Training a model for estimating leukocyte composition using whole-blood DNA methylation and cell counts as reference

Area of Science:

• Genomics
• Epigenetics
• Computational Biology

Background:

• Whole-blood DNA methylation is influenced by leukocyte composition, a confounder in epigenome-wide association studies (EWAS).
• Accurate leukocyte cell counts are often unavailable or impractical for large-scale studies.
• Existing computational methods rely on purified leukocyte reference datasets.

Purpose of the Study:

• To develop and validate a computational model for estimating leukocyte cell proportions directly from whole-blood DNA methylation data.
• To assess if this model can be trained without requiring purified leukocyte samples.
• To improve the accuracy of EWAS by accounting for cellular heterogeneity.

Main Methods:

• Trained a predictive model using whole-blood DNA methylation and five-part leukocyte cell counts from 175 participants.
• Evaluated the model on a separate subset of participants from the London Life Sciences Prospective Population Study (2445 participants total).
• Assessed the correlation between estimated and actual cell counts and the variance explained in DNA methylation levels.

Main Results:

• Achieved high correlations for neutrophils (0.85), eosinophils (0.88), lymphocytes (0.84), and monocytes (0.55).
• Estimated cell proportions explained more variance in whole-blood DNA methylation than raw cell counts.
• The model demonstrated precise estimation capabilities for common leukocyte types.

Conclusions:

• A novel computational approach enables accurate estimation of leukocyte cell proportions from whole-blood DNA methylation.
• This method bypasses the need for cell purification, making it suitable for large-scale genomic studies.
• The developed model enhances the reliability of EWAS by effectively addressing cellular composition confounding.