Robust classification using average correlations as features (ACF).

Area of Science:

• Bioinformatics
• Computational Biology
• Data Science

Background:

• Omics technologies like single-cell transcriptomics often generate datasets with substantial missing values.
• Traditional approaches either discard incomplete data, leading to information loss, or use imputation, which can introduce errors.
• Developing methods for accurate classification in the presence of missing data is crucial for advancing omics research.

Purpose of the Study:

• To introduce a novel, imputation-free classification method for omics data.
• To address the challenges posed by missing values in large-scale biological datasets.
• To provide a flexible and high-performing alternative to existing classification techniques.

Main Methods:

• Developed a new classification approach named average correlations as features (ACF).
• ACF trains machine learning models using inter-class and intra-class correlations.
• The method leverages pairwise correlations as a metric for classification.

Main Results:

• ACF significantly outperforms established methods like K-nearest-neighbor (KNN) and distribution-based classifiers.
• Demonstrated strong classification performance on real-world single-cell RNA sequencing and proteomics datasets.
• Simulation studies confirmed the robustness and effectiveness of the ACF approach.

Conclusions:

• Average correlations as features (ACF) provides a flexible, missing-value-tolerant classification method for omics data.
• ACF enables classification with minimal data loss, overcoming limitations of traditional imputation or data exclusion methods.
• Variants of ACF offer enhanced performance and flexibility, compatible with other machine learning techniques.