Hello everyone, I'd like to present a tutorial on next generation sequencing technologies. My name is Elaine Martis, and I'm at the Ohio State University and Nationwide Children's Hospital in Columbus, Ohio. This is my financial disclosure. When we talk about massively parallel or next generation sequencing, there are a variety of basic pieces of information that we need to understand in terms of how this technology works. First, I want to just introduce the notion that next generation sequencing has transformed biomedical inquiry. You can see a graph here from a review that I wrote in Nature in 2011 celebrating the 10th anniversary of the Human Genome Project that shows the output per instrument from sequencing instruments, and with a remarkable inflection around 2005, which marks the first introduction of next generation sequencing instrumentation. Ever since then, the trajectory or the slope of this line has gone ever upward to the point where we are today, where multiple human genomes can be sequenced literally overnight on the highest throughput instruments available commercially. Part of the reason for these differences is really illustrated by this simple diagram at the bottom of the slide. In panel A, one can see the historical approach to Sanger sequencing that was utilized that included library preparation and bacterial host, plating, picking of subclones, and preparation of DNA, followed by the sequencing reaction, and ultimately the separation of these fragments using electrophoresis and some form of detection, initially radioactivity, ultimately fluorescence. This is really in stark contrast to the overall methodology that's shown in panel B that's utilized for next generation sequencing, namely that we can start from simple genomic DNA, transform some sort of fragmentation, add on adapter-specific sequences that are unique to the sequencing platform, attach these to a surface and amplify them so that the signal from sequencing reactions can be detected by the sequencing instrumentation, and then go forward to produce the sequencing data in situ without the need for electrophoresis. Literally the advantage of next generation sequencing is that it couples together the molecular biology of sequencing with detection, whereas in Sanger-style sequencing we first do a molecular biology step followed by separation and detection. And so it is this combination in next generation sequencing of sequencing chemistry and detection that's taking place simultaneously on the instrument that leads us to call it massively parallel sequencing because of the magnitude of throughput and the simplicity of producing the data. There are a few basic aspects of next generation sequencing that I'd like to present. First of all, the library construction really combines either DNA or RNA fragments with these custom synthetic adapters that I mentioned, using either conventional ligation with enzymes or by transposon insertion, and we'll look carefully at both of these approaches. The resulting library fragments then are amplified on a solid support, and this support can be either a bead or a flat surface, both of which would be covered with covalently attached adapters that are complementary to the library adapters. So they provide a place where the fragments can be amplified selectively using this adapter-adapter hybridization as a starting point for enzymatic amplification. The sequencing reactions, as I've already stated, are coupling the nucleotide incorporation steps with detection, and this occurs typically in a stepwise fashion, allowing us to detect hundreds of millions to billions of sequencing reactions per instrument run, leading to the term massively parallel sequencing. One aspect of next generation sequencing that I'll talk about is the fact that these shorter read lengths are going to require specialized bioinformatic-based analyses, and this is one of the unique aspects of next generation sequencing that is a bit of a complication, but definitely necessary given the amount of information that we can utilize, produce rather, and utilize for biomedical inquiry. So let's talk first about the input DNA needed to make next generation sequencing libraries. In general, these libraries tend to be what are referred to as short insert libraries, where we're focusing on a fragment population that's typically in the range of 200 to 600 base pairs in terms of the input fragmented DNA to the library construction process. We can, of course, use a variety of starting materials for next generation sequencing libraries, and these can be derived from numerous different sources. I've just listed a few here that are commonly used, but there are a multitude of different places we can obtain DNA or RNA and input these to the next generation sequencing libraries. Typically in a clinical laboratory, we're focusing on sources that come from high molecular weight genomic DNA, such as cell lines, blood, fresh or frozen tissue, or buccal swabs. We can also, of course, input PCR products, and these could include products from what's called multiplex PCR, a concept that I'll cover later in the talk. In addition, the reality in the clinical lab is that we often have low molecular weight or degraded DNA, often that's present in very low quantities as well, and this could be from small formalin fixed paraffin embedded tissues, biopsies, for example, or slides from FFPE. It could be from forensic specimens, and in some extreme cases, it could be from very ancient specimens, such as Neanderthal bones or other fossilized DNA. So the first consideration really is, once you've identified the input DNA source, what's the method that's going to be used to isolate the DNA, and then how do I evaluate the quality or intactness and yield of the DNA that's resulted from my preparatory procedures? Let's take a look now at some basic considerations for DNA isolation. First we need to understand what we're working with. Is it tissues or cells? Is the source from plants, animals, or a microbial source? These all invoke differential methods that essentially open up the cell and release the DNA content, depending upon the nature of the cell membrane and any associated tissue-based molecules that need to be dealt with. In addition, we need to understand how the preservation of this tissue, for example, if that's the input, has been performed. So has the tissue been fresh frozen? Is it in some sort of a preservative that would need to be removed? Or is it formalin-fixed and paraffin-embedded? And here the good news is that regardless of the source and the preservation method, there are now numerous protocols as well as commercial kits that are available to isolate DNA for downstream NGS. Some other considerations include the need for scale. In particular, in certain laboratories, because of the throughput that's required or the number of samples per day, we're often needing to think about robotics or other types of instrumentation that allow us to scale from single samples up to multiple or hundreds of samples, depending upon the input need in that particular day. In addition, we may want to study both