MMIL: A novel algorithm for disease associated cell type discovery.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Single-cell datasets frequently lack individual cell labels, hindering the identification of disease-associated cells.
• Accurate cell classification is crucial for understanding disease mechanisms and developing targeted therapies.

Purpose of the Study:

• To introduce Mixture Modeling for Multiple Instance Learning (MMIL), a novel expectation-maximization method for training cell-level classifiers using only patient-level labels.
• To enable accurate cell classification in complex biological datasets, particularly when gold-standard cell labels are unavailable.

Main Methods:

• Developed MMIL, an expectation-maximization algorithm designed for multiple instance learning at the single-cell level.
• MMIL facilitates the training and calibration of various machine learning models, including logistic regression, gradient boosted trees, and neural networks.
• The method was validated on clinically-annotated primary patient samples from Acute Myeloid Leukemia (AML) and Acute Lymphoblastic Leukemia (ALL).

Main Results:

• MMIL accurately identified cancer cells in AML and ALL patient samples.
• The method demonstrated generalization capabilities across different tissues and treatment timepoints.
• MMIL successfully selected biologically relevant features, contributing to a deeper understanding of disease characteristics.
• The framework effectively integrates known cell labels when available, enhancing model training.

Conclusions:

• MMIL provides a powerful and flexible framework for cell classification using patient-level labels, addressing limitations of unlabeled single-cell data.
• This approach significantly advances disease understanding and management, particularly in high-dimensional and sparsely labeled biological contexts.
• MMIL offers a novel solution for leveraging both labeled and unlabeled data in machine learning for biomedical research.