Probe intensity data for array-based resequencing in three regions. In each region, the log₂ intensities of four probes for the sense strand in 15 consecutive base positions are depicted. Each point represents one 25mer probe, coded for the base of interest in its center position (T—black circle; G—red square; C—green diamond; A—blue triangle), with reference sequences shown on the top of each panel. The only variants in these regions are at position 8 as confirmed by Sanger sequencing. The heterozygous AG in region (A) has equal intensities in probes RM_A and AM_G, whereas the heterozygous CA in region (B) has higher intensity in probe RM_C than AM_A, which in turn is higher than the other AM probes. This suggests region (B) is a harder position to call heterozygous. The homozygous variant GG (versus reference AA) in region (C) causes lower probe intensities in all flanking positions, all of which (RMs and AMs) have one off-center base mismatching the target sequence. This phenomenon is called the footprint effect.

Evaluation of PCR amplification products (amplicons) using resequencing arrays. The distribution of the median log ratio D and the median average intensity T for all 17 520 PCR-amplicons (438 amplicons × 40 samples) are shown. The points highlighted in red correspond to amplicons with a lower reference call rate of R < 0.9. All amplicons with the lowest values in D and T coincide with the red points and are removed from the subsequent analysis.

Examples on SRMA base-calling using the single-chip (A) and then multi-chip (B–F) procedure. (A) Shown is a smoothed density plot of the log ratio δ’s: θ_RM – θ_AM, versus the average intensity A’s: (θ_RM+θ_AM)/2 for selected AM’s on a single array, where the darker color indicates higher density of data points. For one position, the X’s in red are where we expect δ to be for the three variant classes RR, RS and SS, respectively. We compared these values to the observed δ (green circle) and called this position as RR. We finalized base-calling at positions (86%) where all samples present high confidence RR calls from the single-chip procedure. We next present five positions highlighted in yellow and of varying shapes (circle, square, diamond, point-up triangle, point-down triangle and star) for this array in (A) and also in individual δ plots (B–F for positions 1–5). These plots represent our multi-chip procedure, which is based on the distribution of δ’s with selected AM’s at one position in all 40 samples. MAC denotes minor allele count. The y-axis is for sense strand (+) and the x-axis is for antisense strand (–). Black is reference, blue is heterozygous variant and red is homozygous variant. The ellipses represent 90% confidence regions of the component distributions in our multi-chip model. (B) Model-based classification across samples identified additional reference-only positions where one cluster of positive δ’s contained all samples. (C) In common SNP positions, each variant cluster contains enough samples to allow our clustering algorithm to estimate the right number of clusters, the varying shapes and locations of the clusters and assign each sample to a cluster. (D) The multi-chip clustering is more accurate than the single-chip calling when the location of δ’s is abnormal. (E, F) In rare SNP positions, we perform classification based on observed reference cluster and pre-specified locations and shapes of the variant clusters. This method identified one or two variant samples (out of 40) as heterozygous and homozygous.

FDR versus FNR for SRMA and GSEQ at varying cutoffs of the quality scores. The cutoffs are 0–1 for the modified silhouette score in SRMA and 0–46 for the quality score (log likelihood ratio) in GSEQ. Each base-caller has three curves: for all SNPs (black), heterozygous SNPs only (blue) and homozygous SNPs only (red). The results for the default cutoffs of 0.67 (diamond) and 3 (circle) are shown for the two base-callers, respectively. As expected, heterozygous SNPs are more difficult to call and have higher error rates than homozygous SNPs. For all three categories, SRMA out-performed GSEQ by at least one order of magnitude.

Candidate gene analysis of mtDNA maintenance disorders. (A) The 39 genes include 10 genes causing mtDNA maintenance disorders (green nodes), 15 disease genes with suspected functions in mtDNA metabolism and/or similarities in their disease phenotypes (purple nodes), and 14 non-disease candidate genes with functions in mtDNA maintenance (grey nodes) (Supplementary Table S2). Gene relationships are predicted through functional interactions (likelihood ratios; LR) (20) and disease-phenotypic similarities (quantitative phenotypic associations; QPA) (21) with the edge colors: red—gene pairs with highest correlation of LR and QPA; blue—gene pairs with LR only; and orange—gene pairs with QPA only (≥0.4). The six genes at bottom left have multiple functional interactions to this network through intermediate genes (data not shown). Five disease genes (asterisk) are associated with dominant inheritance patterns and four of these genes (C10Orf2, OPA1, POLG, SLC25A4) are also causing recessive disorders (17,18). In this study we identified novel DNA variants in several cases in POLG, APEX2, MUTYH, TOP1MT (underlined). (B) The table shows amino acid changes (acronym p. not shown) identified in only the 23 cases and not controls. Five of these 14 genes (in bold) are known to cause mtDNA maintenance disorders (green nodes in A). The color codes indicate PolyPhen’s functional predictions (38) with ‘possibly damaging’ (yellow) and ‘probably damaging’ (orange), while all other amino acid changes are predicted benign (Supplementary Table S6). Positions that escaped array detection are shown in parenthesis, (asterisk) indicate dbSNP positions, and (caret symbol) are cases with rare variants in multiple genes hypothesizing a synergistic effect on the disease phenotype.