Welcome, everyone, to the Utility and Limitations of Diagnostic Transcriptome RNA Sequencing and Mendelian Disorders. Next please. This activity is intended for educational purposes only. Next please. This is the Faculty and Planning Committee Disclosures. Next. This session is approved for one hour of accredited continued education. ACMG is accredited by ACCME to provide CME. Next please. Please complete the evaluation form at the end of this session by visiting the ACMG Genetics Academy. Next. A recording of this webinar will be available in two weeks. Please submit your questions by using the Q&A feature. Use the chat only for indicating technical issues. I am honored to present you Dr. Pengfei Liu. He is an Associate Professor at the Department of Molecular and Human Genetics at Baylor College of Medicine. Next. And these are today's learning objectives. Thank you. Okay. Hello, everyone. I'm Pengfei Liu from Baylor. Thank you so much for the invitation to give this webinar. Today I'm very glad to share with you some of our experience in piloting using RNA-seq as a diagnostic tool to improve the molecular diagnosis of patients with Mendelian disorders. We have a long history in working very hard to try to get to the diagnosis for patients with genetic disorders. And there has been so much advancement in terms of technology over the past 10, 20 years. And we have also learned so much by studying the DNA, which is where the answers lie in, the molecular diagnosis are found in the DNA. When we classify variants that we find in the DNA, we have learned that we need to collect all kinds of evidence. One of them is RNA sequencing can be a potentially very useful high-throughput technology that provide a diagnostic laboratory real-time evidence to classify the variants when we are working on a patient's genetic diagnosis. So this is what I'm going to talk about today. First I want to review some basics about RNA biology. We know about the central dogma where the DNA gets transcribed into a primary RNA. And then several things occur following that, including the splicing, where introns are spliced out, leaving the coding axons. And then there are capping and polyadenylation of the RNA. And then eventually the RNA gets translated into a protein. So today we're going to talk about using RNA sequencing for the molecular diagnosis to facilitate DNA diagnosis. And in our current day and age, this is mostly built on the foundation of DNA diagnosis. So we often corroborate findings from RNA sequencing to DNA sequencing. Again, the primary goal is to facilitate interpretation of that DNA data using the RNA results. There have been quite a few papers over the past five, six years using RNA sequencing in the rare disease field to advance patient diagnosis. We can see that several different types of tissues are used and reaching diagnostic rate from 7.5% all the way to 36%, depending on the tissue type, disease type, and the patient cohort selection. And some of the studies were done with dependency on having the DNA sequencing data and some others also report it is possible to use RNA sequencing data agnostically to find things that are not originally reviewed by DNA sequencing. So first I'm going to give three vignettes to show, to illustrate the three common ways that we utilize RNA sequencing to facilitate molecular diagnosis of the rare disorders. The first one, using RNA sequencing to clarify pathogenicity of variants, often variants of unknown significance that are identified from DNA testing, such as whole genome sequencing or exome sequencing. So this is a 23-year-old male with intellectual disability, epilepsy, and lesion encephaly. Whole genome sequencing was done and de novo variant of unknown significance was identified in the gene DCX, which is associated with the lesion encephaly disorder that is matching to the phenotype. However, this variant, despite the fact that it is de novo hemizygous, it is very much deep intronic. This is 4.5 kb into the intron and there hasn't been previous report. Also predictions from computational programs such as SpliceAI gave a pretty low score on the probability of this variant causing abnormal splicing. So there's a lot of uncertainty here and we really needed additional evidence to help classify this variant, whether this is a VUS or there could be more evidence supporting this variant. So RNA-seq was done using the patient sample and now we see very clearly this is where the deep intronic variant is and we see reads to the left of that sequencing variant suggesting that there is indeed a new cryptic splice acceptor introduced caused by this deep intronic variant. This is consistent with prediction from SpliceAI and the final consequence is that a long fragment of intron is now spliced in to the coding region that is resulting in a frame shift, so loss of function. All of these are supporting the pathogenicity of this variant, helping the final diagnosis of this patient. Now we have the second case example illustrating another utility of RNA-seq where RNA-seq is unrevealing some of the missed diagnosis from the DNA analysis. So this is a six-week-old male with seizures, metabolic abnormalities