DRUG EFFICACY PREDICTION

Overview:

The project aims to build a model to predict drug efficacy of molecules.

A pre-processed HIV dataset with 3 classess (CA - Confirmed active, CM - Confirmed moderately active, CI - Confirmed inactive and benign) is available at: http://moleculenet.ai/datasets-1. The raw data is available here: https://wiki.nci.nih.gov/display/NCIDTPdata/AIDS+Antiviral+Screen+Data
We perform stratified random splitting on the dataset: 80% of the images are in train set and 20% of the images are in test set.

Model Architecture

1. Classical Machine learning with 1D, 2D and 3D molecular descriptors

• We perform feature extraction by contructing a graph structure representing the molecules and calculating 1D descriptors (logP, Ring Count, Aromaticity, etc), 2D descriptors(MQNs, Morgan Fingerprints, 2D Pharmacophore Fingerprints, etc) and 3D descriptors (MORSE, WHIM, etc) for the molecules.
• Using these molecular descriptors, we train base models with different hyper parameters
• The predictions made by these base models are added to the feature space. We then train Level 1 Model Stacking using Cross-Validation on trainset

2. Graph Convolution Network to learn neural fingerprints

• Instead of molecular descriptors, we can train a graph convolutional network on the molecules' graph structures to learn vector representation of molecules. https://arxiv.org/abs/1609.02907

3. Deep learning and Transfer learning to learn image features (representation learning with Convolutional Neural Network)

• Instead of constructing graph representation of molecules, we can use a recurrent neural network on the molecules' SMILES string to learn vector representation of molecules. https://arxiv.org/pdf/1712.02034.pdf

Model Performance

1. Evaluation Metrics

We will look at several evaluation metrics:

• Accuracy: Accuracy simply measures proportion of predictions that are correct. . It ranges between 0 (worst) and 1 (best)
• F1-Score: F1-Score is the harmonic mean of precision and recall. It ranges between 0 (worst) and 1 (best)
• Cohen’s kappa: Cohen’s kappa statistics measures how agreeable the prediction and the true label are. It ranges between -1 (completely disagree) and 1 (completely agree).

2. Performance Report

The models' performance on test set is reported below. The best model is a convolutional neural network with InceptionV3 backbone.

Model	Accuracy	F1-Score	Cohen’s kappa

Confusion Matrix of the best model's prediction on test set:

Instruction