MalariaFlow: A comprehensive deep learning platform for multistage phenotypic antimalarial drug discovery.

Area of Science:

• Computational chemistry and pharmacology
• Machine learning applications in drug discovery
• Parasitology and infectious disease research

Background:

• Malaria drug resistance and transmission pose significant global health challenges.
• Existing machine learning (ML) and deep learning (DL) methods for antimalarial drug discovery require further optimization.
• Current research often overlooks diverse parasite strains and multi-stage activity prediction.

Purpose of the Study:

• To systematically compare the performance of various ML and DL models for antimalarial activity prediction.
• To evaluate models across different Plasmodium parasite phenotypes and life cycle stages.
• To identify advanced computational strategies for accelerating antimalarial drug discovery.

Main Methods:

• Curated a large benchmark dataset of 407,404 compounds and 410,654 bioactivity points.
• Compared fingerprint-based ML (RF::Morgan, XGBoost:Morgan) and graph-based DL models (GCN, GAT, MPNN, Attentive FP).
• Evaluated co-representation DL models (FP-GNN, HiGNN, FG-BERT) for incorporating chemical knowledge.

Main Results:

• The FP-GNN model achieved the highest predictive performance with an AUROC of 0.900.
• Co-representation DL models, particularly FP-GNN, excelled by integrating chemical knowledge.
• Fingerprint-based ML models were effective on larger datasets, but DL models showed better performance with integrated features.

Conclusions:

• FP-GNN demonstrates superior predictive power for antimalarial activity across multiple parasite stages.
• Co-representation DL models offer a promising approach to enhance antimalarial drug discovery by leveraging chemical insights.
• The developed MalariaFlow web server facilitates prediction, screening, and discovery of novel, multi-stage antimalarial drug candidates.