Astir

Astir assigns cell types in single-cell multiplexed imaging and proteomics datasets using a probabilistic framework that integrates marker protein priors and machine learning for automated, reference-free cell-type annotation.

Key Features:

• Probabilistic Modelling: Astir employs a probabilistic model that integrates prior knowledge of marker proteins to assign cells to known or unknown cell types.
• Deep Recognition Neural Networks: Utilizing deep recognition neural networks, Astir performs fast Bayesian inference to enable scalable annotation.
• Scalability and Implementation: Implemented in PyTorch, Astir scales to datasets comprising millions of single cells and supports multiple single-cell technologies and antibody panels, including Imaging Mass Cytometry (IMC).
• Reference-free Annotation: Astir provides cell type annotations without requiring previously annotated reference datasets, enabling discovery of uncharacterized cell types.

Scientific Applications:

• Tumor Microenvironment Analysis: Analyze the spatial architecture of tumor microenvironments, quantify immune cell influx, and assess spatial heterogeneity in patient samples.
• Cancer Research: Support studies of disease processes such as cancer initiation and progression by providing accurate cell type annotations across imaging and proteomics modalities.
• Multiplexed Imaging and Proteomics Studies: Apply to Imaging Mass Cytometry, suspension mass cytometry, and microscopy datasets to annotate cellular heterogeneity across antibody panels.

Methodology:

Astir integrates prior biological knowledge of marker proteins into a probabilistic model and uses deep recognition neural networks to perform fast Bayesian inference via stochastic variational inference; the implementation uses PyTorch.