- Knowledge center
The COVID-19 pandemic, caused by the coronavirus SARS-CoV-2, can only be effectively countered by a combination strategy of quarantining/social distancing and widespread PCR testing for the presence of viral RNA. The gold standard in Belgium is the use of a nasopharyngeal swab, taken by a healthcare professional, and transported to the test lab in a medium that typically keeps the virus alive. As an ISO17025 accredited lab for RT-qPCR services, Biogazelle had the unique skills set and expertise for RT-qPCR quantification of RNA to develop a robust, sensitive, and high-throughput test for SARS-CoV-2 in 10 days. Crucially important was the usage of reagents, consumables, and instruments that allow (semi-)automation and business continuity despite worldwide shortage.
In this webinar, co-organized and hosted by Tecan, Biogazelle co-founder and CSO Prof. Jo Vandesompele, outlined how we have set up and validated the workflow and how we are further improving it for cost-efficiency and performance.
I would like to introduce our speaker. Today's speaker, Jo Vandesompele, is a co-founder and Chief Scientific Officer at Biogazelle, a Contract Research Organization specializing in high-value genomic applications to support pharmaceutical research, clinical trials, and diagnostic test development. He's also Professor in functional cancer genomics and applied bioinformatics at Ghent University, Belgium. He obtained a Master of Science in Bioscience Engineering in 1997 and a PhD in Medical Genetics. He's the author of more than 250 scientific articles in international journals, including some pioneering publications in the domain of RNA quantification and non-coding RNA. I'm very happy to have Jo here, so enjoy with me Jo's webinar.
All right. Thanks very much for the kind introduction. Lovely to be here and thank you for the opportunity to present our story. I hope you can all see my screen.
To get everyone on the same page, COVID-19 is the name of the disease, an abbreviation for coronavirus disease 2019, and SARS-CoV-2 is the name of the virus, short for severe acute respiratory syndrome coronavirus 2. This is a single-stranded RNA virus with a genome size of about 30,000 nucleotides. As of last week (July 9, 2020), 10 million have gotten infected with the disease, and half a million have died. It is worrisome that the number of infections has doubled from last month. Awaiting a vaccine to prevent or a medicine to cure it, the key priority is to prevent its spreading. We can do this by physical distancing, washing our hands, and wearing a face mask when appropriate, combined with massive PCR testing to detect and isolate infected individuals.
Before we go into the details of PCR testing, a few words on who we are. Biogazelle is a contract research organization (a service provider), offering high-value applications to support research, clinical trials, and diagnostic test development for the pharmaceutical and biotech industry. We help accelerate the development of novel therapies through the discovery and validation of all kinds of RNA biomarkers, and the development of tools to assess efficacy, safety, and toxicity of drugs. We hold a unique forefront position in the application of quantitative PCR, digital PCR, and RNA sequencing.
On March 19, I received a call from the minister who invited Biogazelle to participate in the federal testing platform. In less than 2 weeks, on April 2, we had the first version of our platform to process 2,000 samples per day. Since then, we have scaled to 6,000 per day. Doing this in 2 weeks was a formidable task and only possible with the hard work, literally day and night, seven days a week, by a dedicated team. And of course, also only possible because we had years of internationally recognized expertise in PCR test development and clinical trials.
Here's an overview of our relevant qPCR expertise. Biogazelle is ISO17025 accredited for qPCR test development and use in clinical trials. The company was founded on a revolutionary method for qPCR normalization (geNorm) and data-analysis (qbase+), with more than 20,000 citations and thousands of customers worldwide. The company founders co-authored the MIQE guidelines for design, execution, analysis, and reporting of qPCR studies, again with more than 10,000 citations thus far. We have wet-lab validated more than 100,000 qPCR assays, beyond industry standards; the SARS-CoV-2 assay is just one of them. And since April, we are also part of the CSWG, an initiative from JIMB (joint initiative for metrology in biology) from Stanford University, dealing with the development and provision of access to standards, control materials, inter-lab comparisons, and knowledge to perform accurate SARS-CoV-2 tests. The ultimate goal is to build a "COVID-19 Diagnostic Standards Development Partnership".
The SARS-CoV-2 RT-qPCR test consists of 4 main steps, here indicated in pink and executed by Biogazelle. It involves the transfer of the primary sample tube to a 96-well plate to increase throughput downstream, followed by viral RNA purification, RT-qPCR detection of the virus, and data-analysis and authorization.
