Prediction of spurious HLA class II typing results using probabilistic classification

Area of Science:

• Immunogenetics
• Bioinformatics
• Computational Biology

Background:

• High-throughput sequencing methods for HLA genotyping are generally accurate but susceptible to artefacts.
• Artefacts can arise from sample contamination, sequencing errors, or PCR biases, leading to spurious typing results.

Purpose of the Study:

• To evaluate the performance of probabilistic classifiers for detecting spurious HLA genotyping results.
• To assess the utility of population-specific genotype frequencies and workflow-specific predictors in improving typing accuracy.

Main Methods:

• Trained binary logistic regression and random forest models using high-resolution HLA-DRB1, DQB1, and DPB1 typing data from German, Polish, and UK donors.
• Validated models using 10-fold cross-validation and independent datasets.
• Incorporated workflow-specific predictors, such as DNA concentration and PCR reaction volume.

Main Results:

• Models based on genotype frequencies demonstrated high predictive capacity (AUC 0.820-0.893).
• Inclusion of workflow-specific predictors, particularly DNA concentration, significantly enhanced prediction specificity and accuracy (AUC 0.947-0.959).
• Low DNA concentration and low-volume PCR reactions were identified as key sources of error in the Fluidigm chip-based workflow.

Conclusions:

• Probabilistic models utilizing population-specific genotype frequencies are effective in identifying potentially spurious HLA typing results.
• Integrating workflow-specific parameters, such as DNA concentration, substantially improves the reliability and specificity of HLA genotyping quality control.