TranScreen: Transfer Learning on Graph-Based Anti-Cancer Virtual Screening Model

• DOI: 10.3390/bdcc4030016
• Publication Date: 2020-06-29T07:40:07Z
• Authors: Milad Salem; Aminollah Khormali; Arash Keshavarzi ...; Julia Webb; Jiann-Shiun Yuan
• Keywords: Transfer of learning

Deep learning’s automatic feature extraction has proven its superior performance over traditional fingerprint-based features in the implementation of virtual screening models. However, these models face multiple challenges field early drug discovery, such as over-training and generalization to unseen data, due inherently unbalanced small datasets. In this work, TranScreen pipeline is proposed, which utilizes transfer learning a collection weight initializations overcome challenges. An amount 182 graph convolutional neural networks are trained on molecular source datasets learned knowledge transferred target task for fine-tuning. The p53-based bioactivity prediction, an important factor anti-cancer chosen showcase capability pipeline. Having models, three different approaches implemented compare rank them given before results show improvement model cases, with best increasing area under receiver operating curve ROC-AUC from 0.75 0.91 recall 0.25 1. This vital practical via lowering false negatives demonstrates potential learning. code pre-trained made accessible online.