Descriptor free models for anti-malaria drugs discovery
We built deep learning descriptor-free models for anti-malaria and compared them against the high performing models.
Malaria is one of the most severe infectious diseases in the world. Due to the spread of drug resistance to the marketed antimalarial drugs, the discovery and development of new antimalarial drugs is one of the most pressing challenges. QSAR plays a crucial role in the field of drug design; however, to build QSAR models will involve applying a descriptor selection algorithm to find descriptors from large and diverse datasets. In addition, the descriptors are hard to explain as they are related to the target activity and are indirect representations of chemical structures. As a result, it is difficult and time-consuming to find descriptors to build QSAR models from large datasets.
We took efforts in taking descriptor-free models one step forward by using Gated Recurrent Unit (GRU) on SMILES and SELFIES, and Graph Convolution Network (GCN) on the graph data of the SMILES. We compared these models with a QSAR descriptor and ECFP-trained Random Forest model as a control comparison. We also compared our achieved results with the results obtained in other studies of descriptor-free models. The results of this study indicated that descriptor-free models are viable for large datasets thus saving time and computational power to generate descriptors.