PMC

When analyzing family trios, multi-allelic loci can look like de novo CNVs.

Schematics of mean-shift procedure. For each bin, i.e., data point, the mean-shift vector points in the direction of bins with the most similar RD signal. Segment breakpoints are determined where two neighboring vectors have opposite directions but do not point to each other.

Examples of multi-allelic loci. (A) Tri-allelic locus with CN0, CN1, and CN2 is at Hardy-Weinberg equilibrium. (B) Distribution of genotypes across a population can be explained by hexa-allelic locus with CN0–CN5. (C) Tri-allelic locus that is not at Hardy-Weinberg equilibrium, which may indicate natural selection. In this case, the distribution of genotypes peaks around 3, with the likely explanation that an equal proportion of CN1 and CN2 alleles at this locus dominate the population. This, in turn, implies balancing selection.

Distribution of normalized average RD signal for predicted CNVs (for a CEPH daughter) that are >1 kb and pass the q0 filter. The normalization factor is the double (two copies of each chromosome) of the genome-wide average RD signal. Two clear peaks (around 0 and 1) correspond to homozygous and heterozygous deletions. Slight displacement of the second peak (∼0.05) from a value of 1 is the result of read over-mapping in those regions, when choosing a mapping location for nonuniquely mapped reads (see Methods). Peaks for duplications are smeared, which reflects the larger variations in the RD signal and, as a consequence, the greater challenge in detecting and genotyping duplications.

Cartoon demonstration of the adaptive procedure for an increase in bandwidth H_b. (A) When the band is 2, then the largest contribution to mean-shift vector calculations, e.g., for the cyan bin, comes from two neighboring bins. Following the partitioning, two bins within one segment get “frozen,” and bins within it are excluded from partitioning on the next step. New partitioning allows for freezing of more bins that are skipped at the next step when bandwidth equals 4. (B) The deletion region is clearly seen by the eye but could not be detected as a whole at a bandwidth of 2. Only a small portion is detected as CNV and gets “frozen.” After new partitioning with a bandwidth of 3, the region is not frozen anymore and is included for partitioning on the next step (bandwidth of 4), where the complete region of deletion is detected.

Statistics on partitioning the RD signal for a child in a CEPH trio (NA12878) using 100-bp bins and standard parameters (see Methods). (A) Average RD signal distribution in produced segments. The distribution has three clear peaks: around the genomic RD average (no CNVs), half of that (heterozygous deletion), and one and one-half of that (duplication of one haplotype). The average genomic RD signal is ∼ 77 reads. Not all partitioned regions with abnormal RD are called CNVs by the statistical significance test. Therefore, the area under each peak is not representative of the corresponding fraction of CNVs. (B) Distribution of the average RD signal difference for neighboring segments. The distribution is for the absolute value of the difference and shows that either produced segments have similar average signals (peak around zero) or their average signals are approximately half of the genomic average RD signal (second peak), indicating deletion/duplication of one haplotype. (C) Example of partitioning clarifying clusters in D. (D) Distribution of the average RD signal difference at the left and right boundary for each segment. The distribution has several clear clusters. Clusters originate due to various combinations of segments with different RD signals. Clusters 8 and 9 represent cases of enclosed events, such as duplication of a region within duplication.