Fully unsupervised deep mode of action learning for phenotyping high-content cellular images

Area of Science:

• Computational biology
• Bioimage analysis
• Machine learning

Background:

• Phenotype identification from high-content screening images is challenging.
• Existing methods rely on image analysis pipelines, supervised learning, or pre-trained neural networks.
• These approaches often require cell candidate extraction or pre-trained models on non-cellular data.

Purpose of the Study:

• To develop a novel unsupervised deep learning algorithm for clustering cellular images based on Mode-of-Action (MOA).
• To directly process entire images using only pixel intensity values, eliminating the need for cell candidate extraction.
• To correct for batch effects during training and enable discovery of novel MOAs.

Main Methods:

• An unsupervised deep learning algorithm was developed to cluster cellular images.
• The algorithm uses only pixel intensity values as input.
• It incorporates batch effect correction during the training process.

Main Results:

• The method achieved 97.09% accuracy in classifying MOA on a labeled dataset using nearest neighbor matching.
• The approach can be trained on unannotated datasets, enabling the discovery of novel MOAs.
• The algorithm successfully distinguishes treatments by their effect on proliferation.

Conclusions:

• The unsupervised deep learning algorithm offers a powerful and direct approach to cellular image analysis and MOA discovery.
• The method's ability to train on full-resolution images and handle unannotated data simplifies application and expands discovery potential.
• The developed algorithm can identify novel MOAs and annotate unlabeled compounds, advancing drug discovery and biological research.