If parental genotypes are provided, all patient heterozygous variants are examined for the possibility of using the parental variant calls to allocate a variant to either the maternal or paternal chromosome. If the pedigree information allows the phase to be determined, a confidence score of 1.0 is given.

Generally, it is not possible to assign de novo mutations to the correct allele by trio information. However, if the inherited allele is present in both parents, with one being heterozygous and the other being homozygous for said allele, it is probabilistically assigned to the homozygous parent with a confidence score of 0.66.

If DNA or RNA sequencing reads are provided, SmartPhase uses multi-variant spanning reads to aid in the resolution of local haplotypes. If a read spans two variant positions and contains both variants, this constitutes an evidence that both variants lie on the same allele in cis configuration. If two reads span both variant positions, each containing only one of the variants, this is an evidence that the variants lie on opposing chromosomes in trans configuration. To ensure the creation of accurate haplotypes, SmartPhase ignores reads that are not mapped, not part of a proper pair in case of paired-end reads, marked as duplicate reads, or not part of a primary alignment. Further, the user can choose to ignore reads whose mapping quality is lower than that of a defined threshold.

In order to inform about the quality of the inferred haplotype for two variants v₁ and v₂, a confidence score, Confidence(v₁, v₂), is computed by the formula | ∑i=1ntrans(1−∑k=1lv110−qk/10lv1︷Correctedv1+∑k=1lv210−qk/10lv2︷Correctedv22︸AverageInversePhred)︷TransSubscore−min(2∑j=1ncis(1−∑k=1lv110−qk/10lv1︷Correctedv1+∑k=1lv210−qk/10lv2︷Correctedv22︸AverageInversePhred),n)︷CisSubscore |n+2 (1) where n is the number of reads overlapping two variant positions, n_trans is the number of reads supporting a trans configuration where each variant is present in only half of the reads, n_cis is the number of reads containing both variants, q_k is the Phred quality score of a read at a particular position k, and l represents the length of the variant allele being examined in this read (either v₁ or v₂). The confidence score was designed to summarize the strength of the evidence behind a read-based phasing call. For a thorough explanation of how the confidence score is calculated, see Section 1.1 of S1 Appendix. In order to differentiate high quality phasing calls from low quality phasing calls, we derived 0.34 as a threshold for the confidence score as explained in Section 1.2 of S1 Appendix.

Variants are directly phased with their immediate neighbor using the cis and trans subscores, calculated by summing the inverse Phred score corrected evidence counts. This basic strategy permits the creation of seed haplotypes that can be locally extended to neighboring variants if overlapping reads exist. SmartPhase can elongate these haplotypes by applying this strategy within paired-end reads as the two paired-end reads must come from the same haplotype, regardless of physical distance or disjointedness. SmartPhase also leverages RNA sequencing reads that connect distant variants due to the read spanning exon-exon boundaries. To calculate the confidence of non-directly phased variants, the confidence scores of all directly phased variants on the shortest path are multiplied together. This helps to represent the growing uncertainty of phase calls as haplotype blocks increase in size and the distance between variants increases.

If both reads as well as parental variants are provided, the local haplotype blocks created by read-based phasing are combined using variants that were pedigree phased. Any contradictions between pedigree phasing and read-based phasing are resolved according to their confidence scores. All variant pairs not phased by direct evidences again have their confidence scores calculated by taking the product of all directly phased linking variants on the shortest possible path between the variants in the pair.

If parental genotype information is given, certain variant pair constellations can be designated as innocuous based on the assumption that the parents of the patient are healthy. Innocuous variants are those variants that are deemed to be clinically irrelevant as all variant combinations they partake in have been deduced to be non-disease-causing in Section 1.3 of S1 Appendix. Variant pairs are labeled as innocuous if one of the variants is homozygous in a parent or if mother, father, and child all possess the same heterozygous genotype for one of the variants in the pair.