DNA and RNA from the same tissue that we're evaluating in the laboratory. And in this case, we would want to preserve tissue and isolate nucleic acids, both DNA and RNA, from the same tissue block, for example. Finally, as I've already mentioned, the use of formalin fixation and paraffin embedding can cause specific problems. In particular, we need to take into consideration here the age of the specimen because the older the specimen, the more degraded the nucleic acids will be. And there are also specific formats for formalin fix, paraffin-embedded tissues, such as entire blocks, small needle biopsies, curls from the block where the microtome has been used to shave off portions, and other considerations. And so all of these things need to be taken into account when designing a laboratory assay that's going to be utilizing next-generation sequencing as a readout. I mentioned the need to evaluate the input DNA quality and quantity. This is critically important before going forward in library construction. In particular, evaluating DNA quality or intactness can be done cheaply and easily using agarose gels with ethidium staining, and this is ideal for low-throughput applications. There are also commercially available precast gels that can be utilized, and these are very easy to use because you can simply open them up, put them on the device, load up the samples, and do the run. But they also are expensive, and so you have to find a balance there in terms of what the laboratory needs, turnaround time versus cost. And in addition, we also want to be able to consider scalability, as I've mentioned before. So if you have large sample numbers, running small precast gels or even agarose gels can be very cumbersome and time-consuming. We may want to go with a microfluidic approach, such as on the Agilent BioAnalyzer Tape Station, that allows a fairly simple input, a very small amount of DNA or RNA, and an evaluation that happens quickly and can be recorded, such as in a laboratory information management system. DNA quantity is also important. Many clinical labs utilize the NanoDrop platform, which is fine when dealing with high-yield nucleic acid extraction, but has some difficulties at the lower end of yield, where the sensitivity really is lacking. In this case, for single samples or small numbers of samples, we typically utilize the Qubit, which includes an intercalating dye and so is much more sensitive to low amounts of DNA. Or in a multiplex format, we may want to consider, for example, pico-green staining and a 96-well plate reader to allow for high-throughput quantitation, again, in a format that may be electronically incorporated into a LIMS database. If after the evaluation is completed, we have high molecular weight and good concentration, we then proceed directly to fragmentation, which I'll talk about next. The challenge is really in low molecular weight degraded DNA, where you may not even need a fragmentation approach, as well as low concentration, where you may need to perform some sort of amplification or concentration to reduce the volume and increase the concentration of DNA before going forward into the library construction approach. In either one of these cases, there will be some adjustments required to optimize the library construction. In the case of high molecular weight genomic DNA, we want to then perform some sort of fragmentation. This takes a variety of forms, flavors, and scalability. In the simplest sense, back before there were specialized devices available, one might be using a syringe and a needle with a specific gauge, coupled into some sort of high-pressure device that would permit shearing of the DNA as it passed from the syringe through the needle. There are also commercially available devices now that one can use, including the Hydroshear device, the Covaris Ultrasonicator, which also has a G-tube device that is specially adapted to form one fixed paraffin-embedded tissues. And there are many others that are available commercially to perform mechanical shearing. In a case of scalability, often these mechanical approaches are not terribly scalable. They also tend to use a lot or lose a lot of DNA in the process of the fragmentation, and so one needs typically a higher input to get the desired yield on the backside of fragmentation. In these cases, we often want to consider enzymatic methods for fragmenting DNA. These can be simple as restriction enzyme cleavage, which is fine but can be quite non-random, zinc treatment, which damages the DNA backbone and has to be carefully quantitated, or utilizing a proprietary nuclease cocktail, which may be available commercially, often referred to as fragmentase. And again, any of these approaches will require some sort of evaluation up front to really understand the optimums for time and amounts of input enzymes for the proper shearing of DNA to the size range that is optimal for library construction. This is just now an idealized look at adapter ligation-based library construction. So we're assuming here that we're starting with the DNA fragments. Again, these may be from shearing or from enzymatic treatment of high molecular weight DNA, or if starting from DNA fragments from other preparatory methods, including low-input, already degraded DNA that does not require an initial shearing approach, or PCR products, for example. Typically, we perform a series of steps that are enzymatic in nature to blunt the ends of these input fragments by either chewing back or filling in accordingly. And the processed fragments then have a phosphorylation step that adds phosphate groups to the end of the fragments, allowing the ligation of these specific adapters that I mentioned earlier, which are synthetic DNAs that are unique to the platform which sequencing will take place on. Once we have the adapter ligation step completed, there typically are some PCR amplification steps and cleanup steps that follow, including a final step to quantitate the library and really understand what the library concentration is prior to going forward into other methods. Other methods to follow can include simply going forward to sequence the entire library, which is considered whole genome sequencing, or in the case where the input DNA was from a whole genome, or one might proceed with the whole genome library to perform subsetting or specific isolation of components of the library, such as the exome, which is a word that indicates isolation of all coding genes, protein coding genes, or one may have specific genes that one is interested in, which is a panel or gene panel approach using hybrid capture. And I'll talk in a few minutes about the hybrid capture-based approach to subsetting the genome. I also mentioned that transposons can be used for library construction. This is a very straightforward approach that simply loads up the transposons with the synthetic adapters that one needs to add on to the ends of the resulting fragments and allows the transposons to interact with the genomic DNA and essentially tag the genome by creating breaks in the DNA and simultaneously adding in the beginning of these unique