Optimal deconvolution of transcriptional profiling data using quadratic programming with application to complex...

Area of Science:

• Computational biology
• Genomics
• Biostatistics

Background:

• Molecular profiling aids biomarker discovery but complex human samples obscure analysis.
• Heterogeneity in clinical trial samples can negatively impact statistical studies.
• Accurate quantification of cell types is crucial for interpreting complex biological data.

Purpose of the Study:

• To develop and validate a computational method for deconvoluting cell type proportions from mixed-species transcriptional data.
• To address the challenge of sample heterogeneity in clinical trial molecular profiling.
• To improve the accuracy of statistical analyses on complex biological samples.

Main Methods:

• Utilized a linear latent variable model to represent mixed cell population expression as a weighted average of pure cell types.
• Employed quadratic programming for efficient, non-negative estimation of cell fractions.
• Applied the method to diverse platforms and benchmark datasets with known mixing fractions.

Main Results:

• Demonstrated accurate prediction of cell type mixing fractions on benchmark datasets.
• Successfully estimated proportions for over ten circulating cell types, including rare populations (<10%).
• Identified accurate leukocyte trafficking changes with Fingolomid (FTY720) treatment, consistent with prior methods.

Conclusions:

• The developed method effectively deconvolutes cell proportions from complex transcriptional data.
• This approach offers a robust solution for analyzing heterogeneity in clinical trial samples.
• The method enhances the reliability of molecular profiling for biomarker discovery and disease understanding.