Generic Multi-Task Learning Framework for Structural Mixed Data using Deep Neural Network Feature Sharing Architecture-Application to Predictive Tasks in Vietnamese Herbal Medicine
Abstract
While the development of new drugs is costly, time-consuming, and often accompanied by safety issues, drug repurposing, where old drugs with established safety are used for medical conditions other than originally developed, is an attractive alternative. In this project, our purpose is to use a Graph convolutional neural network to solve several tasks in drug discovery. First, drug-mediated toxicity is a heavy burden to the pharmaceutical industry, leading to safety-related failures in development and the high cost of drug discovery. Second, we consider the anti-cancer screening task which assesses the positive or negative of drug response to different types of cancer in humans. The success of deep learning on high-throughput chemical structure screening has offered unprecedented opportunities to detect toxicity compound candidates or negative responses to cancer cells. However, capturing graph structure and a very high number of toxicity/cancer type tasks of chemical structure has been challenging. We proposed to build a multitask graph neural network which is based on AdaShare capsule networks, graph convolutional neural network, and multi-task learning to predict the anti-cancer and toxic effect which represent in the chemical structure graph. We implemented two methods for toxicity testing prediction and anti-cancer screening by using the adaptive sharing features. The first method is adaptive sharing based on residual graph network and policy network and the second method is Capsule graph network with Task routing network. The graph-based AdaShare showed a potential performance on Tox21 datasets. Single task CapsGNN shows good performance. Multitask CapsGNN has been implemented and is ready for the training process, but the results have not been obtained yet. We continue to work on this and test on NCI and Tox21 datasets. We expect that the capsule mechanism of Multitask CapsGNN achieves better performance with a smaller number of parameters.
Document Details
- Document Type
- Technical Report
- Publication Date
- Nov 09, 2021
- Accession Number
- AD1154600
Entities
People
- Le Ly
Organizations
- Ho Chi Minh City International University