adapter sequences onto the resulting fragment ends. These then go through subsequent steps to add on the additional components of the adapters, a PCR amplification step, and then you have essentially sequencing-ready fragments. So the transposon-based approach is typically used in situations where there is low input DNA and allows one to forego several of the steps that were presented in the prior slide using the adapter ligation and to produce a library with relatively low amounts of DNA loss in the process just because the number of steps have been minimized. So this is a frequently used, highly automatable approach to producing NGS libraries. In the process of making libraries and in the surface-based amplification that I've talked about already, we're typically using enzymatic amplification, or PCR, of the library fragments in these different processes. And while PCR is a terrific vehicle for amplifying DNA, we do know that in the construction of NGS libraries and the amplification on surface, we can introduce some specific problems. In particular, PCR in the library construction approach can introduce preferential amplification of certain fragments. This can happen in particular if there is a wide distribution of fragment lengths where very short fragments will be preferentially amplified. And in addition, in low input DNA amounts, jackpotting occurs because of the lack of complexity in the fragment population. So we need to pursue approaches that remove the duplicate reads that result from this preferential amplification, and I'll talk about that a little bit later. The good news is that it can be largely done in bioinformatics space following the production of the NGS data set. However, we do want to be aware of this, and in cases where possible, limit the amount of PCR amplification in the library construction process as much as possible. In fact, there are now commercially available kits that use only one or two PCR amplification steps, and these are commonly referred to as PCR free library construction kits. PCR can also introduce false positive artifacts because enzymatic amplification can introduce substitution errors. The polymerases are not perfect. If this occurs early in the PCR cycle, the error may then propagate through the amplification of the fragment and would appear to be a true variant just based on its prevalence in the fragment population. I'll talk in a minute about how the use of unique molecular identifiers or UMIs has largely addressed this early substitution error problem. Finally, as we have, we'll look in the next few slides around the amplification on surface prior to the sequencing process. This bridge amplification or cluster formation can also introduce some biases, and typically this shows up in amplifying high and low GC fragments with lower avidity and efficiency than fragments with a more uniform distribution of ATGC. As a result, we can identify that coverage will be reduced at these loci, and we need to take this into account in evaluating variants by eye. I want to talk just briefly about multiplexing and the need for molecular barcoding before I describe how it's done. This has really been brought on largely by the increase in throughput on next generation sequencers, which allows us to pool together libraries or reaction products onto the sequencing flow cell to produce more data in a given run. However, we need to, prior to pooling, allow ourselves to have the ability to differentiate which libraries come from which set of fragments, or from which sample rather, and so I'll talk about how that's done using molecular barcoding. These barcoding schemes really permit the addition of specific sequences onto each library and enable equimolar pooling or multiplexing as it's called. The ability to specifically label the fragments then is dealt with post-run or post-sequencing by using a demultiplexing software that essentially identifies the index sequences or the molecular barcode sequences that have been placed onto each fragment and bins these according to the barcode combinations so that we recapitulate the original fragment set of reads from each individual sample and then we can proceed with the individual analyses for those samples. In addition, because next generation sequencing is now being used to provide very deep sequencing of libraries, for example, in the application referred to as liquid biopsy, or even in non-invasive prenatal testing, we have to use unique molecular identifiers to separate true mutations from sequencing noise during data analysis. So let's take a look at how that works. This is just a brief look at these unique molecular identifiers, or UMIs, added on to the adapter, the platform-specific adapters that are being used in this case for the Illumina sequencer. So on the upper panel, you can see the fragment that is readied for adapter ligation, and you can see the adapter in the Y adapter that's commonly used for the Illumina platform that contains this small addition on the P7 arm, which is the unique molecular identifier or index. So this, in particular, is added during the ligation step where we have a one-to-one relationship between the adapter and the fragment. Therefore, the fragment is labeled initially with the index on the adapter, as you can see on the bottom panel on both ends, and this UMI now uniquely labels that fragment as we go forward into the bridge amplification and sequencing reactions. And we'll come back to this later when we look at the stepwise approach to sequencing. This is really the initial introduction of this UMI. Now how do we use this as we come through the sequencing process? We really get a dual functionality out of the UMI-based approach. First, as you can see from these aligned reads onto the genome, we can perform deduplication, as I mentioned, to remove any PCR duplicates. And what you can see in this diagram is that the top five molecules all have a unique UMI that has been assigned to them, and so they are indeed from unique fragments in the population of the library. However, the molecules grouped together in group six below are all the same molecule. They all have the yellow UMI, and so here we have only a total of six unique DNA molecules, the top five, and only one of the duplicates that are grouped together from the same molecule. The rest of those reads would be discarded and not considered further in variant identification. The next step, then, is variant identification, and we want to use the UMIs shown here in a group of molecules that all came from the same read but have different start and stop points, so they're not duplicates. So this is post-duplication. We're now looking at the elimination of base-calling errors or polymerase errors, and we see an example of both in this diagram. The sequencing error is the little blue dot. It's occurring at the end of the read, where we commonly see the signal-to-noise sometimes compromising our ability to get a true base call, and this sequencing error would be removed and not considered further because of the fact that it's a one-off. It's only happening in one of the reads in this population