Deep Learning Models for Activity Prediction Against the Low-data COVID-19 Targets

OBJECTIVE In response to Corona Virus Disease 2019(COVID-19), reusable drugs and new drugs against the low-data COVID-19 targets (with <300 known inhibitors) need to be discovered. METHODS Employing MolMapNet, a deep learning architecture that outperformed the state-of-the-art deep learning models on pharmaceutical benchmark datasets, new deep learning models were developed for predicting pharmaceutical properties with broadly-learned knowledge-based molecular representations. Predicted activities against 6 low-data COVID-19 targets with 34, 51, 81, 155, 161, 241 known inhibitors respectively. Compared with machine learning and deep learning models(with 5 478-10 000 known inhibitors) trained with targets in higher datasets. RESULTS Tested under the 10-fold cross-validation, our models predicted the activity values of the test-set inhibitors of these 6 targets with RMSE 0.442-0.917, MAE 0.358-0.749, and R² 0.436-0.761. CONCLUSION The screening of approved drugs for potential drug repurposing agents against COVID-19 identified 3 drugs that are consistent with the literature-reported experimental findings. These indicate the potential of our deep learning method for the low-data targets against COVID-19 and other diseases.