An adaptive method of defining negative mutation status for multi-sample comparison using next-generation sequencing.

Background: Multi-sample comparison is commonly used in cancer genomics studies. By using next-generation sequencing (NGS), a mutation's status in a specific sample can be measured by the number of reads supporting mutant or wildtype alleles. When no mutant reads are detected, it could represent either a true negative mutation status or a false negative due to an insufficient number of reads, so-called "coverage". To minimize the chance of false-negative, we should consider the mutation status as "unknown" instead of "negative" when the coverage is inadequately low. There is no established method for determining the coverage threshold between negative and unknown statuses. A common solution is to apply a universal minimum coverage (UMC). However, this method relies on an arbitrarily chosen threshold, and it does not take into account the mutations' relative abundances, which can vary dramatically by the type of mutations. The result could be misclassification between negative and unknown statuses.

Methods: We propose an adaptive mutation-specific negative (MSN) method to improve the discrimination between negative and unknown mutation statuses. For a specific mutation, a non-positive sample is compared with every known positive sample to test the null hypothesis that they may contain the same frequency of mutant reads. The non-positive sample can only be claimed as "negative" when this null hypothesis is rejected with all known positive samples; otherwise, the status would be "unknown".

Results: We first compared the performance of MSN and UMC methods in a simulated dataset containing varying tumor cell fractions. Only the MSN methods appropriately assigned negative statuses for samples with both high- and low-tumor cell fractions. When evaluated on a real dual-platform single-cell sequencing dataset, the MSN method not only provided more accurate assessments of negative statuses but also yielded three times more available data after excluding the "unknown" statuses, compared with the UMC method.

Conclusions: We developed a new adaptive method for distinguishing unknown from negative statuses in multi-sample comparison NGS data. The method can provide more accurate negative statuses than the conventional UMC method and generate a remarkably higher amount of available data by reducing unnecessary "unknown" calls.