Background

The absence of heterozygosity (AOH) is a kind of genomic change characterized by a long contiguous region of homozygous alleles in a chromosome [1]. Several underlying mechanisms of AOH have been reported, such as meiotic segregation errors [2], parental consanguinity [3], or complex chromosomal rearrangements [4]. AOHs do not necessarily have clinical consequences, however, they may cause serious pathogenic effects when it is related to imprinting effects [5] or autosomal recessive disease mechanisms [3]. For example, more than 25% of patients with Prader–Willi syndrome are caused by isodisomy (the inheritance of both homologs from a single parent and only one homolog of that parent is present) or heterodisomy (the inheritance of both homologs from a single parent and both homologs of that parent are present) [6]. Sahoo et al. found that whole-genome uniparental isodisomy (UPD) caused pregnancy loss in ~ 1% of cases [7]. In a study of rare autosomal trisomy by genome-wide noninvasive prenatal testing, the author found that 4.16% of cases with rare autosomal trisomies originate from uniparental disomy [15], while variants on the X chromosome were phased by Eagle2 (without the pedigree-based correction) [16]. Due to this inconsistency in variant phasing, the probability calculation of CNVseq-AOH may be influenced. So, we separately calculated the concordant rate on autosomes and the X chromosome.

For the 784 AOHs on autosomes, in general, the prediction sensitivity of CNVseq-AOH increased with depth (Fig. 2a). As expected, the sensitivity of CNVseq-AOH was 100% (784/784) when the depth was >  = onefold (Supplementary Table 1). With a depth of 0.5-fold, the sensitivity reached 99.9%. Only one AOH with an overlap of 47% was missed by CNVseq-AOH (Supplementary Table 1). The sensitivity of CNVseq-AOH reached 96.23% even with a depth of 0.1-fold, which is far lower than current studies, which need 4-to-fivefold depth [11, 17]. For the 79 AOHs on the X chromosome, the sensitivity of CNVseq-AOH was 59.49% (47/79) with a depth of 0.1-fold. The prediction sensitivity of CNVseq-AOH also increased with depth (Fig. 2b). However, even with a depth of threefold, the prediction sensitivity is still not 100%. There were 6 AOHs missed by CNVseq-AOH with an overlap ranging from 18%-44%. These 6 AOHs were located in similar regions on the X chromosome (Supplementary Table 1), which were also missed by CNVseq-AOH when using 0.5-fold and onefold depth. We further calculated the SNP numbers per 1 Mb on all the chromosomes in the 1KGP. The number of SNPs per 1 Mb on the X chromosome (mean of 26,839.9) was significantly less than the number of autosomes (18,439.9) (T-test, with P-value of 2.94E-12). One reasonable explanation for the relatively low sensitivity for AOHs on the X chromosome is that, compared with autosome, the variant information in the phasing results of the X chromosome in 1KGP was insufficient to calculate the probabilities for resampled reads.

Fig. 2
figure 2

Performance of CNVseq-AOH for 784 AOHs on autosome (a) and 79 AOHs on X chromosome (b) respectively

Full size image

For SNP-based microarrays, a threshold of >  = 10 Mb has been suggested for reporting AOH [18]. In the real clinical setting, AOH larger than 10 Mb in one chromosome is recommended for reporting the possibility of UPD [19, 20]. There were 291 AOH regions larger than 10 Mb in the 1KGP. For these AOHs, CNVseq-AOH can predict AOHs with a sensitivity of 100% (291/291) when the depth was >  = 0.5-fold (Supplementary Table 1). With a depth of 0.1-fold, the sensitivity reached 99.66% (290/291). CNVseq-AOH provided a prediction sensitivity of 94.5% (275/291) even with a depth of 0.05-fold.

For 0.1-fold LP-WGS data, it takes an average of 11 min to process a single sample using an 8-core CPU with 8 GB of RAM (from data alignment to reporting), including an average of 10 min for alignment, 25 s for feature learning, and 10 s for AOH prediction and reporting.

