Vicente Yepez explains how RNA sequencing catches disease-causing variants that DNA tests miss. Two tools, Outrider and FRASER, found pathogenic variants in 16% of unsolved mitochondrial disease cases, including a synonymous variant that triggers exon skipping and a deep-intronic variant that creates a cryptic exon. The DROP pipeline automates the entire analysis.
Chapters
Hello, I am Vicente Yepes from the Technical University of Munich. Welcome to my talk where I will discuss how we have implemented RNA sequencing in a clinical rare disease diagnostic setting. So first of all, what are rare diseases? They are defined as life-threatening chronically debilitating conditions. In order to be called rare, they have to affect less than 1 in 2,000 people. There exist between 6,000 and 8,000 of them, and collectively they affect more than 350 million people worldwide. Most of them have unfortunately no treatment. They are predominantly caused by variants in a single gene, therefore the current approach is aimed at sequencing DNA to find the causal variants using, for example, whole exome sequencing. But this fails to pinpoint the pathogenic variants in two scenarios. First of all, the variants might not be captured, for example when variants are deep intronic or in large deletions. Whole genome sequencing can be a solution, but not all— not the consequence of all variants have not been characterized yet. For example, low impact variants such as UTRs or synonyms are usually not prioritized, and not all high impact variants have been associated to a disease already, such as variants such as stop, splice site, frameshift, or large deletions. And when they are novel, they require functional validation, which can be for provided, for example, using RNA sequencing. As such, RNA sequencing has established itself as a complementary tool for diagnostics since 2017, the pioneering works of Cummings and Kremer. Since then, other studies have focused using different diseases, focusing on different diseases using a variety of tissues, and their strategy is similar, which consists of detecting aberrant expression, aberrant splicing, monolithic expression, and even coding variants in RNA sequencing data. For us, we detect aberrant expression or expression outliers using a method, a statistical method and program called Outrider. First, gene expression data is known to correlate and covariate. Therefore, Outrider uses a denoising autoencoder to remove the unknown technical biases the common variation. Here we see two heatmaps, the left one of the raw counts and the second one at the right of the normalized counts, and we see how OTTRADER was indeed successfully able to remove the covariation. Afterwards, it fits a negative binomial distribution per gene, thus obtaining a— yielding a p-value per sample gene combination. Expression outliers are indicative of rare variants, and this has already been shown by other groups using control samples, for example GTEx, and in our case this also holds true, but especially for underexpression outliers, and for variants that are known to cause nonsense mediated decay, such as stop frameshift of splice. Now, aberrant splicing is a harder problem than aberrant expression. First of all, it can take multiple forms, such as exon skipping, exon truncation, exon elongation, exon creation, and intron retention. In order to detect them, We use two types of reads: these split and non-split reads spanning junctions. These are later converted into these metrics: percent splicing, or psi for short, and splicing efficiency, or theta. These metrics are fitted using a program and a statistical method called phraser, which also uses a Noising Autoencoder, and which yields a p-value per sample junction metric combination. In a similar way, but in a much lower proportion than for expression outliers, splicing outliers can also be associated with rare variants, as we see here. As expected, most of these variants are in the splice region, but also in intronic and other coding regions. This reveals that these regions, coding and intronic, are also important in the sense that they should also be inspected, as they can also cause aberrant splicing. So these two methods to detect our expression and splicing, they guide diagnostics. First of all, you identify outliers, then you go case-by-case inspection, adding genetic information and phenotypic information. And this, in our cohort of more than 300 mitochondrial disease patients, has led to the diagnostic of 33 of the unsolved ones, which represents a diagnostic rate of 16%. Moreover, we have 11 candidates where, where we are doing follow-up experiments and trying to detect variants that we haven't found yet. Our study is now available in medRxiv since a couple of months. I will now present two cases, the first of which is of a synonymous variant in the gene TWINK, Because it was synonymous, if it was detected by whole exome sequencing and it's a rare variant, but it was not prioritized. RNA sequencing reveals an exon truncation, clear also in this Affinity plot, and this eventually created a stop codon which caused also aberrant expression, as we can see also in this plot. The gene is associated with liver deficiency, which fits the phenotype perfectly, thus yielding to a solved case. A second case is in the gene NDOF5, which is a well-established mitochondrial disease gene in a patient suffering from mitochondrial disease. Whole exome sequencing revealed a start-loss variant heterozygous, but one and only one heterozygous variant is not enough to solve because this gene is known to have a recessive mode of inheritance. Therefore, a second variant in the other allele is needed. RNA sequencing revealed the second variant that was missed by whole exome sequencing because it was intronic, deep intronic, and it clearly shows how the variant activated a splice site and created a new exon in a weak splice site. Therefore, now we have the two variants in different alleles, thus leading to these cases being solved. I present only 2 cases, but as I said before, RNA sequencing helped us solving 33 cases in different genetic regions, for example in the UTRs, even in the hazard promoters, and that have been caused by large deletions missed by whole exome sequencing, in the splice region, direct splice sites, and even in the last exon where variants are suspected to escape nonsense-mediated decay. Not only we have diagnosed cases, but we built also a computational pipeline called DropTheDetectionOfRNAOutliers pipeline, which can be used by other users. So it takes as input raw sequencing data in the shape of BAM files and VCF files, plus some configuration, and it creates— it performs all the three modules of aberrant expression, aberrant splicing, and monolithic expression. It creates the necessary tables as input for each of them, plus quality control plots, and fits and generates the results table for each module. Moreover, it provides plots and guidance to analyze the individual results. After setup, the whole pipeline is executable with only one command, and it's publicly available in our GitHub page. We also address the problem of too few samples. So why is this? Because Altradar and Fraser, they require at least 50 samples in order to perform the statistical fit. So one idea is to combine samples from different centers. We tested if this idea is possible by combining our patients with controls from GTEx, and we saw that the majority of outliers were recovered both for expression and for splicing, as long as we use the same tissue, in this case skin fibroblasts. As such, we provide a resource of count matrices for blood and for fibroblasts which can directly be integrated with DROP. So the take-home of my talk is that RNA sequencing has increased our diagnostic rates by providing functional evidence for identified variants, and by identifying new candidates that were missed by whole exome or whole genome sequencing. Outrider and Fraser were developed as specialized methods to detect aberrant expression and aberrant splicing respectively. And we created an integrated workflow that uses them to detect these aberrant events from raw sequencing files, and that includes quality control and all intermediate steps. Finally, at least 550 samples are required. We provide a resource of counts that can be incorporated into DROP and into the analysis of labs just venturing into this. Finally, I would like to acknowledge my group in the Technical University of Munich, especially my PI Julian Gagneur, also the group of Holger Brockisch in the Helmholtz Zentrum Munich, and from there especially Mirjana Gusic, and all the patients, clinicians, and other people and families included in this study, and you for your attention. Thanks a lot.