and is representing essentially a polymerase error. In the lower example with the three red dots, we have different reads. They're all reporting the same error that's obviously happened very early in the process of making the library amplification, again, because they all contain the same UMI, and they are clearly indicating a false positive error because of the similarity of the error occurring in multiple fragments. So by this manner, we can do error correction and consensus read construction and come up with the true SNV, this green indicated change in the nucleotide sequence that would then go forward into the variant-calling process. I want to turn my attention now to some platform-specific comments. We'll start first with the Illumina sequencing instrument and its approach using sequencing by synthesis. In the Illumina process, as I've already alluded to, once we have the library fragments with the specific adapters and quantitated, we can then introduce them using the instrument fluidics onto the flow cell for amplification of individual fragments. And again, this individual fragment amplification is occurring so that we have sufficient signal for the instrument optics to be detecting the sequencing reactions as they occur. So as I mentioned earlier, quantitating the library fragments is really important. We don't want to overload the flow cell. We don't want to underload the flow cell. We want to have just the right concentration, and there are commercial kits that are available for high-precision quantitation of libraries that are also scalable. Once we have introduced these appropriately quantitated and concentrated library fragments onto the flow cell, they then are amplified in situ, and this uses covalently attached complementary adapters that I mentioned earlier, and you can see these in the diagram on the top. This is occurring through a bridge amplification step, which is a polymerase isothermal reaction that occurs to amplify the molecules on the surface of the flow cell. And as you can see, we get each cluster being derived ideally from a single library fragment if the quantitation of the library has been done appropriately. How does the sequencing chemistry then occur once we've got our amplified library fragments? Well, this is just sort of a close-up look of an idealized fragment. You can see it's been positioned on the surface. It's been primed on the right-hand side with a primer that is providing a free three-prime hydroxyl for the incorporation of the first nucleotide, and the polymerase has complexed with the two strands. That's the hatched purple oval, and so what we can then do is essentially add in the labeled nucleotides as can be seen on the left-hand side of this diagram. They're shown in different colors because, indeed, they are labeled with different fluorescent groups, and so we can determine the nucleotide identity that's been incorporated during excitation and emission because the emission will happen at a unique wavelength for that nucleotide. There are some specific other attributes about these nucleotides that are shown in these diagrams. First of all, you can see that they are fluorescently labeled, as I already stated, but importantly, on the three-prime end of the sugar moiety, there's a blocking group, which does not allow a subsequent incorporation once the nucleotide has been incorporated, and this is really the secret behind sequencing by synthesis because what we want to do is have a single nucleotide extension on the end of each three-prime hydroxyl for the primer sequence by the polymerase, and we then want to be able to detect the nucleotide that was incorporated at that particular step, and so this stepwise progress goes according to the words that are written here. We have nucleotide incorporation. We go through a detection step where we can identify the nucleotide that's been incorporated on each fragment population across the flow cell, and once those detection steps have been completed, we then go forward to chemically deblock the three-prime hydroxyls so that we can add in the next nucleotide in the next incorporation step, and we also want to cleave and wash away all of the fluorescent groups from the incorporated nucleotides because we don't need that fluorescence interfering with subsequent steps in the process, so at the bottom, you can see now the three-prime end that's available, the cleaved floor, and this is really the process by which this massively parallel sequencing takes place. A couple of points here because one could look at this diagram and think, well, this is great. I can just sequence forever and not have to worry about limiting my read length. Why are the read lengths limited by the sequencing instrument, and the truth is that a couple of things lead to increased signal-to-noise as we go through the sequencing reaction process. Firstly, in some cases, we may not always have a three-prime blocking group, so we could incorporate two rather than one nucleotide in a given step. This would lead to that particular fragment being so-called out of phase with the rest of the fragments, and indeed, as we go through the 150 or so steps that are happening in the normal Illumina sequencing process, we can see examples of some fragments that fall out of phase, and over time, this leads to an increased signal-to-noise. In addition, there may not be in a specific step a nucleotide incorporation that occurs, and so this, in addition, would also be a result of causing that fragment to become out of phase with the remaining fragments that have the same nucleotide sequence. Furthermore, the cleavage of the floor is never 100%, and so there's also some carryover of signal-to-noise that would occur, and so then that result is that over time, we increase this signal-to-noise and to the point, ultimately, where we can't get a clean read out of the latest incorporated nucleotide, and this is really what limits the overall sequencing reaction read length on the Illumina platform. So sequencing by synthesis, now, this is just showing a specific set of processes that happen writ large as we go through the sequencing by synthesis. Essentially, what we're doing here is we're illustrating that we're getting a stepwise incorporation and detection down the length of the fragment with the first priming step. We then go through a series of steps that read the different indices, these different UMI sequences that we already talked about that have been incorporated in the primers. Here you can see that we're sequencing first index one. The read product gets washed away. We then differentially prime and sequence the second index, and that information is included with read one, and then we go forward with the final read two amplification sequencing process using sequencing by synthesis as we did for all of the other reads. So now we have, for a given location on the flow cell, the combination of reads one and two and their corresponding indices that can go forward into the bioinformatic process of alignment and variant detection. One of the changes that's happened over time with many of the Illumina instruments