A Structure-Based Deep Learning Framework for Correcting Marine Natural Products' Misannotations Attributed to...

Area of Science:

• Marine Natural Products Chemistry
• Bioinformatics
• Drug Discovery

Background:

• Marine natural products (MNPs) are crucial for drug discovery but often misannotated in databases due to host-microbe symbiosis.
• Inaccurate data impedes biosynthetic studies and artificial intelligence (AI)-driven discovery of novel therapeutics.

Purpose of the Study:

• To develop a structure-based workflow for classifying the origins of MNPs and correcting misannotations in marine datasets.
• To enhance the accuracy of natural product databases for improved AI-driven drug discovery.

Main Methods:

• Integrated a two-step data cleaning strategy using CMNPD and NPAtlas compound data.
• Employed a microbial-pretrained graph neural network to detect label inconsistencies and filter structural outliers.
• Classified compound origins based on chemical structure and predicted microbial associations.

Main Results:

• Achieved 85.56% balanced accuracy in origin classification.
• Identified 3996 compounds with predicted microbial origins conflicting with Animalia labels.
• Detected biologically coherent structural patterns within putative symbiotic metabolites.

Conclusions:

• The developed framework offers a scalable quality-control method for natural product databases.
• Improved accuracy supports more reliable biosynthetic gene cluster (BGC) tracing and host selection.
• Facilitates more effective AI-driven discovery of marine natural products.