Self-Supervised Learning for Drug Discovery Using Nematode Images: Method and Dataset.

Area of Science:

• Computational biology
• Parasitology
• Computer vision

Background:

• Parasitic worms cause significant human and livestock diseases, necessitating the search for new drug candidates.
• Natural product extracts are screened using the nematode C. elegans as a model organism.
• Microscopy image analysis of C. elegans is crucial but time-consuming.

Purpose of the Study:

• To create a novel dataset of C. elegans microscopy images for natural product extract screening.
• To develop an automated image classification method to accelerate phenotype analysis.
• To improve the efficiency of identifying potential anti-parasitic compounds.

Main Methods:

• A dataset of 12,717 C. elegans microscopy images was curated, with a portion labeled by experts.
• A two-stage Semi-supervised Mix-up Barlow Twins Nematode Classifier (MBT-NC) was proposed.
• The MBT-NC integrates self-supervised learning (Barlow Twins) with supervised classification for feature representation and task execution.

Main Results:

• The MBT-NC dataset comprises 12,717 C. elegans microscopy images.
• The MBT-NC model achieved superior performance in binary, six-class, and 27-class classification tasks.
• Outperformed fully supervised and other self-supervised methods by 3.2%, 1.0%, and 2.2% in test accuracy, respectively.

Conclusions:

• The MBT-NC classifier effectively analyzes C. elegans phenotypes for drug discovery.
• The publicly available dataset and model advance computer vision applications in healthcare and parasitology.
• This research accelerates the identification of natural products with potential anti-parasitic activity.