is transitioning from random fragment cluster amplification across the surface of a flow cell to patterning of the flow cell, which is shown here. This is essentially nanomachining of small pits into the surface of the flow cell that provide the highest packing density of areas where fragments can be amplified, as well as assigned locations that don't require a multiple series of steps in the early parts of the sequencing process, as had been required in the past, to identify each of the fragment populations on the surface of the flow cell. So the net result of the pattern flow cell is that you get a higher data output because these are packed physically as closely as they possibly can be. You generate more sequencing reads as a result, and then the run times are actually faster because the need to locate fragment clusters is no longer required by the Illumina software. So this has really added a lot of facility to the different Illumina platforms. One other change that's occurred over time is differential labeling of the nucleotides for different platforms. So as you can see, depending upon the platform for Illumina, one either uses four differentially labeled nucleotides in the four-channel chemistry on the left, or two-channels chemistry where we just have a differential labeling of A, T, C, and G is unlabeled. So we just have two channels that need to be detected. Or on the iSeq platform, actually just a single chemistry is utilized with an intermediate chemistry step. And so this results in essentially faster data acquisition times for the two-channel and one-channel chemistry-based approaches because fewer detection steps are required to gather the information from a smaller number of floors being used to label the nucleotides that are incorporated in this stepwise fashion that I've talked about. So just to finish up with Illumina, this is just a representation, it's certainly not comprehensive, of the different platforms that are offered by this company. And what you can see is that it's essentially a range of different output from small desktop sequencers to the large high-throughput NovaSeq 6000 and HiSeq X10 that are used for population-scale sequencing or very high-throughput sequencing. Some platforms, including the MiSeq, also have longer read lengths available. The typical read length is about 2x150 base pairs, as you can see, but MiSeq can go up to 300 base pair paired-in reads. And the unique aspects of Illumina are that it is short-read sequencing, but high-accuracy reads are produced. Sequences derive from both ends of the library fragments, if desired. And there's a range, as I'm showing here, of capacity and throughput. Illumina is also offering cloud computing capabilities in their base space for bioinformatic analysis of the sequencing data once it's produced. I want to switch gears now to a second platform that's highly differentiated from the Illumina approach that we've talked about, which is a platform called Ion Torrent Sequencing, and this is sequencing by pH sensing. So let's take a look at how this chemistry works. Really, this takes advantage of the fact that in the incorporation of a nucleotide during polymerase addition, we release a hydrogen ion so that in a situation where you have multiple fragments of the same sequence being sequenced simultaneously, like we do in massively parallel approaches, multiple hydrogen ions would be released during the incorporation of a nucleotide. And so really, the way that this platform works is to produce a composite flow cell, which is seen on a side view here, idealized, obviously, where the upper level is a so-called sensing layer, where the molecular biology of sequencing is going to take place, and the lower layer is essentially a pH meter, a sensor plate that's meant to detect these changes in hydrogen ion concentration that result when a nucleotide is incorporated. This platform uses amplification of fragments on a bead. The bead, as I said earlier, is decorated with complementary adapters to the ones that are used in the Ion Torrent library construction process. This on-bead amplification takes place in a specialized process called emulsion PCR, where the carefully quantitated library fragments are emulsed together with beads, as well as polymerase and dNTPs, and essentially cycled together in these mini-reactors that are provided by the emulsion micelles to amplify each individual fragment on the surface of the beads. The beads are then reclaimed, placed onto the surface of the Ion Torrent flow cell, as shown here, and the sequencing reaction proceeds, essentially using differential flows of unlabeled native dNTPs. So the dNTPs are flowed in a specific sequence, C followed by T, followed by A, followed by G, for example, and if a nucleotide incorporation is taking place in each individual well that contains a unique set of the same fragment that's been amplified on the surface of that bead, then there will be a release of hydrogen ions, and that incorporation will be detected and recorded for that specific location on the flow cell, and so on. So what you get is essentially undefined read lengths, for example, if a C nucleotide flow occurs and there are four Gs in the sequence of that fragment population, you would get the release of multiple hydrogen ions, resulting in not just a single G being incorporated, C being incorporated, rather, but four Cs on each one, according to each one of the Gs that are present in sequence. So you can, so they're essentially the flow number is defined by the user, not the read length, and so you can get differential read lengths coming out of this sequencing process. One thing that you might be wondering is, in the case where I have four Gs in a row, how do I know that those are four Gs? And the basic answer is that the pH meter has a relatively large dynamic range, and so the larger amount of hydrogen ions that are produced from four Cs being incorporated against four Gs in sequence would be quantitated by the pH change in the instrument. However, the dynamic range is not an infinite, and we can see situations where multiple nucleotides in a row can give differential numbers in the case of the incorporation and sensing by the pH meter. In particular, we also need to be able to quantitate what a single nucleotide incorporation looks like in order to understand multiple nucleotide incorporations. And this is really done in the earliest stages of the sequencing on Ion Torrent, because there is a defined sequence of nucleotide flows, and for the Ion Torrent adapters, a defined single nucleotide incorporation in the first four flows. And so by incorporating, as you can see is illustrated here, in the initial nucleotide flows, a single C against this G, a single T against the A in the primer, a single A against the T, and a single G against the C, we can define what a single nucleotide incorporation looks like. And this provides a template or reference for all subsequent incorporations in terms of what the number of nucleotides added was. The platforms that are available from Ion Torrent are shown in this slide. As you can see, along the left-hand