cascAGS: Comparative Analysis of SNP Calling Methods for Human Genome Data in the Absence of Gold Standard.

The development of third-generation sequencing has accelerated the boom of single nucleotide polymorphism (SNP) calling methods, but evaluating accuracy remains challenging owing to the absence of the SNP gold standard. The definitions for without-gold-standard and performance metrics and their estimation are urgently needed. Additionally, the possible correlations between different SNP loci should also be further explored. To address these challenges, we first introduced the concept of a gold standard and imperfect gold standard under the consistency framework and gave the corresponding definitions of sensitivity and specificity. A latent class model (LCM) was established to estimate the sensitivity and specificity of callers. Furthermore, we incorporated different dependency structures into LCM to investigate their impact on sensitivity and specificity. The performance of LCM was illustrated by comparing the accuracy of BCFtools, DeepVariant, FreeBayes, and GATK on various datasets. Through estimations across multiple datasets, the results indicate that LCM is well-suitable for evaluating callers without the SNP gold standard, and accurate inclusion of the dependency between variations is crucial for better performance ranking. DeepVariant has a higher sum of sensitivity and specificity than other callers, followed by GATK and BCFtools. FreeBayes has low sensitivity but high specificity. Notably, appropriate sequencing coverage is another important factor for precise callers' evaluation. Most importantly, a web interface for assessing and comparing different callers was developed to simplify the evaluation process.