Abstract:

The precise prediction of drug-target interactions is essential for drug discovery (DTI). Deep learning (DL) models have recently shown encouraging outcomes for DTI prediction. However, using these models may be difficult for bioinformaticians with little experience using DL as well as computer scientists who are fresh to the biomedical field. We present DeepPurpose, a comprehensive and user-friendly DL library for DTI prediction, in this paper. DeepPurpose provides training of custom DTI prediction models using chemical and protein encoders and more than 50 neural architectures, in addition to many other useful features. We demonstrate the state-of-the-art performance of DeepPurpose on various test datasets. Identification of potential Drug-Target Interactions is a critical step in the drug discovery and repositioning process as the efficacy of the presently available antibiotic treatment is declining. DeepPurpose uses an encoder-decoder framework for DTI prediction. The sources for DeepPurpose are a compound SMILES string and two protein amino acid sequences. The outcome of DeepPurpose is a number that evaluates the binding activity of the input compound protein pair. DeepPurpose, in particular, encodes the input protein and compound using a variety of deep learning encoders to obtain their deep embeddings, then concatenates and sends them into a decoder for a different deep neural network that tries to determine whether the input protein and compound bind.

Description:

Deep Learning, DTI prediction, DeepPurpose, CNN, drug encoders and decoders