Self-supervised Learning for Drug Synergy Prediction with Small Data Set
Conference proceedings article
Authors/Editors
Strategic Research Themes
Publication Details
Author list: Klaikeaw, Tanakrit; Piyayotai, Supanida; Phunchongharn, Phond; Termsaithong, Teerasit;
Publisher: Institute of Electrical and Electronics Engineers Inc.
Publication year: 2023
Start page: 625
End page: 630
Number of pages: 6
ISBN: 979-835032353-5
Languages: English-Great Britain (EN-GB)
Abstract
Drug combination is a promising approach to enhance treatment effectiveness and mitigate the adverse effects associated with individual drugs. However, the substantial number of possible combinations and the time-intensive nature of cell response testing present formidable challenges. To overcome these challenges, deep learning techniques have been employed to effectively process complex data. Nevertheless, the reliance of deep learning models on extensive labeled medical data poses significant barriers. To overcome this limitation, the utilization of unlabeled data has emerged as a viable alternative. Therefore, we investigated the application of self-supervised learning to capture the molecular structures of drugs in conjunction with transfer learning, to predict the synergistic effects resulting from drug combinations on various types of cancer cells. It is assumed that the integration of self-supervised learning techniques enhances the predictive performance concerning the synergistic effects exhibited by drug pairs. The experiments were conducted by tuning parameters, including token size (i.e., 1, 3, 5, 10, and 37) and model architecture (i.e., Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and Convolutional LSTM) with the self-supervised learning process. Additionally, we examined the influence of training data size (ranging from 100 to 5% of the total data) and model architecture (CNN, LSTM, and Convolutional LSTM) on drug synergy prediction. The findings demonstrated that, among the self-supervised learning parameters, the Convolutional LSTM model with a token size of 37 showed the highest accuracy of 83.07%. Furthermore, with the knowledge of chemical structures acquired by self-supervised learning, the self-supervised learning technique reduced the mean square error (MSE) by 0.32-4.22% without the self-supervised process. ฉ 2023 IEEE.
Keywords
Drug combination, Drug Synergy, Self-Supervised Learning
Contact Information