Prediction of spurious HLA class II typing results using probabilistic classification

Abstract While modern high-throughput sequence-based HLA genotyping methods generally provide highly accurate typing results, artefacts may nonetheless arise for numerous reasons, such as sample contamination, sequencing errors, read misalignments, or PCR amplification biases. To help detecting spur...

Full description

Saved in:
Bibliographic Details
Published in:Human immunology 2016-03, Vol.77 (3), p.264-272
Main Authors: Schöfl, Gerhard, Schmidt, Alexander H, Lange, Vinzenz
Format: Article
Language:English
Subjects:
Alleles
Allelic dropout
Allergy and Immunology
Area Under Curve
Classification
Computational Biology - methods
Gene Frequency
Genotype
Genotyping error
Genotyping Techniques
Histocompatibility Antigens Class II - classification
Histocompatibility Antigens Class II - genetics
Histocompatibility Testing - methods
HLA class II
Humans
Random forest
Reproducibility of Results
ROC Curve
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Abstract While modern high-throughput sequence-based HLA genotyping methods generally provide highly accurate typing results, artefacts may nonetheless arise for numerous reasons, such as sample contamination, sequencing errors, read misalignments, or PCR amplification biases. To help detecting spurious typing results, we tested the performance of two probabilistic classifiers (binary logistic regression and random forest models) based on population-specific genotype frequencies. We trained the model using high-resolution typing results for HLA-DRB1, DQB1, and DPB1 from large samples of German, Polish and UK-based donors. The high predictive capacity of the best models replicated both in 10-fold cross-validation for each gene and in using independent evaluation data (AUC 0.820–0.893). While genotype frequencies alone provide enough predictive power to render the model generally useful for highlighting potentially spurious typing results, the inclusion of workflow-specific predictors substantially increases prediction specificity. Low initial DNA concentrations in combination with low-volume PCR reactions form a major source of stochastic error specific to the Fluidigm chip-based workflow at DKMS Life Science Lab. The addition of DNA concentrations as a predictor variable thus substantially increased AUC (0.947–0.959) over purely frequency-based models.
ISSN:0198-8859
1879-1166
DOI:10.1016/j.humimm.2016.01.012