Additional AOHs detected by CNVseq-AOH

Compared to AOHs detected by CMA, additional AOHs were detected by CNVseq-AOH. We analyzed additional AOHs detected by CNVseq-AOH with a depth of 0.1-fold. A total of 267 additional AOHs were detected in the 409 samples by CNVseq-AOH, approximately 0.65 AOHs for each sample. The number of the additionally detected AOHs decreased with the length of AOH (Supplementary Fig. 1). Using high-coverage data, we further validated these AOHs by visualization using an in-house script. The results showed that, 50.56% (135/267) additional AOHs were true positives (Supplementary Table 2; Supplementary Fig. 2). In the clinical setting, a threshold of > 10 Mb was recommended for reporting the possibility of UPD [19, 20]. Using a threshold of > 10 Mb, only 5 additional AOHs were detected by CNVseq-AOH for the 409 samples with 0.1-fold depth (Supplementary Table 2).

Interestingly, we found an AOH region (seq[GRCh38] hmz(6)(p12.3q12) chr6:g. 47568317_64568317hmz) using CNVseq-AOH, which crossed the centromeric regions of chromosome 6 in this case (Fig. 3c, d). Although with sufficient markers for this region (Fig. 3a), no AOH was reported in this region by CMA, which indirectly reflects the detection performance of CNVseq-AOH for regions crossing the centromeric regions. This AOH was further validated using high-coverage data, which also showed positive signals in this region (Fig. 3b).

Fig. 3
figure 3

AOH region (seq[GRCh38] hmz(6)(p12.3q12) chr6:g. 47568317_64568317hmz) detected by CNVseq-AOH in chromosome 6 of HG01980. A Marker of HumanOmni2.5–4 (SNP-based microarray kit used in the 1KGP) in chromosome 6. Marker density in this region of the sample represents sufficient markers for the chr6:g. 47568317_64568317 region. However, no AOH was reported in this region by CMA, which indirectly reflects the detection performance of CNVseq-AOH for regions crossing the centromeric regions; b The number of heterozygous SNP number (yellow line) and all SNP number (green line) in this region calculated using high coverage WGS data. The heterozygous SNP number (yellow line) in the chr6:g. 47568317_64568317 region was close to 0, which indicated potential AOH events in this region; c Log-likelihood ratio for haploid and diploid in each bin. Each dot represents the mean log-likelihood ratio in each bin. For potential AOH regions, the log-likelihood ratio tends to approach 0, with relatively sparse blue dots above 0; d AOH prediction likelihoods of CNVseq-AOH using 0.1-fold depth. The higher the confidence, the closer it is to 1, indicating that the region we predicted is likely to be AOH regions; e Chromosome 6

Full size image

RNN VS. Hidden Markov model

RNN and Hidden Markov Model (HMM) are both widely used models for processing sequential data. HMM, a probabilistic model, is particularly effective for problems involving time series data. Currently, no published literature employs the HMM method for the detection of AOH, hence it cannot be cited. In this study, HMM with Gaussian emissions (the “hmmlearn.hmm.GaussianHMM” module in Python) was used for AOH prediction. We established an HMM model with 5 hidden states and a full covariance matrix, and compared it with CNVseq-AOH for AOH prediction. As a result, the prediction sensitivity of CNVseq-AOH is better than the HMM-based method with differing depths (Fig. 4).

Fig. 4
figure 4

Method comparison for the 863 AOHs. a Performance of CNVseq-AOH; b Performance of HMM

Full size image

Validation of CNVseq-AOH with 6 clinical cases

We further applied CNVseq-AOH on 6 clinical cases with previously detected AOHs (Table 1). A mean depth of 0.573-fold (raw reads) was obtained for each sample. Uniquely aligned high-quality reads (UAHRs) reads were used for the detection of AOH. A UAHR was defined as a read that was uniquely aligned to the human genome reference with a quality value of more than 20 per base (containing no partial adapter sequences and no more than 5% that were not determined in the read length).

Table 1 Validation of CNVseq-AOH with 6 clinical cases
Full size table

As a result, CNVseq-AOH detected all the 15 AOH regions (Table 1). In some cases with multiple known AOHs (Case 1, Case 2, and Case 5), a greater number of AOH regions were detected by CMA, probably because several AOH regions were split into sub-regions by CMA.