side, there are a variety of different Ion chips, and these produce differential numbers of reads. And there's also the Ion Chef hardware that's shown in the middle. The Ion Chef is really the device that allows one to automate the emulsion PCR steps that I mentioned earlier that are used to amplify the library fragments on the surfaces of the beads. And then one has either the Ion Torrent S5 instrument to produce data on, or the more recently introduced Ion Torrent GeneXus platform, which is more of a clinical sequencing platform, if you will. The Ion Torrent is essentially utilized for gene panels as well as for human exome sequencing. So it has a higher throughput, whereas the GeneXus would be more for gene panels. In particular, because of the way the sequencing occurs, which I just talked about, and because of the fact that individual nucleotides get introduced onto the Ion chip at a time, this approach has a very low substitution error rate. However, as I indicated, in runs of multiple nucleotides of the same identity, the production of insertion deletion appearances may be problematic, but can be largely overcome with coverage to get the accurate enumeration of homopolymer runs. This platform does not use paired-end reads. It uses a single primed sequencing that can, as I mentioned, have differential read lengths. It tends to be a relatively inexpensive fast turnaround platform for data production, and over time, the improved computational workflows for data analysis have emerged, including highly automated analysis workflows on the GeneXus platform. I want to turn my attention now to different aspects of how we subset the genome in next generation sequencing so that we don't have to sequence the entirety of the genome if that's not required. There are a number of different approaches to this. I alluded earlier to the use of multiplex PCR amplification. This is used, for example, in some of the Ion Torrent kits in the platform that I just talked about, where one essentially identifies genes of interest and produces PCR primers that will amplify across that gene of interest, producing overlapping fragments so that one can orient these fragments in alignment subsequent to the sequencing reaction and understand the variation that may lie there compared to the human reference genome. This requires the design of amplification primer pairs, a tiling approach to overlap the fragments, and then grouping primer pairs according to, as much as possible, shared GC content, TM for the priming reaction, and then specific reaction conditions. We then use these optimized pools of multiplex PCR primers to amplify these loci, pool together the products from each set once the amplification is completed, and go forward with the conventional library approach by ligating on these platform-specific adapters onto the ends of the PCR products prior to quantitation and sequencing. This is a facile approach that's been used clinically. There are some downsides in terms of being able to identify unique primers for some sequences in the human genome which are known to be problematic, but this can be a rapid, inexpensive way to produce gene panel results or hotspot panel results from clinical samples. Direct genomic selection also has emerged since almost the beginning of next-generation sequencing to subset the genome. These initial reports of direct genomic selection, including one from this group at Cold Spring Harbor in Nature Genetics in 2007, and another from the group at Baylor in 2008 in Nature Methods, initially utilized hybrid probes on the surface of an array to select out the fragments from the whole genome library, then elute those off differentially and proceed to sequencing. Of course, nowadays we don't use array-based probes. Rather, we do solution-phase hybrid capture, which is significantly more scalable, more facile, and occurs in a matter of hours as opposed to over a multitude of days as those early microarray approaches are required. This is, again, now an idealized figure on the right showing the target enrichment process where the input includes a whole genome library, as well as these synthetic probes that can be either DNA or RNA probes that have attached to them a biotin. These are the little blue circles on these probes shown in the figure. These probes or baits, as they're often referred to, are mixed together under idealized hybridization conditions with the whole genome library, allowing the probes with specific sequences for the genes or exons of interest to hybridize to the corresponding library fragments that contain those sequences from the whole genome library. We then allow these hybrids to be removed from solution using magnetic beads that have streptavidin on their surface and, of course, can complex with the biotin and selectively elute these out of solution using the application of a magnetic field to pull down the magnetic beads. The remaining unhybridized probes, as well as genomic fragments from the library, are then removed or washed away. We can then specifically elute off the captured fragments. These are quantitated, amplified, and proceed to sequencing, largely yielding just the genes or hot spots of interest that we have from our custom capture reagent. This is customizable, as I've mentioned, or, in the extreme, we can capture every coding exon in the genome, idealistically, using an exome capture reagent. Just a little terminology around hybrid capture. I already referred to these probes or baits as being the biotin-linked oligonucleotides, the synthetic DNAs or RNAs that are designed to capture specific genomic loci. The target or pond is really the fragmented whole genome library or input library that has been prepared for NGS sequencing and is available for capture. Capture is the process of this target enrichment by nucleic acid hybridization between the probes and the library fragments. We also incorporate in this process blocking reagents that include the addition of repetitive DNA for blocking these repeat sequences in the human or other mammalian genomes to avoid a spurious capture of those sequences. We also include library-specific adapter oligos to minimize off-target capture due to daisy-chaining of adapters. These are sort of additional details about how we achieve highly accurate capture of the fragments that we're interested in sequencing. Then another aspect of hybrid capture, of course, is evaluating the coverage that one obtains on the desired targets in the genome. So what are the breadth and depth requirements that we need to accurately call genomic variation at high specificity and sensitivity, which of course we want to be able to do. So this is again now just a picture depicting various aspects of coverage definitions for hybrid capture. We can see idealized at the bottom of this figure the target sequence that we want to capture and the areas just adjacent to it on either side at around 500 nucleotides. These areas of wingspan are defined largely by the fragment size population in the library. But in terms of definitions, on-target reads are the blue reads that are aligning specifically to the