Interpretation, extrapolation and perturbation of single cells.

Area of Science:

• Computational biology
• Systems biology
• Genomics

Background:

• Single-cell analyses are advancing from descriptive atlasing to inferring causal relationships.
• Machine learning methods are crucial for analyzing large-scale observational and interventional single-cell datasets.

Purpose of the Study:

• To review and connect machine learning approaches for single-cell data analysis.
• To provide a unifying ontology for selecting appropriate methods for biological questions.
• To identify future computational directions in the field.

Main Methods:

• Review of machine learning methods including representation learning, causal inference, mechanistic discovery, disentanglement, and population tracing.
• Categorization of methods based on modeling concepts, assumptions, and downstream tasks.
• Development of a unifying ontology for method selection.

Main Results:

• A comprehensive overview and connection of various machine learning approaches for single-cell data.
• A proposed ontology to guide researchers in method selection.
• Identification of promising future research avenues and underutilized data properties.

Conclusions:

• Machine learning is essential for uncovering mechanistic relationships in single-cell data.
• A structured approach and ontology can facilitate the application of advanced computational methods.
• Further research into computational methods and data properties will advance single-cell biology.