consistent with the cobalamin pathway defect, triaxone sequencing was initially performed but it was negative. Then we attempted RNA sequencing using the patient fibroblast and we actually found a decreased expression of the MMAC gene as an outlier expression. So this is consistent with the clinical phenotype of the cobalamin pathway defect, however, we don't see any variant from exome sequencing on the MMACHC gene. Then whole genome sequencing was done. Actually there's another variant from another gene, a de novo hemizygous variant affecting the gene HCFC1. This is a pathogenic variant. We understand the reason why this was not identified in the exome is that the region happened to be poorly covered. So what is this HCFC1 and what is the downstream effect of exome expression in the pathway? This gene is actually a transcriptional co-regulator of the MMACHC gene. So HCFC1 sits right upstream of the MMACHC gene. Based on the literature, patients with mutation in the HCFC1 gene, they can have abnormal expression in the cobalamin pathway genes. This is consistent with finding from our RNA-seq data, this is data from our patient, showing many genes overlapping with the predicted pathway defect. So this is a good example where a variant may be difficult to find from DNA sequencing from the exome data, but now we can use RNA sequencing to give us clues, to guide us to where the molecular diagnosis may be. Now this is the third example where we use RNA sequencing data in a more complex way to not only find the diagnosis, in this case the diagnosis is already there, but we can learn more about the disease mechanism. This patient is a two-year-old female with severe global developmental delay epilepsy, and whole genome sequencing was done. We found two sets of diagnosis. The first one is a compound heterozygous pathogenic variant in the WWOX gene, and the second variant is the variant in question here, which is a GRIA4 de novo deep intronic variant. This is known to be associated, the gene GRIA4 is known to be associated with neurodevelopmental disorder, which might have overlap with the patient's clinical phenotype presentation. So now the question is, with this variant being a variant of unknown significance, is this really a secondary diagnosis, compound diagnosis, that is contributing together with the WWOX gene, or this may not be really contributing? What do we know about the gene GRIA4 regarding disease mechanism? GRIA4 has been shown in the literature to have autosomal dominant general function disease mechanism. This is slightly different from the other GRIA genes, where dominant and recessive general function and loss of function mechanisms are both available. But GRIA4, based on the current report, we only know of the general function mechanism. So we want to explore in the RNA-seq data, does that support the general function mechanism? First, before showing you the RNA-seq data, we have predictions, computational predictions from SPLICE-AI. It's predicted that there are new SPLICE donor acceptor sites resulting in this new cryptic axon that is supposed to be in-frame. So it is possible to cause some kind of function effect. However, RNA-sequencing data showed actually slightly different results. First, we see many different splicing consequences, and second, more importantly, we are seeing a reduced gene expression compared to the controls. So this is suggesting that the dependent variant created some cryptic splicing events that resulted in a frame shift consequence, leading to loss of function. So this is not consistent with the known disease mechanism, therefore is not providing this variant being a very strong candidate for a second diagnosis. So I've shown you the three use cases of RNA-seq. The first one, where we already have a VUS from genome sequencing, and we are doing the targeted analysis on the RNA-seq data. This is the most common use case currently, because this is the type of analysis that we know very well how to do. The other two types of analysis, the second one, where RNA-seq can find something that's missed by DNA, we are starting to learn more about it. And hopefully, we will have more and more examples in the future to know how to do this. But this is not the majority of the current use case. And the third one, regarding disease mechanism, this is going to be an emerging topic as we move more into disease therapy in the future. Now we will have more use cases potentially over there. So now I'm going to talk about some of our experience in applying moving RNA sequencing from a translational research test into a clinical test. We have actually just finished clinical validation of RNA sequencing in a cochlear environment under the project of UDN, the Undiagnosed Diseases Network. So we tried to validate RNA sequencing using two types of patient samples, cultured skin fibroblasts and whole blood. The experimental protocols and the bioinformatics pipelines are listed here with some of the QC metrics that we established from the validation work. And this manuscript has been recently put onto the preprint. And you can see more details if you search for this preprint in MedArchive. So before we start, we wanted to