Discussion

RNN, known as recurrent neural network, is a very popular class of neural network. RNN is especially useful with sequential data. The neuron in RNN can use the internal state to “memory” previous input information, combining the information of the current input, to determine the next output state. RNN was widely used in natural language processing (NLP) [21]. However, the application of RNN in human genomic research is still rare. In this study, we described an RNN-based method, CNVseq-AOH, for predicting the absence of heterozygosity using LP-WGS. To the best of our knowledge, CNVseq-AOH is the first application combining population-based haplotype information, adjustable sliding windows, and RNN in genetic testing. CNVseq-AOH shows the feasibility of using ultra-low sequencing depth for the detection of clinically significant AOHs and demonstrates its potential in genetic testing.

One of the key innovations of CNVseq-AOH is the use of population-based haplotype information. Based on our testing, population-based haplotype information greatly influenced the feasibility of CNVseq-AOH. For the 409 samples in the current study, ancestry-matched populations (or genetically similar population) were used for analysis. We further compared the sensitivity using ancestry-matched populations for feature learning and using all available haplotype information from multiple ethnicities for feature learning at 0.1-fold. As a result, using a threshold of 50% overlap with the AOHs detected by CMA, the sensitivity of CNVseq-AOH reached 93.40% (806/863) when using ancestry-matched populations for feature learning. When switching to the strategy using all available haplotype information from multiple ethnicities for feature learning, the sensitivity is only 60.95% (526/863). Simultaneously, when employing a strategy using all available haplotype information from multiple ethnicities for feature learning, the accuracy is also significantly impacted (Supplementary Fig. 3). Not all the populations are captured in the 1KGP. The number of samples in a specific population varied a lot. This may influence the accuracy of our method and impede the wide application of CNVseq-AOH. Expanding the data collection to include new populations and samples may solve the problem.

One limitation of CNVseq-AOH is that it cannot be used for the detection of mosaic AOH. So, we did not include the 4 cases with mosaic AOH for testing in the first place. Based on the signals for these 4 cases (Supplementary Fig. 4), CNVseq-AOH possesses the potential for predicting mosaic AOH. This may require a different model and a great number of ascertained positive cases with mosaic AOHs for model training, which is an interesting topic but beyond the scope of this study. Another limitation of the current study is the performance of CNVseq-AOH for the detection of AOHs on the X chromosome. With a depth of 0.1-fold, a detection sensitivity of only 59.49% was achieved for the 79 AOHs on the X chromosome in the 1KGP. In phase three of the 1KGP, variants on the X chromosome were phased without the pedigree-based correction using Eagle2 (http://ftp.1000genomes.ebi.ac.uk/vol1/ftp/data_collections/1000G_2504_high_coverage/working/20201028_3202_phased/README_SNV_INDEL_phasing_111822.pdf), resulting in less number of SNPs per 1 Mb. So, the information of biallelic SNPs in the VCF file for the X chromosome is insufficient to calculate the probabilities for resampled reads. Actually, this is not a limitation of CNVseq-AOH, which means that, with sufficient information in the reference panel, CNVseq-AOH also possesses the potential to provide high prediction sensitivity for AOHs located on the X chromosome. Next, we plan to reanalyze these samples to optimize the performance of CNVseq-AOH for the detection of AOHs on the X chromosome.

In this study, we investigated sequencing depth on model performance in the 1KGP. In general, the prediction sensitivity of CNVseq-AOH increased with sequencing depth. However, data in the 1KGP was generated using various sequencing parameters (different sample types, library construction protocols, sequencing platforms, etc.), so the evaluation of sequencing depth may be biased. For clinical laboratories, depth evaluation using real clinical samples and uniform sequencing parameters is necessary before clinical application.

Conclusions

In summary, we developed a method for predicting the absence of heterozygosity using LP-WGS data, which overcomes the sparse nature of typical LP-WGS by combing population-based haplotype information, adjustable sliding windows, and RNN. Next, we plan to apply our method to clinical pregnant women who underwent prenatal diagnosis, thereby further evaluating the performance and potential utility of CNVseq-AOH under realistic clinical scenarios.