could we build GraphConvModel ourselves? Of course! The first step is to create a TensorGraph object. This object will hold the “computational graph” that defines the computation that a graph convolutional network will perform.

from deepchem.models.tensorgraph.tensor_graph import TensorGraph

tg = TensorGraph(use_queue=False)

Let’s now define the inputs to our model. Conceptually, graph convolutions just requires a the structure of the molecule in question and a vector of features for every atom that describes the local chemical environment. However in practice, due to TensorFlow’s limitations as a general programming environment, we have to have some auxiliary information as well preprocessed.

atom_features holds a feature vector of length 75 for each atom. The other feature inputs are required to support minibatching in TensorFlow. degree_slice is an indexing convenience that makes it easy to locate atoms from all molecules with a given degree. membership determines the membership of atoms in molecules (atom i belongs to molecule membership[i]). deg_adjs is a list that contains adjacency lists grouped by atom degree For more details, check out the code.

To define feature inputs in TensorGraph, we use the Feature layer. Conceptually, a TensorGraph is a mathematical graph composed of layer objects. Features layers have to be the root nodes of the graph since they consitute inputs.

from deepchem.models.tensorgraph.layers import Feature

atom_features = Feature(shape=(None, 75))
degree_slice = Feature(shape=(None, 2), dtype=tf.int32)
membership = Feature(shape=(None,), dtype=tf.int32)

deg_adjs = []
for i in range(0, 10 + 1):
    deg_adj = Feature(shape=(None, i + 1), dtype=tf.int32)

Let’s now implement the body of the graph convolutional network. TensorGraph has a number of layers that encode various graph operations. Namely, the GraphConv, GraphPool and GraphGather layers. We will also apply standard neural network layers such as Dense and BatchNorm.

The layers we’re adding effect a “feature transformation” that will create one vector for each molecule.

from deepchem.models.tensorgraph.layers import Dense, GraphConv, BatchNorm
from deepchem.models.tensorgraph.layers import GraphPool, GraphGather

batch_size = 50

gc1 = GraphConv(
    in_layers=[atom_features, degree_slice, membership] + deg_adjs)
batch_norm1 = BatchNorm(in_layers=[gc1])
gp1 = GraphPool(in_layers=[batch_norm1, degree_slice, membership] + deg_adjs)
gc2 = GraphConv(
    in_layers=[gp1, degree_slice, membership] + deg_adjs)
batch_norm2 = BatchNorm(in_layers=[gc2])
gp2 = GraphPool(in_layers=[batch_norm2, degree_slice, membership] + deg_adjs)
dense = Dense(out_channels=128, activation_fn=tf.nn.relu, in_layers=[gp2])
batch_norm3 = BatchNorm(in_layers=[dense])
readout = GraphGather(
    in_layers=[batch_norm3, degree_slice, membership] + deg_adjs)

Let’s now make predictions from the TensorGraph model. Tox21 is a multitask dataset. That is, there are 12 different datasets grouped together, which share many common molecules, but with different outputs for each. As a result, we have to add a separate output layer for each task. We will use a for loop over the tox21_tasks list to make this happen. We need to add labels for each

We also have to define a loss for the model which tells the network the objective to minimize during training.

We have to tell TensorGraph which layers are outputs with TensorGraph.add_output(layer). Similarly, we tell the network its loss with TensorGraph.set_loss(loss).

from deepchem.models.tensorgraph.layers import Dense, SoftMax, \
    SoftMaxCrossEntropy, WeightedError, Stack
from deepchem.models.tensorgraph.layers import Label, Weights

costs = []
labels = []
for task in range(len(tox21_tasks)):
    classification = Dense(
        out_channels=2, activation_fn=None, in_layers=[readout])

    softmax = SoftMax(in_layers=[classification])

    label = Label(shape=(None, 2))
    cost = SoftMaxCrossEntropy(in_layers=[label, classification])
all_cost = Stack(in_layers=costs, axis=1)
weights = Weights(shape=(None, len(tox21_tasks)))
loss = WeightedError(in_layers=[all_cost, weights])