If a variant is not visible in the alignment of the reads as given by the provided mapping file, this variant was most likely called as a result of read realignment done by the used variant calling program. As a consequence, these variants are designated as not found by SmartPhase. As the HaplotypeCaller (HC) tool of the Genome Analysis Toolkit (GATK) [11] is currently one of the most widely-used variant calling algorithms, we implemented the ability to incorporate phasing information returned by HC when no variant evidences were found within the reads and the variant could not be phased by trio phasing. If HC calls variants through read rearrangement, these variants are usually physically phased at the same time and the used local haplotype information is provided in the resulting variant file. The phase of otherwise missing variants is adopted from the variant files and given a confidence score of 1.0.

SmartPhase resolves haplotypes in genomic intervals of interest. Genomic intervals can either be directly defined by the user, or are generated by creating regions enveloping potential compound heterozygous variant pairs of interest. SmartPhase accepts up to two variant specifying files encompassing all variants and those that have been filtered to be deemed clinically relevant. The all variants file is used to create haplotype blocks and usually corresponds to the result of variant calling. The filtered variants file is optional, but can be used to narrow the scope of the output as only those variants specified in this file are printed in the final result. These variants generally constitute the set of variants that were filtered for clinical relevance according to allele frequency, predicted functional impact and other criteria. As many mapping files containing DNA or RNA sequencing reads as desired can be provided to be used during read-based phasing.

The phases and confidence scores of all heterozygous variant combinations within the given or created intervals are the results of SmartPhase. To fully capture the complexity of phased variant pairs, innocuous pairs, and variants that were not found in the mapping data, we developed a bitwise flag system to efficiently store all necessary information in a single number. The classification of variant pairs according to the defined criteria is visualized in Fig 1. Table A of S1 Appendix shows the possible combinations of bits and the corresponding final flag.

As described in Section 2 of S1 Appendix, we simulated WES data of the widely used CEU and YRI trio and phased their heterozygous variants in genes on chromosome 1 and 19 using SmartPhase and WhatsHap. We generated a set of 26, 638 potential heterozygous variant pairs distributed over 2, 922 genes with 4.21 heterozygous variants per gene on average (see Table B of S1 Appendix). Table 1 shows an overview of the main results of the benchmark (complete data in S1 Table).

Phasing in read only mode results in low amounts of phased pairs for both phasing tools. In comparison to WhatsHap, SmartPhase clears 2.8%–3.6% less pairs because variant alleles are often not found in the mapped reads as reported in the variant file. Variant calling tools, like the GATK HaplotypeCaller, rearrange read alignments internally before calling variants. As SmartPhase does not realign reads in contrast to WhatsHap, it performs worse in areas of uncertain mapping when only read information is provided. The results for phasing based on read and trio information demonstrate the power of SmartPhase, as the combination of innocuous labeling and phasing clears all input variant pairs. This corresponds to 10.9%–20.4% more variant pairs cleared in comparison to WhatsHap.

SmartPhase generated error-free predictions in the combined read & trio mode and all errors in read only mode are labeled as low quality. WhatsHap has an average error-rate of 1.03% (0.21%–1.89%) with a remarkable increase of the error rate in combined read & trio mode. Although this number is quite low, it can have detrimental consequences if even only one pair is wrongly predicted as being in cis configuration when in reality it is disease-causing. This emphasizes how crucial a confidence score is in generating trustworthy and accurate predictions.

Another advantage of SmartPhase is its runtime which is on average five times faster than WhatsHap, independent of using only read or both read and trio information. While the absolute difference is minor for the limited, simulated data, the runtime becomes particularly relevant in a clinical setting where variant pairs from all chromosomes of hundreds of patients must be phased.

We validated SmartPhase on a cohort of clinical WES data that consists of 121 trio and 800 singleton patients without parental genotype information. As detailed in Section 3 of S1 Appendix, we selected a set of 116, 613 potential compound heterozygous variant pairs after filtering for rare and protein-altering heterozygous autosomal variants. On average, we identified 126.62 ± 161.25 variant pairs per individual.

The underlying studies that generated human exome sequencing data that were used for the manuscript were approved by the Ethics Commission of the Medical Faculty of the LMU Munich. The reference numbers of the ethics votes are 346-11, 381-11, 387-11, 438-11, 486-11, 303-12, 187-13 BB, 353-13, 66-14, 501-14 and 806-16. The consent was obtained in written form.