target or mostly to the target in some of the adjacent wingspan sequence. And in particular, once we go through the deduplication step, we can see that we produce in the darkest bars on-target unique sequences. Off-target sequences can also be produced. These can be due to the size of the library input fragments, as I mentioned, and these can also be unique sequences. They can go through a deduplication process and in some cases may provide additional evidence of variation. And then we inevitably get some fragments that come along for the ride. The off-target in the dark orange bars, as indicated, these can be higher or lower depending upon a number of things. One is the input amount of DNA from the library. The other is the amount of genome that we hope to capture with the targeted designed probes that we're using, where smaller amounts of target capture space lead typically to higher amounts of off-target sequence being captured. Then we can use a variety of approaches in bioinformatics to actually evaluate what we were able to obtain from the coverage of the areas that we were targeting. So these are just a couple of examples. The one on the left shows poorer quality coverage. This is evaluated post-alignment to the human genome and deduplication. And on the right-hand side, a more idealized and better quality yield in the dark blue areas of unique on-target sequences that were obtained from this better quality DNA sample that was used. And then at the bottom, you can just see how we break out different percentages. These are comparing a normal and a tumor from the same individual and really go in detail to identify unique on-target, duplicate on-target, and so on for the different fragment alignments that have been obtained post-exome capture. So that's depth evaluation. We also want to evaluate how broadly we've covered the areas that we're after. And this is now shown on the left, a threshold of 20-fold coverage showing for each probe the breadth of coverage at that 20-fold depth. And then on the right-hand side, showing 100-fold depth and the breadth of coverage that results. And again, these are two samples, a tumor and a normal, where the dark green is indicative of perfect coverage and the dark red is indicative of zero coverage. And so you can see these are quite high-quality results that are giving a large number of total probes with perfect coverage, even at a threshold of 100-fold coverage depth. We can also look at probe coverage and outlier analysis in aggregate. This is just showing a probe coverage distribution, including low-level outliers, of which there are just a tiny number on the lower right-hand side of this diagram, and very high-coverage probes on the upper left-hand side of this diagram. These basic quality metrics can provide just at-a-glance confirmation that the quality of the data produced from a hybrid capture approach is sufficient to go forward into variant identification, or conversely, can indicate that there's some problem where there may be more coverage required to be produced by additional sequencing, and so on and so forth. One last comment about hybrid capture is that the placement of probes where they're designed really matters. Here's an example from a paper that we published a couple of years ago, looking at breast cancer specimens captured through a very targeted approach on a small number of genes, where we were trying to compare the mutational spectrum of these breast cancer samples to the long-term outcome. This involved old-age FFPE blocks from different women with a breast cancer diagnosis. What we were able to identify in this particular assay was that there were a series of donor site mutations on the end of this second exon in this gene. CBFB is the name of the gene. In particular, these were highly prevalent in this set of breast cancers that we had identified, but had been completely missed in the breast cancer sequencing by TCGA, which was also done by our group. When we went back to evaluate TCGA samples for the prevalence of this donor site mutation, we were able to identify that it was there, but because of the probe placement in the exome capture reagent we used at the time, it had been almost completely missed because of the placement of that probe on exon two. This is just one example of where you can be sometimes missing things using a subsetting approach for the genome if the probes are not designed carefully. I want to just spend a little bit of time now talking about what happens after sequencing post-data generation analyses for NGS data. Firstly, we would not be able to produce these data, we would not have been able to transform biomedical research with NGS if it hadn't been for the availability of the human genome reference sequence. This is really our keystone for identifying variation in these short read sequences, because we first need to align these sequences to the human reference sequence in order to identify where the fragments came from and identify the variation that we may be interested in in the context of understanding causation for human diseases, for example, or in diagnostic mechanisms in the clinical labs. We first align reads, we go through a series of steps, which I'll talk about in just a minute in particular, and then we identify variants in comparison to the human reference genome. We subsequently would overlay the locations of genes onto these sequence variants that have been identified, and this allows us to interpret whether the variants that are identified are going to result in changes to the encoded proteins, and this is, of course, a critically important next step that is sometimes definitive, but in many cases the power of NGS is that we have very comprehensive capability to identify variants, the majority of which we may not actually understand in terms of not just the changes to the protein, but actually changes to the protein function. And so this is where this concept of variants of unknown significance emerges and sometimes is vexing in the context of trying to help diagnose a patient. So let's take a look sort of at a semi-silly representation of short realignment to the human genome, but literally this is the way that I think about this, and Susan will talk about this in much more detail in her comments, but really the human, these small fragments of the human genome, which are NGS reads, the puzzle pieces at the top, are really going to need to be put together into this crazy simplistic picture down below of the human genome representing where the tree is probably the exome, the one and a half percent of the genome that codes for protein coding genes, the blue sky is largely repetitive sequence, and then the green grass is the rest of the genome that's not repetitive but doesn't really have a coding function, if you will. So we have all of these different small puzzle pieces and we've got to figure out where they go. Some pieces obviously are going to be easier to place than others, and a lot of pieces look exactly like each other. And so we've come up with a variety of ways of dealing with this bioinformatically in terms of repeat content