define what are the basic analytical elements that you analyze in the RNA-seq data. Just like we focus on analyzing variants, point mutations, single nucleotide or indel variants in the DNA data, in the exome genome data, in RNA-seq data, we don't focus on the variants, the point mutations or the indels, but rather we focus on two other basic elements. One is gene expression. The other is splicing pattern. What do each of these mean? I'm going to show you some simple examples in the next few slides. Gene expression level. This is basically zooming onto a gene and counting how many NGS release are there in the gene. And then you normalize that to have a number that you can compare between the patients. This is an example with a patient on the top showing this gene, that there are close to 3,000 sequencing reads mapped to this gene, compared to a control where there are about twice the number of reads, 6,000 reads, in the same gene. And both of these are sequenced at similar depths, so they're kind of normalized already. And the top is showing reduced gene expression. This is what we mean by gene expression level analysis. The splicing pattern example, here we have some other concepts, one of them called percent splice in. That is to indicate what other percent of reads at a given junction that are spliced to support the given splicing junction in your patient. So in this example, we had the patient data on the top showing the sashimi part of this gene. And we have a DNA variant here that is causing skipping of that exon highlighted here. So compared to the controls, we can see in the controls, if you focus on this junction, 100% of the reads that is going to the left are supporting this canonical splicing pattern. But in a patient, only about 50% of the reads are supporting this pattern, and the other 50%, they are showing skipped exon. So this is what we mean by splicing pattern. Again, the concept of PSI, percent splice in. We use numbers, metrics like this to evaluate every junction. So in our clinical validation, the first thing we do after we define the two basic elements of analysis is that we need a benchmark, we need a gold standard sample. It looks like we often use the genome in a bottle sample when we validate whole genome sequencing because there has been well-established gold standards using DNA sample from several HapMap samples provided by the gold standard results from genome in a bottle team. But we don't have a final product of that yet for RNA sequencing, even though we know that the genome in a bottle team have generated lots of useful data, and the team is working on finalizing the benchmark. So to complete our validation, we made a provisional one, again, focusing on the two basic elements, expression benchmark and splicing benchmark, using data released, raw data that are released by the genome in a bottle consortium, and we use them to compare against the sequencing results we have applied in our RNA validation so we can calculate sensitivity, specificity, just like what you do for whole genome sequencing validation. And as you can see, these numbers are satisfactory to us. We also calculated reproducibility, comparing repeats between different batches or within the same batch. Again, these are all showing convincing results. So next, I wanted to introduce an important concept here, which is called reference ranges. This is something that we don't talk about too much in DNA sequencing, but I want to make an analogy. When we evaluate DNA sequencing quality metrics, QC metrics, we often say that this set of WGS data is sequenced perhaps to 40X, and it can cover, let's say, 95% of the genome over 20X. This is always some kind of important metric to mention and to define, especially in a clinical assay. We needed something like this for RNA sequencing, for the transcriptome analysis as well, to tell you the limitation of the assay, what are the regions, what are the genes that we can confidently analyze, and what are the genes that we cannot, so that when you get a negative result, you know that perhaps there are many genes, especially genes of interest, that may not be interrogated well enough for this assay. So for this purpose, we are recommending, we established reference ranges for expression and splicing, again, the two basic elements. I'm going to give you some examples to illustrate what we mean by reference ranges, by expression and splicing. So now we are focusing one random gene, VTR1, and we wanted to see what is the reference range of expression of this gene among controls. So in order to do this, we first have to sequence a good number of controls, and then we count how many reads are there mapping to this gene in all of these controls and normalize them. So you can see that most of these patients, they have around 1,500 to 1,200 reads mapping to this LCTR gene. So we can define a statistical range, and we have a boundary that tells us above this or below this, we have statistical meaningful results to say that your patient may be an outlier compared to this. These are, again, the lower and the upper boundaries for this gene only. Next, we have another gene. This is a gene that is actually not well