SoftHybrid: A Hybrid Imputation Algorithm Optimized for Single-Cell Proteomics Data

Area of Science:

• Proteomics
• Computational Biology
• Single-cell analysis

Background:

• Missing values (MVs) are a major challenge in proteomics, especially in single-cell studies due to low sample amounts and detection limits.
• Existing imputation methods often struggle with missing-not-at-random (MNAR) data and are primarily designed for bulk analyses, leading to compromises in data quality.

Purpose of the Study:

• To develop a novel, data-driven imputation framework for single-cell proteomics that addresses the limitations of existing methods.
• To improve the accuracy and reliability of proteomics data analysis by effectively handling MNAR data.

Main Methods:

• Introduced SoftHybrid, a framework that jointly models missingness and protein abundance to estimate MNAR probability.
• Enabled continuous weighting between missing-at-random (MAR) and MNAR imputation strategies without requiring external priors.
• Developed a fully unsupervised application for imputation.

Main Results:

• SoftHybrid demonstrated superior performance compared to existing methods on benchmark datasets, particularly at low input levels.
• The method matched or exceeded the performance of other approaches at the minibulk level.
• Preserved proteomic structure and abundance accuracy, leading to enhanced recovery of biological signals.

Conclusions:

• SoftHybrid offers a robust and unsupervised solution for imputing missing values in single-cell proteomics data.
• The framework improves the biological insights obtainable from complex proteomic datasets.
• SoftHybrid is available as an R package, facilitating its adoption in the research community.