Fig. 2From: De novo design and bioactivity prediction of SARS-CoV-2 main protease inhibitors using recurrent neural network-based transfer learningOverview of the ULMFit approach. Initially, a general chemical model is trained to learn the “chemical language” contained in a collection of input molecules. The learned features can then be transferred to a target-task and adapted to the idiosyncrasies of the data. These “chemical models” can be used to generate molecules on demand. The last step consists of using the fine-tuned features to train a classifier that predicts bioactivityBack to article page