expressed in blood. That's why you see much lower numbers and a much wider, very wide range of distribution among controls. But we can still calculate a reference range, but this time the range is just very wide. We have two reference ranges that are wider in the distribution, and you continue to do this for many, many other genes. You do this for all the genes in the transcriptome. So you get a distribution of your data, and in the end, the data will look like this. So most of the genes with expression reference range, the lower end will form a curve and the upper end will form a curve like this. Similarly, we may do that for splicing. Well, sorry, this is the actual result we get for expression outlier. We can see our data using patient blood and fibroblast. We have a lower reference range. For most of the genes, they are at 0.83, meaning that we can detect 17% expression reduction for those genes. And for other genes, some of the genes we can detect very sensitive gene expression reductions. Some other genes, not so much. We can only detect a very traumatic expression reduction. Similarly, this side is for gene expression increment over expression, and the blood and the fibroblast data are shown here. That is gene expression reference range. Then we did similar analysis for splicing. This time, we calculated the statistical range for every junction. There are two example genes showing for each splice junction, we have a range. We can detect a reduction of close to maybe 15% splicing change and a much smaller number for the increment. And then we got all the numbers for all the genes in the transcriptome and made this plot, which is the splicing reportable range. So after defining the reference range, after defining the reference ranges for all the genes in the transcriptome, we started to challenge the validation with clinical samples with positive findings of abnormal RNA expression and abnormal splicing. And we did this within the UDN, and we were very fortunate to have a very diverse type of different abnormalities, such as axon skipping, cryptic axon, cryptic splice donor, cryptic splice acceptor, et cetera, et cetera. So we picked a good number of sample of each type, fibroblast and blood samples, representing different types of abnormalities, and challenged the pipeline. So here we used outlier analysis, as I alluded earlier, when we compare the patient sample versus a control, if the patient sample has a lower or higher expression, that may be an expression outlier, or if the patient sample is showing a different pattern of splicing that is outside of the reference range that we just defined for the gene, then that is also called an outlier for the splicing. So we did this. We analyzed the outliers for the expression and splicing for the positive samples, and we see that most of the findings have been identified. Some of these diagnostic findings were not captured within the respective type of RNA sequencing, possibly because the gene is not expressed very high in that tissue type. We tried to challenge everything and get all the samples we could, but it included some very difficult genes with low expression. Or at other times, there are genes where you may be looking for abnormal splicing, but that can be degraded by LMD. That makes it very challenging to find, or expression outlier that you expect to be resulting from a frame-shift variant, but sometimes the LMD efficiency is not high, so you are not detecting that very easily. So this is a summary of the metrics we got regarding the sensitivity to detect clinically positive findings. So I call this a targeted analysis mode, because we already know the diagnosis when we do this analysis, and we are just focusing on the gene of interest, the splice junction of interest, and try to determine if that is captured by the pipeline. However, as is the challenge with any MGS test, even though we have very high number of sensitivity here, at the same time, we have too many variants to review. And that often is not feasible for a clinical lab to review hundreds or thousands of variants, of outlier variants, per sample. So we have to come to a—we have to do some compromise and have a balance between sensitivity and specificity. In this case, we can increase the filtration stringency so that we have fewer variants to review per sample. Now, after that, we have dozens of candidates to review for every sample, but some of the most challenging diagnostic findings are missed during this, what we call the agnostic analysis mode. So what we propose here is that when actual clinical samples come for transcriptome analysis, the agnostic analysis mode is performed for every sample. In this case, we will find the—most of the diagnosis, but perhaps we'll miss the very challenging ones. And so you expect sensitivity, something like this, but we always recommend that the analysis to be done in conjunction with the DNA data. And if you have DNA data that already give you a clue to look at some of the genes or some of the variants very closely, you will get a sensitivity closer to this. So when we do the actual clinical interpretation, we recommend to do the combined analysis of agnostic analysis, looking