being masked largely, paired in reads obviously giving us some opportunity to perhaps have one end in a unique sequence and one end in a repeat, but still allowing us to map that read onto the genome and call variants from us. So let's talk just for a minute about the concepts of alignment to the human genome of short reads versus the ability to assemble them and where this brings us in NGS data analysis. So if you're sequencing small low complexity genomes, for example, viral isolates or bacterial isolates, these actually assemble reasonably well from whole genome data sets generated by next generation sequencing. Why? Because they're small and because they're low complexity, so very low repetitive content, single chromosome, for example. This happens quite readily and you can get highly contiguous assemblies. However, as we increase the complexity and repetitive elements, the size of the genome and chromosomal numbers, this really challenges our ability to take billions of short reads and reassemble them de novo to recapitulate entire human genomes and chromosomes. So in general, we have to resort to alignment to the reference genome, as I've already talked about, and this is our first step to identifying variants. There are approaches that are emerging now that are using what's called k-mer based assembly. These are being developed for de novo short read assembly and are enabling some novel content discovery because we're trying to use these approaches to assemble whole human chromosomes. So over time, this may emerge, but generally speaking, right now today, the alignment approach is favored for variant discovery using NGS short reads. Here's just an idealized workflow for variant calling. So I've already said that having this reference genome sequence is the key. We have to align reads using this approach where we take the FASTQ generated reads, as you can see from the diagram at the bottom of this slide, we map them or align them onto the reference genome to create what's called a BAM format file. This then is evaluated by position sorting and marking of duplicates. So these PCR duplicates that I referred to at the beginning of the talk get marked and they are no longer considered in the next step, which is variant calling. The output of the variant calling is the VCF or variant call format file, and this then provides the basis for comparing variants across different samples or looking at, for example, a primary and a recurrent tumor, etc. And I should also point out that RNA sequence data is aligned and analyzed in many of the same ways, especially the alignment step, the position sorting and duplicate marking, and then goes forward to other analytical pipelines that are specific to RNA-based analyses. One last comment, which is really, you know, we're used to in Sanger sequencing evaluating data by eye, and so how do we do that with next generation sequencing data? Well, the good news is we have the IGV, Integrated Genomics Viewer, excuse me, that can be utilized for this approach where we're able to now go in and evaluate chromosome by chromosome or in a specific locus how our next generation sequencing reads have aligned to the human genome. And in particular, this approach will highlight bases that do not match the reference sequence by color, and you can get down to the level of resolution shown here where you're even able to identify a single variant that's been identified and also identify low quality base calls that may look like variant but are of poor quality and probably false positives, etc. So IGV is used extensively throughout labs that are pursuing next generation sequencing. This just gives you an idea of how IGV might be used to visualize a tumor versus normal. Here we've identified a somatic single nucleotide variant in the lower panel of tumor specific reads that's entirely absent in the normal reads that are shown at the top in the IGV, and these in particular are in a gene polymerase epsilon 2 that's shown at the bottom with a corresponding protein sequence. This variant is occurring right near the end of an exon in this gene. Here's a look at an IGV view of a somatic insertion deletion event. This in particular is a deletion event of about 15 nucleotides, again highly unique to the tumor which is shown the reads from which are shown on the lower panel, and this is in a gene PDGFR beta. It's actually occurring in this juxtamembrane segment that's indicated in the pictorial representation of the PDGFR beta protein, and this is a well-known activating alteration that causes PDGFR beta receptor tyrosine kinase to activate its pathway. We can also evaluate, given the right probes into a hybrid capture or from whole genome sequencing, copy number alterations. These are now the human chromosomes shown in aggregate where the top panel is indicating log2 copy number changes that have been identified from read alignment and evaluation using GATK and VARSCAN2 to identify copy number altered regions. The variant allele frequency down below is reflective of areas of loss of heterozygosity, for example, similar to what we do from microarrays, and now this just shows some close-up arm level and focal alterations in copy number. In particular, on the left-hand side of the slide, chromosome 20 p arm deletion and an amplification near the center mirror on the q arm, and then on chromosome 17, an entire arm level loss of the p and part of the q arm with the genes that are involved being indicated in that central panel. Again, these are ways to evaluate the data by eye that we're generating from next-generation sequencing. So just to conclude this presentation, I want to say that widespread use of NGS platforms for biological research has really extended now to include their use in clinical laboratories, and this is really the reason for putting together this series of presentations. We have a variety of platforms to choose from and a variety of approaches to prepare DNA or RNA going into next-generation sequencing-based clinical assays, and they each have their strengths and weaknesses, so they require a very careful consideration prior to designing and implementing an assay in the clinical space. Because of the size and complexity of the resulting data sets, we really need to validate approaches to ensure the quality of the data and to design analytical pipelines that have the appropriate sensitivity, specificity, and reproducibility that are needed for clinical testing. Similarly, we have to help our variant analysts and our clinical directors to visualize data and really understand and characterize what is being identified by these assays to provide underlying data support and evaluation of variants going into a clinical sign-out report. So I want to thank you for your attention, and we'll turn next to Susan, who's going to talk more about bioinformatic analysis of these data. Forgot about that last slide.

next-generation sequencing

NGS

sequencing technologies

biomedical research

Sanger sequencing

library construction

amplification

sequencing reactions

Illumina sequencing

Ion Torrent sequencing

hybrid capture-based approaches