ELF: Extract Landmark Features By Optimizing Topology Maintenance, Redundancy, and Specificity.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Feature selection is crucial for high-dimensional data with limited samples, especially in single-cell biology.
• Challenges include high dimensionality, noise, data sparsity, redundant features, and poor sample labeling.

Purpose of the Study:

• To propose a novel model, Extract Landmark Features (ELF), to address challenges in feature selection.
• To simultaneously maintain sample relationships, minimize feature redundancy, and maximize feature specificity.

Main Methods:

• ELF employs a nonlinear combinatorial optimization approach.
• A heuristic algorithm based on a greedy strategy is proposed to solve the optimization problem.

Main Results:

• ELF demonstrates outstanding performance on single-cell RNA-seq datasets (direct reprogramming and hepatoblast differentiation).
• It selects hundreds of landmark genes to maintain cell correlativity, showing the importance of topology maintenance, redundancy removal, and specificity.
• ELF also proves effective in image datasets for identifying pivotal pixels in handwriting and faces.

Conclusions:

• ELF is a valuable tool for obtaining interpretable results by identifying key features.
• It effectively clusters samples while revealing essential biological functions and patterns.
• The model's general applicability extends beyond biological data to image analysis.