at the whole transcriptome, and the targeted analysis, focusing on the BUSs that are prompted—suggested by the DNA sequencing data. And we also, in the manuscript, we summarized some of the known factors that may impact the variability of the sequencing result. And I have just shown you that we expect that RNA sequencing data to be variable for some of the genes, and that's why we need to define the reference ranges and have a confidence for every gene. So every effort should be made to control these variables, and we are only in the very early phase to learn what are the potential contributing factors to the variabilities. And I'm sure the field will learn more and more as we process more RNAse examples. So now I'm going to give a quick quiz regarding what I just mentioned for the RNAse validation experience. So I'm just going to walk through the question. I don't need to have interaction with the audience at the moment. I'm going to show you the rationale why I ask you this question. This is a pretty simple, straightforward question that I often give to fellow residents or graduate students when I teach about RNA sequencing. Which of the following analytical schemes is commonly used in the clinical interpretation of diagnostic transcriptome data? Abnormal expression outlier, abnormal splicing outlier, monolithic expression, all of the above. So the answer to this question is pretty simple. Just told you all of the above, all the three, abnormal expression, abnormal splicing, and monolithic expression, which we didn't talk too much about today, but that is very similar to the abnormal expression. All of the three, they are actually used, they're actually being used for analysis. But why do I ask this question here? What I really wanted to ask is the following philosophical version of the question. Again, it's the same options, but I'm asking you what is the most important component to you under different circumstances? So the first circumstance is that, imagine you're a physician treating the patient and you are trying to find the answer, what is the diagnosis? In this case, I think the abnormal gene expression outlier is sometimes, is often very important consideration because this gives you a very clear-cut endpoint of whether there is increased or reduced expression of the gene. And you can predict gain of function, loss of function in a very straightforward way. So in that circumstance, I see perhaps this outlier gene expression is the most important result you are trying to look for. What about this one, the abnormal splice junction? This is actually also important, especially during the analysis phase. The reason is that even though we often want to find the abnormal gene expression outliers, it is not very easy to do that because there can be substantial variability in gene expression and your patient may have a result that is very difficult to discern from the controls. But abnormal splicing is different. So I'm showing you some example data to illustrate this point. Here we are looking at gene expression distribution, where you see that the controls, they are scattered around here and you have a patient that may have a heterozygous loss of function, thereby reduced gene expression. But this is not too much lower compared to some of the lower end control samples. So sometimes it can be difficult to find this if there are additional variables in the experiment that may mask this result. However, for splicing outliers, based on our experience in the validation, you actually often see this. You may see some splice size where in controls, typically you don't expect much of splicing over there. But in your patient, there may be a cryptic splice site. You will have a very different splicing consequence. Or when most patients are showing canonical splicing and you have a mutation disrupting that splice site, your patient will be a clear outlier compared to the controls. Even though sometimes when NMD, non-mediated decay, is involved, we could see the effect being not so dramatic. But still, I think for most of the splice size in the transcriptome, they tend to be either very much close to 100% supporting the canonical splice site or there's no splicing if there hasn't been a cryptic splice site created. So the splicing data is sometimes an easier way for you to get to that result to identify the gene expression outlier. And similarly, monoallelic expression is used in a similar way. Sometimes the gene expression outlier is difficult to discern, but you may have a clearer data if you check the allelic expression. OK, finally, I'm going to spend a couple more minutes to talk about future directions beyond the current clinical practice of what can we do to make RNA sequencing better in providing diagnosis to patients with genetic disorders. So this has to do with the challenges that are intrinsic to any gene expression profiling or assays, which is for gene expression, it is always tissue specific. And in the genetic testing scenario, we have clinically accessible tissue, which is blood most of the time and fibroblasts, skin and fibroblasts sometimes, and muscle biopsy can be difficult, but there's not so much of other options