Sophia Gibson: 'Results suggest that episignatures can be discovered directly from sequencing data, capture signals that arrays miss, and carry over across sequencing platforms, but there is still work to do before they are ready for clinical use.'
[Editor's Note: Sophia Gibson, a Ph.D. Candidate in the lab of Danny Miller, M.D., Ph.D., is the first author on the recently released preprint paper, "EpiSignaLR: an open resource for genome-wide episignature analysis from methylation-aware short- and long-read sequencing.' She also spoke on the paper October 26 at the fourth annual Long-Read Seqencing Symposium. Here, she summarizes the findings.]
Episignatures are reproducible, disorder-specific patterns of genome-wide DNA methylation. They have been described for more than 100 Mendelian conditions, most of which are chromatinopathies, caused by loss of function of chromatin-modifying genes. Episignatures are most often used clinically in two scenarios: reclassifying variants of uncertain significance (VUS) in episignature-associated genes, and suggesting diagnoses for individuals who remain unsolved after exome or genome sequencing. Clinical testing, however, still relies almost entirely on methylation arrays, which analyze DNA that undergoes a chemical process called bisulfite conversion before being analyzed. That means a separate assay, usually ordered after a non-diagnostic or uncertain sequencing result, with analysis limited to the ~930,000 CpGs on the latest array, or roughly 3% of those in our genome.
Long-read sequencing is an emerging technology that can simultaneously identify CpG methylation differences, along with sequence variation. In principle, episignature analysis could be analyzed along with data already generated for diagnosis. Earlier work has shown that array-derived classifiers can be applied to methylation-aware sequencing data, but those classifiers work better for some conditions than others (and it turns out to be really hard to predict which ones are going to work well, and which ones won’t). As an example, the array-derived methylation classifier correctly identified only about a third of individuals with SETD1B-related neurodevelopmental disorder (SETD1B-NDD) we have sequenced in the Miller Lab on the Oxford Nanopore Technologies (ONT) long read sequencing platform. My work focused on whether signatures discovered directly from long-read sequencing would perform better, and whether they would work with data generated on other sequencing platforms.
We started by calling differentially methylated regions (DMRs) genome-wide between our SETD1B-NDD cohort and unaffected controls. About a quarter of these DMRs contained no array probes, which suggests that array-based discovery has missed disease-relevant signal. The most stable DMR in the SETD1B-NDD signature was within a gene named PPM1D, which itself causes a distinct neurodevelopmental disorder (Jansen–de Vries syndrome). Surprisingly, this DMR is not covered by any array probe, meaning array-based approaches would not have used it to classify individuals with SETD1B-NDD.
We built a long-read-derived methylation classifier for SETD1B-NDD within a nested cross-validation framework. We were happy to see that our classifier correctly classified independent cases from our collaborators at the University of British Columbia and UMC Utrecht, even though the Utrecht samples were processed with different DNA extraction, library preparation, and basecalling methods.
We used the same approach to develop classifiers for Kabuki syndrome and Kleefstra syndrome, with promising preliminary results. These classifiers also transferred to data generated on the PacBio platform and to methylation-aware short-read sequencing using the Illumina 5-base technology. Unlike bisulfite, 5-base converts only methylated cytosines and does so without damaging the DNA, so sequence variants and methylation can be called from a single library.
We combined our three sequencing-derived classifiers with 34 array-derived signatures in a single open workflow called EpiSignaLR, which is available for any researcher to use. It takes an ONT modBAM file, a PacBio BAM file, or a 5-base methylation gVCF as input and returns a call for each signature. Because methylation is already captured in these genomes, EpiSignaLR adds episignature screening for low additional computational cost, with no additional sample or laboratory work necessary, so it can be run on existing genomes from unsolved cases or alongside a primary diagnostic analysis.
Together, these results suggest that episignatures can be discovered directly from sequencing data, capture signals that arrays miss, and carry over across sequencing platforms, but there is still work to do before they are ready for clinical use. The main limitation is cohort size. Each signature was built from roughly 10 cases and 10 controls, and calibration, particularly for Kabuki syndrome, will need larger, age-matched, platform-matched cohorts. Because these conditions are individually rare, no single center can assemble those cohorts alone. The individual-level data behind many published episignatures is also unavailable, so they can’t be independently evaluated or recalibrated for new platforms. For these reasons, we will be releasing all methylation data, signatures, and code from this study, and we plan to keep doing so. We believe the main barrier to clinical sequencing-based episignatures is now access to data, not the technology. If you care for individuals with these or other episignature-associated conditions that you think would like to participate in our research study, or have methylation-aware genomes from unsolved cases, we would love to hear from you. We are recruiting individuals with known diagnoses, as well as individuals with uncertain results in episignature-associated genes, and are able to send collection kits to families houses at no cost.
The full study is available as a preprint on medRxiv
