Performance comparisons between clustering models for reconstructing NGS results from technical replicates.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• High-throughput DNA sequencing generates vast amounts of data, necessitating robust methods for accurate variant identification.
• Individual sequencing results can be noisy; combining data from multiple replicates and employing statistical models can improve callset quality.

Purpose of the Study:

• To evaluate and compare the performance of five different statistical clustering models for reconstructing high-performance DNA sequencing callsets.
• To identify optimal models for improving DNA variant calling accuracy using technical replicates.

Main Methods:

• Utilized three technical replicates of the NA12878 genome for analysis.
• Compared five models: consensus, latent class, Gaussian mixture, Kamila-adapted k-means, and random forest.
• Assessed performance using sensitivity, precision, accuracy, and F1-score metrics.

Main Results:

• The Kamila model achieved the best overall performance, increasing precision to over 99% while maintaining high sensitivity (98.8%).
• Gaussian mixture and random forest models also showed high precision (>99%) but with reduced sensitivity.
• Non-supervised clustering models generally improved sequencing performance compared to supervised models, particularly in precision and F1-score.

Conclusions:

• The Kamila and Gaussian mixture models offer significant improvements in precision and F1-score for DNA variant calling.
• These models are recommended for callset reconstruction from biological or technical replicates for applications in diagnostics and precision medicine.