beyond these three. However, the disease gene of interest may not be expressed or it may not be expressed well sufficiently in these clinically accessible tissues. So this is a key problem that we need to address to improve the clinical utility of RNA sequencing. And I'm summarizing some of the current efforts that have been shown to be effective in addressing this problem. The first one is CRISPR activation. This paper was published earlier this year in the American Journal of Human Genetics, where they focused on the two genes, MPZ1 and SPAST, related to neurological disorders. And these genes are neurological genes that are not well expressed in patient fibroblasts. But if you have patient fibroblasts, you can actually utilize the CRISPR activation mechanism, design, get RNA to target the promoter so that you can activate these genes in the patient-derived fibroblasts. And they have shown that before activation, this is the sequencing result if you just sequence the fibroblasts. There's very low expression, small number of reads, but after CRISPR activation, the gene MPZ1 can be much better detected, so allowing you better opportunity to reclassify a variant to evaluate BUSs in the gene. So that's CRISPR activation. What are the alternatives? One alternative way is very simple. We know that the gene is not well expressed in clinically accessible tissues. But it is expressed at some level. What if we sequence it very deep? In this case, we sequenced patient blood sample. I think blood is actually equally or perhaps more challenging sample to deal with than fibroblasts because many of the neurological disease genes, they are really lowly expressed or not expressed in blood. But if we sequence it to a higher depth compared to the conventional sequencing depth, this is 350 million reads compared to the most of the current papers that I show at the beginning of the talk. They sequence between 50 to 150 million reads. So this is deeper, we see better detection, and this is ultra deep. This is two billion reads, and we see pretty reasonable detection of these junctions. That's similar to what you see in CRISPR activation. Another way is to take the patient sample, patient clinically accessible sample, and perform differentiation. In this case, we made induced neuron in the patient sample, and you can see that by sequencing this induced neuron at regular depth, 100 million reads, we can detect pretty good gene expression of the MPC gene. Now, if you focus on this first axon, this is actually an important axon. You realize that this axon is only detected in CRISPR activation, but not in the fibroblast, and not in the deep RNA-seq using blood. This is actually because that first axon is from the transcript that is tissue specific, that is neuron specific. The transcript with that axon one is perhaps not expressed, or very, very lowly, extremely lowly expressed in blood in fibroblast. So there is some advantage in either using the induced neuron or CRISPR activation. Compared to just sequencing deeper, if the transcript of interest that you have is tissue specific, really, really challenging transcript. So speaking of induced neuron, this is work that we published earlier this year using patient fibroblast and convert them to induced neuron. So Dr. Shenglan Li led this work inside of the UDN. We piloted this on 71 families, and eventually got a 25% diagnostic rate with positive molecular findings. And five of these cases, they are in genes with low or no expression in fibroblast. So if we didn't do the induced neuron transdifferentiation, we wouldn't get any RNA data to classify these variants for that five cases. So the utility of induced neuron transdifferentiation is essential for that five cases. Very importantly, we designed the protocol so that it is applicable for clinical lab application. We noted that the traditional IPS approach might be too costly, or it might take too much time, and also it may take too much training. All of these are prohibiting these technologies to be implemented in the high-throughput diagnostic operation. So we opt to pick this transdifferentiation protocol that is involving a relatively low cost for the cell culture part, and a reasonable and shorter time compared to IPS, and also simpler training for the technologist that can be performed at a relatively larger scale. We also suggested QC experimental, QC metrics. At the beginning and at the end of the transdifferentiation and RNA sequencing analysis, to ensure that we are providing sufficient quality monitoring and quality check for the induced neuron transdifferentiation process. This work has, I have talked about it in a webinar from AJHG earlier this year. If you're interested, you can go there and find more details if you Google about this from the AJHG webinar. Now, this work is led by Dr. Sun Zhao regarding the deep, ultra-deep RNA sequencing that I told you earlier. So the idea is that if we can sequence the transcriptome at a much higher depth, perhaps we can detect more genes. We can make the RNA-seq data more useful so that it can have the potential to classify more genes. And indeed, this is the case. We sequenced several samples, including