Method

Data Collection

Compound and training data was collected from the PubChem Online Database. PubChem comtains chemical information for more than 9 billion compounds and the results of thousands of experiments performed in labs across the world. For this analysis, we used experimental data from PubChem Bioassay 1285 [3]. This assay evaluated 193,771 compounds in their ability to inhibit levels of Amyloid Precursor Protein (APP). Of these 193,771 compounds, we were able to use the PubChem compound database to extract relevant chemical information (SMILES) for each of the compounds.

For the nomination data, chemical information for 10,000 compounds were downloaded from the ChEMBL database, another well known chemical library. The purpose of the nomination set is to nominate new compounds that have yet to be tested in a lab for their ability to reduce levels of the protein. For this reason the 10,000 compounds chosen were confirmed to not exist in the 193,771 compounds included in the original data.

Data Preparation

Morgan Fingerprints

Simple feed forward network requires input in the format of a matrix. So we have used API from RDKit library. It takes as input a SMILES string and provides an object of molecular structure. The object is transformed into vector format by method rdMolDescriptors.GetMorganFingerprintAsBitVect of RDKit library. This output is called morgan fingerprint.

DeepChem is a Python library/framework that is often used in the Deep Learning and Chemistry fields. The framework has integrated many deep learning models and functionalities from both PyTorch and TensorFlow. Following the extraction of standard chemical data from PubChem, graph representations of the molecules were needed to pass into our Graph Neural Network models. DeepChem provides a function, ConvMolFeaturizer() which easily converts approriately formatted graph representations by passing in SMILES representations of each compound into the function.

Due to the nature of drug discovery, the large majority of compounds evaluated for activity on proteins do not show any activity or impact on levels of the protein. This leads to a highly imbalanced dataset. In order to account for this, prior to training our models we used the DeepChem function BalancingTransformer() to minimize the impact of class imbalance on classification tasks. The BalancingTransformer function accomplishes this by balancing the class weights of both classes, to ensure similar impact on model training from both classes.

The data was split into training, validation, and test sets in a 80:10:10 ratio. Only 0.89% of tested compounds showed activity against the Amyloid Precursor Protein, thus we split the data using the RandomStratifiedSplitter() function within DeepChem. This ensured the training, validation, and test set had equally proportional data in terms of class balance.

Model Initialization and Training

The Support Vector Classifier from the scikit-learn Python library was used as the baseline model to compare with our deep learning models. The model was initialized using the following parameters:

• RGB kernel
• class_weight: balanced (to accomodate for class imbalance)
• Squared L2 penalty C = 10.0

Aside from the SVC model, we evaluated three different deep learning models using the listed optimization parameters:

Feed forward Network Model

• Two Feed forward network of size 1024 and 256
• Each feed forward Network is followed by dropout of 0.2
• ADAM optimizer
• Learning Rate 0.0001
• 100 epochs
• Cross-Entropy Loss Function

GraphConvModel
The GraphConvModel provided by DeepChem was used as the baseline Graph Neural Network model for our analyses. It is built on top of the tensorflow Python library. From this model we tuned the parameters to determine which architecture provided the best performance on our validation set. The architecture of the final model included:

• Four graph convolutional layers
• One feed-forward fully connected layer
• Dropout (p = 0.5)
• Batch Normalization
• ADAM optimizer
• Learning Rate 0.001
• 50 epochs
• Cross-Entropy Loss Function

Custom Graph Neural network

The custom GNN was tested with multiple different hyperparameters to find the best performing model. Among these tested parameters include:

• Five different variations with number of graph neural network layers = 2, 3, 4,5 and 6
• Feed forward fully-conntect layers of size 16 and 32 and a sigmoid function
• ADAM optimizer
• Learning Rate 0.0001
• 100 epochs
• Cross-Entropy Loss Function