Exponential family measurement error models for single-cell CRISPR screens

Area of Science:

• Genomics
• Computational Biology
• Systems Biology

Background:

• CRISPR genome engineering and single-cell RNA sequencing are powerful tools for biological discovery.
• Single-cell CRISPR screens integrate these technologies to link genetic perturbations to gene expression changes.
• Analyzing these screens presents significant statistical challenges, hindering the full potential of the technology.

Purpose of the Study:

• To identify and address statistical limitations in current single-cell CRISPR screen analysis methods.
• To introduce a novel, robust statistical method for analyzing single-cell CRISPR screen data.
• To provide a scalable computational framework for applying the new method to large datasets.

Main Methods:

• Theoretical and real data analyses were used to evaluate existing methods like thresholded regression.
• A new method, GLM-EIV (GLM-based errors-in-variables), was developed, extending classical errors-in-variables models.
• A computational infrastructure was built for cloud and high-performance cluster deployment.

Main Results:

• Standard methods like thresholded regression exhibit attenuation bias and a bias-variance tradeoff.
• GLM-EIV effectively handles exponential family-distributed responses and noisy predictors, accounting for confounding variables.
• Application to two large-scale datasets revealed new biological insights.

Conclusions:

• GLM-EIV overcomes the limitations of existing methods for single-cell CRISPR screen analysis.
• The developed computational infrastructure enables efficient analysis of large-scale single-cell CRISPR screen data.
• This work facilitates deeper understanding of regulatory networks and disease mechanisms through advanced statistical approaches.