blood, fibroblast, IPS, and lymphoblastoid cell lines at various sequencing depths. From the 15 million read at the lower end at the current sequencing practices, all the way to over 1 billion read, and this is after UMI deduplication. And we see that you have more genes detected at a higher depth. So there is certainly a lot of benefit if we have the technology to give us sequencing data of higher depths. And we actually see the curve is pretty deep at the current sequencing depth of 50, 100 million reads, suggesting that if we go just a little bit higher, we may have a pretty dramatic benefit of detecting more genes. So this is the top part of the panel shows the number of increments for gene detection. And this bottom part shows tissue-specific gene expression. And you can see that for every sample type, tissue-specific genes can be better detected showing a darker red color for most of these categories. So this is something that's very interesting to us because we didn't know this before getting the data for the tissue-specific genes, we didn't know whether they are expressed at all if we are not using the matching tissue. But this data is showing us some very promising result that we will see benefit in the tissue-specific transcripts as well if we sequence deeper. This is a case example to illustrate the utility of ultra-deep RNA sequencing. Here we have a patient with a variant of unknown significance in the minus 17 locus of intronic region in the ITPR1 gene. And we wanted to perform RNA sequencing to classify the splicing significance, but this gene is not expressed very high in the fibroblast. So you can see at regular depth, we don't have enough reads to make a conclusion. So initially we made induced neuron, this obviously helped, and we can detect this abnormal junction and also abnormal intron retention thereby improving the precision of this diagnosis. Additionally, if you sequence the fibroblast just at a very high depth, you actually don't need to go through all the cell culture and transdifferentiation step, and you'll get similar results. So this is a good case to show the utility of both induced neuron and also ultra-deep RNA sequencing if you happen to have the gene to be expressed, that can be sufficiently detected at the ultra-high level. So in summary, I have shared some of our experience utilizing RNA-seq to help genetic diagnosis. I talked about our clinical validation experience and also some future perspectives and directions to improve the problem of tissue-specific gene expression technologies. I'd like to thank everyone who contributed to this project, people in my lab, people in our diagnostic lab, the MDML and the beta genetics diagnostic lab. I'd like to thank the funding agencies, the UDN and the Gregor Network, and the physicians who sent us the patient and the patients who participated in the study. Thank you all. I'd like to take questions. Thank you, Dr. Liu, for the informative innovation work. So we do have several questions here. More is technique-based. So one question is, since several splicing sites are generally a little bit leaky, would your test detect an increase in intronic inclusion at a junction where normally there are some rays going into intron anyway? In other words, do you look for an increase in intronic inclusion relative to controls when the splicing sites are leaky? Yes, that's a good question. We use the bioinformatics program called DROP for the outlier analysis, including the suppression and the splicing outlier. So this question falls into the category of the splicing part. That module is called FRASER. And in FRASER, we can indeed detect these type of abnormalities. It will calculate the percent splicing and also the percent of rays that are really into the introns. And we will have metric from the, established from the controls, and if the patient is statistically different from that, that will be flagged. So I think it was asking intron retention, right? Intron inclusion, yeah. Intron inclusion, yes. That can be detected even if the gene is leaky, right? If the gene is covered high enough, we should be able to detect differences even though the gene is leaky. Okay, thank you. To reduce the workload, and have you tried looking just at consensus genes found by both tools, like FRASER or TREADER? Yes, so that, because this is a clinical test, we try to be comprehensive. And we realized that these analytical tools are very often complementary. So meaning that we may find the diagnostic abnormality only in one of them. And actually, this is also predicted based on the biology. For example, if we have an abnormal splicing event that is expected to lead to a frame shift, and then thereby gene expression reduction, then these two events, the abnormal splicing and the gene expression reduction, they are kind of difficult to detect at the same time, right? If you have very high level NMD, you will not see the splicing easily, but you will see the gene expression reduction easily. Or the other way around, if the NMD is not very high, you will likely see the abnormal splicing, but not the abnormal expression. So there are many examples like this where we really benefit from a