Guidelines for cell-type heterogeneity quantification based on a comparative analysis of reference-free DNA...

Area of Science:

• Computational biology
• Genomics
• Cancer research

Background:

• Tumor cell-type heterogeneity impacts cancer progression and treatment response.
• DNA methylation in tumors can reveal cell-type proportions, but confounders like age and sex interfere.
• Existing reference-free algorithms for inferring cell-type proportions lack comparative evaluation.

Purpose of the Study:

• To comparatively evaluate computational pipelines for inferring tumor cell-type proportions from DNA methylation data.
• To identify critical steps and best practices for accurate cell-type proportion inference.
• To develop a standardized benchmark pipeline for community use.

Main Methods:

• Simulations were used to assess pipelines based on MeDeCom, EDec, and RefFreeEWAS software.
• Confounder adjustment, feature selection, and determining the number of cell types were investigated.
• Cattell's rule applied to scree plots was used to determine the optimal number of cell types.

Main Results:

• Removing confounder-correlated methylation probes reduced inference error by 30-35%.
• Selecting cell-type informative probes yielded similar error reduction.
• MeDeCom, EDec, and RefFreeEWAS showed comparable performance after preprocessing; performance improved with larger sample sizes and greater cell-type variation.
• Method sensitivity to initialization was observed, suggesting averaging solutions or optimizing initialization.

Conclusions:

• Accounting for confounders and performing feature selection are crucial for accurate cell-type proportion inference.
• A benchmark pipeline, implemented in the R package medepir, was developed to standardize validation and encourage further development.
• The medepir package facilitates community-driven improvement of computational methods for analyzing tumor heterogeneity.