Now that we’ve successfully defined our graph convolutional model in TensorGraph, we need to train it. We can call fit(), but we need to make sure that each minibatch of data populates all four Feature objects that we’ve created. For this, we need to create a Python generator that given a batch of data generates a dictionary whose keys are the Feature layers and whose values are Numpy arrays we’d like to use for this step of training.

from deepchem.metrics import to_one_hot
from deepchem.feat.mol_graphs import ConvMol

def data_generator(dataset, epochs=1, predict=False, pad_batches=True):
  for epoch in range(epochs):
    if not predict:
        print('Starting epoch %i' % epoch)
    for ind, (X_b, y_b, w_b, ids_b) in enumerate(
            batch_size, pad_batches=pad_batches, deterministic=True)):
      d = {}
      for index, label in enumerate(labels):
        d[label] = to_one_hot(y_b[:, index])
      d[weights] = w_b
      multiConvMol = ConvMol.agglomerate_mols(X_b)
      d[atom_features] = multiConvMol.get_atom_features()
      d[degree_slice] = multiConvMol.deg_slice
      d[membership] = multiConvMol.membership
      for i in range(1, len(multiConvMol.get_deg_adjacency_lists())):
        d[deg_adjs[i - 1]] = multiConvMol.get_deg_adjacency_lists()[i]
      yield d

Now, we can train the model using TensorGraph.fit_generator(generator) which will use the generator we’ve defined to train the model.

# Epochs set to 1 to render tutorials online.
# Set epochs=10 for better results.
tg.fit_generator(data_generator(train_dataset, epochs=1))
Starting epoch 0
Ending global_step 251: Average loss 530.84
TIMING: model fitting took 6.949 s

Now that we have trained our graph convolutional method, let’s evaluate its performance. We again have to use our defined generator to evaluate model performance.

metric = dc.metrics.Metric(
    dc.metrics.roc_auc_score, np.mean, mode="classification")

def reshape_y_pred(y_true, y_pred):
    TensorGraph.Predict returns a list of arrays, one for each output
    We also have to remove the padding on the last batch
    Metrics taks results of shape (samples, n_task, prob_of_class)
    n_samples = len(y_true)
    retval = np.stack(y_pred, axis=1)
    return retval[:n_samples]

print("Evaluating model")
train_predictions = tg.predict_on_generator(data_generator(train_dataset, predict=True))
train_predictions = reshape_y_pred(train_dataset.y, train_predictions)
train_scores = metric.compute_metric(train_dataset.y, train_predictions, train_dataset.w)
print("Training ROC-AUC Score: %f" % train_scores)

valid_predictions = tg.predict_on_generator(data_generator(valid_dataset, predict=True))
valid_predictions = reshape_y_pred(valid_dataset.y, valid_predictions)
valid_scores = metric.compute_metric(valid_dataset.y, valid_predictions, valid_dataset.w)
print("Valid ROC-AUC Score: %f" % valid_scores)
Evaluating model
computed_metrics: [0.83463194036351052, 0.86218739964675661, 0.84894031662657832, 0.80217986671584707, 0.70559942152332189, 0.79751934844253025, 0.8103057689046107, 0.71659210162938414, 0.80849247997327445, 0.72071717294380933, 0.83433314746710274, 0.78304357554399506]
Training ROC-AUC Score: 0.793712
computed_metrics: [0.78221936377578793, 0.78993055555555547, 0.81705388431256543, 0.77777071682765631, 0.66802272727272727, 0.67197702777122181, 0.64295604015230179, 0.72305596655628368, 0.74692724275959499, 0.63050611290902547, 0.80023473616134511, 0.73880275624461667]
Valid ROC-AUC Score: 0.732455

Success! The model we’ve constructed behaves nearly identically to GraphConvModel. If you’re looking to build your own custom models, you can follow the example we’ve provided here to do so. We hope to see exciting constructions from your end soon!