Discovering a sparse set of pairwise discriminating features in high-dimensional data.

Area of Science:

• Computational Biology
• Developmental Biology
• Genomics

Background:

• High-dimensional biological data, such as single-cell RNA sequencing (scRNA-seq), offer unprecedented insights into complex processes like cell differentiation.
• Extracting meaningful biological insights and mechanistic understanding from these datasets remains a significant challenge.
• Identifying key molecular regulators driving developmental transitions from scRNA-seq data is particularly difficult.

Purpose of the Study:

• To develop an unsupervised method for identifying informative features in high-dimensional biological data.
• To uncover a low-dimensional subspace where clusters representing distinct biological states become linearly separable.
• To improve the analysis of scRNA-seq data and facilitate the discovery of key regulators in developmental processes.

Main Methods:

• Feature identification based on cluster discrimination.
• Development of an unsupervised method to find a low-dimensional subspace for clustering.
• Ensemble averaging of discriminators trained on proposed cluster configurations.
• Application to single-cell RNA-seq data from mouse gastrulation.

Main Results:

• Identification of 27 key transcription factors from 409 tested in mouse gastrulation scRNA-seq data.
• 18 of the identified transcription factors are known regulators of cell states.
• Discovery of a low-dimensional subspace revealing clear signatures of known cell types previously unclassifiable.

Conclusions:

• Identifying informative features is crucial for effective statistical analysis and predictive modeling in high-dimensional biological data.
• The proposed unsupervised method can uncover hidden biological structures in complex datasets.
• This approach enhances the ability to classify cell types and infer regulatory mechanisms in developmental biology.