more comprehensive approach if you look at all the algorithms. So different samples may have varying sequencing yields, even within the same run. Do you normalize the yields or rates before performing any analysis? Yes, that is done before performing the outlier analysis. So we get the sequencing reads, and we do the alignment and generate the basic metrics. Then with those numbers, we start the normalization, which is actually part of the job pipeline. So they also ask, and the transcriptome is complex with variable expression dynamics. It's influenced by factors such as age and the tissue type. How do you select appropriate controls? This is a really key challenge regarding controls. We all know from our experience from DNA sequencing that control databases such as Nomad, Exact, when they came out, it was really, really helpful. That was critical in making our diagnostic tool more effective. We don't have that yet for RNA sequencing. And there have been many active discussion regarding building such a control database. And I think we're in the early phases. So now for a clinical test, we often want to compare data that are generated through the same pipeline, especially for a test like this with so much variability, that there's supposed to be more variability for RNA sequencing compared to DNA sequencing. So whether you will require generating control from data that are sequenced in your own lab, by your own sequencer in the clinical lab, or are you able to utilize publicly available control data? This is something that's worth future investigation. In our current RNA-seq validation effort, we sequenced a good number of controls using the same pipeline. So this is really challenging to do. First, you have to assemble the panel, and you also have to afford the effort and cost to generate all the control data. But we thought that was a good thing to do for the first attempt in the transcriptome clinical test. Thank you. Another question is, do you have any data on how many false positives you get with actual deep sequencing in genes that aren't expressed well in blood compared to the genes that are expressed well in blood? Could the actual deep sequencing just amplify the differences in splicing events that happen naturally, rather than being an indication of a splice event associated with a variance? Yeah, this is a very good question. We have this research effort going on right now. Hopefully, we will finalize the paper in a few months, and we'll see the results. Yeah, but this is a really interesting question to look at splicing phenomenon for the genes that are not usually interrogated in the regular depths. How many of them are artifacts? How many of them are actual splicing events that we don't call them artifacts, but they are not supportive of the canonical splicing? We are currently investigating this. I think time-wise, and the last question, which was very practical, any sense yet for what criteria could be used by a best-set clinician to identify good candidates on, say, a commercially obtained patient's axon for RNA sequencing? For example, if I have a patient in my clinic with a negative axon or genome for epilepsy or intellectual disability, but there was a ooze identified in a candidate gene, should I be thinking of RNA-seq, plus or minus transdifferentiation? And if so, how would I pursue that today? Or is this only through research or UDN right now? Yeah, so first, I wanted to say that this test is suitable for all patients without a clear diagnosis because we are able to do the agnostic analysis. We don't have to have a DNA candidate. We could just look at the RNA data and identify some outliers, as I showed. So this is the same answer, I guess, for any patient with any phenotype, as long as it is undiagnosed. But it will help if DNA data is there to facilitate the analysis in conjunction with RNA sequencing data. So we always recommend to do the DNA data because that's the primary sequencing technology. We always have to get a very high-quality DNA sequencing data, but RNA-seq is recommended for any patients without a clear diagnosis. There's indeed a challenge in how this is going to evolve into the clinical testing in terms of insurance payment. This is currently not covered by insurance. That's why we are piloting this within the UDN environment, but we are also planning to open this up to other, to general hospitals upon future discussion with the hospitals. I think that's the only way for us to generate data to convince the insurance companies and policy makers that there is good potential to use RNA sequencing to improve molecular diagnosis. Thank you. Hope that day comes soon. So this recording will be available at ACMG Genetic Academy in two weeks. So you can access that to review. The 2025 ACMG Annual Meeting, Registration and Abstract Submission will be open October 1st. Welcome everyone to participate. Thank you for everyone joining this education webinar. Thank you, Hong. Thank you everyone for your attention. Thank you everyone.

RNA sequencing

Mendelian disorders

genetic diagnosis

transcriptome

clinical practice

gene expression

CRISPR activation

bioinformatics

tissue specificity

case studies