Automated classification of cellular expression in multiplexed imaging data with Nimbus.

Area of Science:

• Computational Biology
• Bioinformatics
• Digital Pathology

Background:

• Multiplexed imaging is crucial for analyzing tissue spatial topography in health and disease.
• Accurate cell phenotyping requires enumerating marker combinations, often using unsupervised clustering.
• Existing methods may require dataset-specific retraining, limiting broad applicability.

Purpose of the Study:

• To develop a deep learning model for predicting marker positivity in multiplexed imaging data.
• To create a large-scale dataset (Pan-Multiplex) for training and validating the model.
• To enable robust cell subtype identification by integrating model predictions with clustering algorithms.

Main Methods:

• Construction of the Pan-Multiplex (Pan-M) dataset with 197 million marker expression annotations across 15 cell types.
• Development of Nimbus, a pretrained deep learning model for classifying cell marker expression (positive/negative).
• Validation of Nimbus predictions against staining patterns and comparison with existing methods.

Main Results:

• Nimbus accurately predicts marker positivity across diverse cell types, tissues, and microscopy platforms without retraining.
• The model captures underlying staining patterns and matches or surpasses the accuracy of previous approaches.
• Integration of Nimbus predictions with clustering algorithms robustly identifies cell subtypes.

Conclusions:

• Nimbus provides a powerful, generalizable tool for analyzing multiplexed imaging data.
• The open-sourced Nimbus model and Pan-M dataset facilitate community-driven advancements in spatial biology.
• This approach enhances the efficiency and accuracy of cell phenotyping and subtype discovery in complex biological samples.