Importantly, the entire workflow is more complex and is, in fact, a multiparty workflow, involving health care professionals, logistic partners to make test kits (swab and tube), distribute and collect, bring it to the lab, and IT partners for sample registration and reporting of result to the doctor. During the first six weeks of testing, primary samples were preprocessed at four different external sites, with a team of more than 100 trained people. As of mid-May, Biogazelle is doing this mainly by itself, amongst others through automation of the primary sample reformatting part.
Of note, Biogazelle is just 1 of the five federal test labs; GSK, Janssens, UCB, and University of Liège also contributed. Finally, the platform operates under the auspices of the federal institute for public health Sciensano, the federal agency for medicines and health products (FAMHP), and the national reference laboratory for SARS-CoV-2 testing at KU Leuven. Every is coordinated by the taskforce, supported by Deloitte.
To visualize the complex workflow, the VIB made a nice 3 minute video. Let's watch.
The four other labs in the taskforce use a commercial workflow based on magnetic bead viral RNA purification and a CE-IVD approved multiplex RT-qPCR assay. At Biogazelle, given our expertise in PCR test development and its use in clinical trials, we were tasked to setup an alternative platform to mitigate potential risks of supply chain issues. The guiding principles were the following:
First of all, we wanted our platform to be flexible and modular, such that different components from different suppliers could work together, each optimized for their specific part of the workflow.
The platform had to be high throughput with the aim of at least 6,000 per day, and smart scalable. This means that we introduce more units of a given module, or more instruments in a given module if needed.
In terms of business continuity, we have established a strategic stock, standing orders, drop shipments and validated alternative suppliers. So we really needed a dedicated procurement officer to manage all that. For instance, for the RT-qPCR mix and the RNA extraction, we have two validated suppliers to mitigate supply chain issues.
We also wanted our platform to be autonomous, so no A to Z commercial solution of which most of them are relatively slow and/or generally face worldwide shortages for instruments and consumables.
Finally, our platform had to be high quality by default. Obviously, it's a diagnostic test, but built for constant innovation and improvements through internal cross-validation.
Our platform consists of 16 instruments to process 6,000 samples per day. For further scaling, we don't need to linearly increase the count of each instrument. For instance, our Tecan EVO100 qPCR setup robot can handle many more plates. For qPCR, we have not used the stacking module of our CFX384 qPCR instruments but validated it for stacking of a few plates (for DNA, we stack many more). Extraction scales per 2000 samples per centrifuge. These are examples of what we call modular, flexible, and smart-scaling.
The most challenging step in the entire process is the transfer of the individual patient tube into a 96-well plate for downstream high-throughput processing in standard SBS format (8 rows, 12 columns). Not only because this is a very time consuming process, but also because of safety concerns. The inside and outside of the tube may be contaminated with active virus, posing a biosafety risk for the operators. Therefore, everything is processed by skilled operators in a biosafety cabinet class II.
During the first six weeks of testing, we did everything manually and trained several teams. One team of 3 operators could process 500 samples in a shift of 4 hours. Processing means the transfer of an aliquot of swab transport medium to a deep-well plate, followed by the addition of lysis buffer, the first step of RNA extraction, which inactivates the virus and makes it safe to continue with the 96-well plate outside the cabinet.
To process 3000 samples during an 8 hour working day, 18 operators are needed and 9 biosafety cabinet class II. This procedure is laborious, slow, error-prone, and needs many operators. Because of highly focused work, shifts were limited to 4 hours. For all these reasons, we have introduced automation to replace the manual sample reformatting steps.
However, automation is only a part of the solution as it comes with its own challenge, i.e. placing it in a BSL3 lab or enclosing it in a biosafety cabinet. At Biogazelle, we don't have such a lab, and a cabinet enclosure is expensive with lead-times of 6-8 weeks. We therefore looked for an alternative solution, and that is to inactivate the virus prior to handling on our robot. Heat inactivation is an often applied method, but the efficacy depends on the viral transport medium. A wide range of temperatures and durations are recommended in practice, but often without validation that the SARS-CoV-2 virus is truly inactivated. Some studies even suggest not to do it, or to adjust the settings in each lab.
We introduced an alternative solution, i.e. collect the swab in a virus inactivating and RNA preserving solution. This solution not only makes it safe to work with the transfer fluid, but it also stabilizes the RNA, resulting in higher detection sensitivity, tested for up to 5 days. We successfully tested several options and ultimately settled on DNA/RNA Shield from Zymo Research. We add 0.002% of methylene blue to aid visualization. For molecular testing, the virus does not need to be kept alive. I find it thus very strange that virtually all commercial swab solutions attempt to do just that.
