A self-supervised machine learning pipeline for extracting information from live cell images at multiple doses and...

Area of Science:

• Cellular Biology
• Computational Biology
• Drug Discovery

Background:

• Live cells are dynamic systems, but traditional assays destroy them, missing temporal changes.
• Live brightfield imaging is label-free and scalable but suffers from low contrast and batch sensitivity.
• Existing computational methods for brightfield imaging are limited, hindering its potential.

Purpose of the Study:

• To develop a robust computational pipeline for analyzing live cell dynamics using brightfield imaging.
• To address the challenges of low contrast and batch sensitivity in brightfield microscopy.
• To extract subtle, time- and dose-dependent cellular responses for drug development.

Main Methods:

• Developed Live Cell Dynamics (LCD), a transformer-based pipeline using self-supervised learning.
• Implemented novel plane-agnostic augmentation and cross-batch sampling to handle imaging variations.
• Evaluated performance on phenotypic activity and Mechanism of Action (MoA) classification across multiple compounds and timepoints.

Main Results:

• The LCD pipeline significantly outperformed baseline methods in activity and MoA classification.
• Demonstrated the ability to detect compound polypharmacology from multi-dose/timepoint data.
• Enabled unsupervised nuclei detection and counting, highlighting the system's versatility.

Conclusions:

• LCD provides a scalable, cost-effective solution for analyzing live cell states from brightfield imaging.
• This approach facilitates the training of foundation models for detecting subtle cellular changes.
• The developed methods advance live cell imaging analysis for accelerated drug development.