A Deep Learning Methodology for Screening New Natural Therapeutic Candidates for Pharmacological Cardioversion and Anticoagulation in the Treatment and Management of Atrial Fibrillation

Tim Dong*, Rhys Llewellyn, Melanie J Hezzell, Gianni D Angelini

*Corresponding author for this work

Research output: Contribution to journalArticle (Academic Journal)peer-review

Abstract

Background:
The treatment and management of atrial fibrillation poses substantial complexity. A delicate balance in the trade-off between the minimising risk of stroke without increasing the risk of bleeding through anticoagulant optimisations. Natural compounds are often associated with low-toxicity effects, and their effects on atrial fibrillation have yet to be fully understood. Whilst deep learning (a subtype of machine learning that uses multiple layers of artificial neural networks) methods may be useful for drug compound interaction and discovery analysis, graphical processing units (GPUs) are expensive and often required for deep learning. Furthermore, in limited-resource settings, such as low- and middle-income countries, such technology may not be easily available.

Objectives:
This study aims to discover the presence of any new therapeutic candidates from a large set of natural compounds that may support the future treatment and management of atrial fibrillation anywhere using a low-cost technique. The objective is to develop a deep learning approach under a low-resource setting where suitable high-performance NVIDIA graphics processing units (GPUs) are not available and to apply to atrial fibrillation as a case study.

Methods:
The primary training dataset is the MINER-DTI dataset from the BIOSNAP collection. It includes 13,741 DTI pairs from DrugBank, 4510 drug compounds, and 2181 protein targets. Deep cross-modal attention modelling was developed and applied. The Database of Useful Decoys (DUD-E) was used to fine-tune the model using contrastive learning. This application and evaluation of the model were performed on the natural compound NPASS 2018 dataset as well as a dataset curated by a clinical pharmacist and a clinical scientist.

Results:
the new model showed good performance when compared to existing state-of-the-art approaches under low-resource settings in both the validation set (PR AUC: 0.8118 vs. 0.7154) and test set (PR AUC: 0.8134 vs. 0.7206). Tenascin-C (TNC; NPC306696) and deferoxamine (NPC262615) were identified as strong natural compound interactors of the arrhythmogenic targets ADRB1 and HCN1, respectively. A strong natural compound interactor of the bleeding-related target Factor X was also identified as sequoiaflavone (NPC194593).

Conclusions:
This study presented a new high-performing model under low-resource settings that identified new natural therapeutic candidates for pharmacological cardioversion and anticoagulation.
Original languageEnglish
Article number1323
Number of pages18
JournalBiomedicines
Volume13
Issue number6
DOIs
Publication statusPublished - 28 May 2025

Bibliographical note

Publisher Copyright:
© 2025 by the authors.

Research Groups and Themes

  • Bristol Heart Institute

Fingerprint

Dive into the research topics of 'A Deep Learning Methodology for Screening New Natural Therapeutic Candidates for Pharmacological Cardioversion and Anticoagulation in the Treatment and Management of Atrial Fibrillation'. Together they form a unique fingerprint.

Cite this