Predicting drug responses from drug chemical features by using deep neural networks

• Main Author: Krithika Ramamoorthy
• Other Authors: Jagath C Rajapakse
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision
• Online Access: https://hdl.handle.net/10356/147943
• Institution: Nanyang Technological University

Increasing incidence of numerous forms of cancer leads it to be the deadliest disease globally. Hence, it is important to predict drug responses because when doctors are aware which types of drugs will cure cancer for the patients, suitable treatments can be given on time. Predicting anti-cancer drug response of a cancer patient precisely continues to be a serious issue till now. Due to insufficient accessible information and shortcomings in algorithms, this is considered to be challenging. The mechanisms of cancers are also extremely complex that currently there are no ideal cancer drug procedures at present. Hence, we decided to develop feature matrices using Simple Molecular Input Line Entry System (SMILES) notation of drugs where we combined multiple types of data in order to provide a more holistic molecular perspective on cancer research. The SMILES string is expressed in a distributed representation which is named as SMILES feature matrix. SMILES comprises 2 groups of characters which are atomic symbols and SMILES original symbols. The symbols of atoms are derived from periodic tables. For the original symbols to represent in bonds in SMILES notation, it is applied using OpenSMILES documentation. One feature vector comprises 42 features of which 21 features are allocated for symbols to indicate while the remaining 21 features are assigned for original SMILES symbols. The atomic substance quantities such as degree, charge and chirality were taken into consideration by calculating their numerical values using RdKit software. Hence, various software and documentations are utilised in order to develop feature matrices which are the input for deep learning. With the development of deep learning, the main aim of this project is to use different types of neural networks such as deep neural networks (DNN), convolutional neural networks (CNN) and graph convolutional network (GCN) to predict the response of cancer to various anticancer drugs using chemical features by converting the SMILES feature matrix into a low-dimensional feature vector. Hyper-parameter tuning and early stopping are performed in order to find out the optimal model. So that clinicians will be able to take a more advanced approach in molecular level to prescribe anti-cancer drugs that are likely to be more powerful.