[Paper Review] DeepPurpose: a Deep Learning Library for Drug-Target Interaction Prediction and Applications to Repurposing and Screening

Kexin Huang, Tianfan Fu+4 more|arXiv (Cornell University)|Apr 19, 2020

Computational Drug Discovery Methods47 references17 citations

TL;DR

DeepPurpose is a user-friendly deep learning toolkit for drug-target interaction (DTI) prediction that enables both non-computational and computational researchers to perform virtual screening and drug repurposing using five pre-trained models with minimal code. It supports 15 drug and target encodings, 50+ deep learning architectures, and benchmark datasets, achieving state-of-the-art performance on standard DTI benchmarks and identifying clinically relevant candidates for COVID-19 repurposing.

ABSTRACT

Accurate prediction of drug-target interactions (DTI) enables drug discovery tasks, including virtual screening and drug repurposing, which can shorten the time to identify promising drug candidates and provide cures to patients. Recently, there is a growing number of research that developed deep learning (DL) models for DTI. Despite their superior performance, these research models are difficult to use in real drug discovery practice due to the complexity of deploying the research code as well as the restricted data formatting, model capacity, and evaluation setting. We present DeepPurpose, a comprehensive and easy-to-use software toolkit for DL based drug-target interaction (DTI) prediction with applications to drug screening and repurposing. The unique feature of DeepPurpose is that it enables non-computational drug development scientists to identify drug candidates based on five pre-trained DL models with only a few lines of codes. Further, computer scientists can use DeepPurpose to train customized DTI prediction models with 15 drug and target encodings and 50+ novel DL architectures. To tackle method development challenges, DeepPurpose also supports various data split settings and preloads five benchmarking datasets. We demonstrated that DeepPurpose allows users to obtain state-of-the-art prediction performance on several benchmark datasets. We also presented several case studies, including a study on drug repurposing for COVID-19, where promising drug candidates currently investigated in clinical trials are ranked high in DeepPurpose predictions.

Motivation & Objective

• To address the gap between advanced deep learning DTI models and practical usability in real-world drug discovery.
• To provide non-computational scientists with a simple interface to predict DTIs using pre-trained models with minimal coding.
• To offer computational researchers a flexible framework for training custom DTI models with diverse encodings and architectures.
• To standardize evaluation and model development by integrating multiple data split strategies and benchmark datasets.
• To support critical applications such as drug repurposing, demonstrated through a case study on COVID-19.

Proposed method

• The toolkit implements 50+ deep learning architectures for DTI prediction, including attention-based and graph neural network models.
• It supports 15 different drug and target encodings, such as SMILES, fingerprints, and protein sequences, to represent molecular and target features.
• DeepPurpose provides five pre-trained models that are fine-tuned on benchmark datasets and ready for inference with minimal user input.
• The framework supports various data split strategies, including random, scaffold-based, and time-based splits, to ensure robust model evaluation.
• It includes preloaded benchmark datasets such as Davis, KIBA, and PDB, enabling standardized performance comparison.
• The system is designed with a modular API that allows users to easily swap encodings, architectures, and evaluation protocols.

Experimental results

Research questions

• RQ1 Can a unified deep learning toolkit significantly lower the barrier to entry for non-computational drug discovery scientists in DTI prediction?
• RQ2 To what extent can pre-trained models in DeepPurpose achieve state-of-the-art performance on standard DTI benchmark datasets?
• RQ3 How effective is DeepPurpose in identifying repurposed drug candidates for emerging diseases like COVID-19?
• RQ4 Can the toolkit support flexible model development with diverse encodings and architectures for custom DTI modeling?
• RQ5 How does the integration of multiple data split strategies improve model generalization and evaluation reliability?

Key findings

• DeepPurpose achieves state-of-the-art performance on multiple benchmark DTI datasets, including Davis, KIBA, and PDB, through its combination of advanced architectures and robust data handling.
• The five pre-trained models in DeepPurpose enable non-expert users to predict DTIs with just a few lines of code, significantly reducing the barrier to entry for drug discovery teams.
• In a case study on drug repurposing for COVID-19, DeepPurpose ranked several drugs currently under clinical investigation highly, demonstrating its real-world predictive relevance.
• The toolkit supports 15 drug and target encodings and 50+ deep learning architectures, enabling high customization for computational researchers.
• DeepPurpose's support for diverse data split strategies enhances model evaluation reliability and generalization across different biological and chemical contexts.
• The integration of benchmark datasets and standardized evaluation protocols allows for consistent and reproducible model comparison across studies.