Considering our good experience with the Tecan EVO series of liquid handlers, we acquired an EVO200 for primary sample transfer and lysate preparation. This 2 m long robotic system has a barcode scanner to scan the patient tubes and the destination plate, and an 8 channel liquid handling module with liquid sensing to aspirate from the patient tube and dispense into a 96-deep-well plate, that has been pre-filled with lysis buffer using the 96-multichannel arm (MCA). This robot can process 3,000 tubes in an 8 hour shift by 2 operators. This is a reduction by 9-fold compared to the manual procedure. Introduction and validation were pretty straightforward; we had experience with the software, executed several dry runs, did so-called water runs, and then defined liquid classes using various concentrations of glycerol to mimic viscous swab fluids.
To select a suitable RNA extraction kit, we evaluated cost, easy of use, throughput, guaranteed availability and of course, performance. We went to great lengths in comparing 5 different RNA extraction methods on a standardized set of viral stock dilutions and positive and negative patients samples. Patient samples and viral stock were collected in 4 different transport media that are routinely used, including liquid Amies (eSwab buffer) and phosphate buffer saline. The results were quite shocking, with different RNA extraction methods showing very different performance; the transport buffers also differed greatly, with clear interactions between kit and buffer. On the left, I show you the serial dilution series of the inactivated viral stock dilutions from 10-3 to 10-5 for 2 RNA kits and 4 buffers. Cq values in the Y-axis, log10 of the dilution factor in the X-axis. The linearity and slopes are all excellent, but the intercept values differ greatly. The best and worst kit/buffer combination differs more than 5 PCR cycles, equivalent to a 30-fold difference in RNA detection sensitivity.
For RNA extraction, we finally settled on filter plates from Norgen Biotek or Zymo Research, processed in a centrifuge. We use a centrifuge with 4 positions for 96-deep-well plates. The epMotion dispensor is used to transfer the lysate to the filter plate, and a Viafill is used for quick plate dispensing of washing and elution buffers. The Mantis is used to add 4 µl of spike-in control and carrier RNA. Carrier RNA is essential to improve extraction efficiency, especially of low concentrated samples.
One operator can do 374 samples in 90 minutes. Three operators can thus process 6,000 samples in an 8-hour working day. Of note, other automated or semi-automated solutions either require much higher investment and/or do not reach the same throughput.
For qPCR setup in a 384-well plate, we use a Tecan EVO100 robot with a 96 pipetting head (or MCA). 14 µl of mastermix is dispensed, followed by 6 µl RNA sample. This process takes about 7 minutes, but with all preparations, barcode scanning and entry into our LIMS system, and plate sealing, we count on 20 minutes for 384 samples.
Of note, in other projects, we do 5 µl qPCR reactions, with the same robot (3 + 2 µl), but to maximize detection sensitivity, we opted for 20 µl reaction volumes. As indicated before, for business continuity reasons, we have two validated mixes, namely Bio-Rad’s iTaq one-step RT-qPCR for probes and Takara’s PrimeScript III one-step RT-qPCR.
The actual qPCR is done using 6 CFX384 qPCR instruments from Bio-Rad. This is a bit overkill, as 4 would be sufficient. We don’t even use the stacking module, but have tested it, and for a one-step RT-qPCR, it seems we can stack one or 2 plates at room temperature without loss in sensitivity.
Data analysis is a huge challenge. You can imagine if you have to look at 6,000 curves on a single day and interpret them for diagnostic accuracy, you need support. So we use the FastFinder software from UgenTec that allows automated data transfer from the instrument to the cloud, uniform interpretation with a little bit of artificial intelligence, and importantly, it has numerous checks. It looks at the positive and negative controls. It looks at the values of the internal control, does trend analysis and neighborhood analysis to check for possible cross-contamination. The data analysis occurs at two levels. There is the interpretation level and the authorization level, and it allows coupling with our LIMS system and that of the clinical laboratory or hospital.
Our platform is approved by the federal institute for public Health Sciensano, the federal agency for medicines and health products (FAMHP) and the national reference laboratory. We also operate in a ISO17025 accredited lab, and accreditation for the test is pending. We participated in a European quality assessment scheme and process daily blind proficiency samples at a rate of 1/300 patient samples. Every morning, we get QC results from the national reference lab. To safeguard our quality, we have multiple controls in each experiment, namely an internal spike-in RNA that controls for RNA extraction and RT-qPCR of each sample, and 1 positive and negative workflow control per batch of 93 samples. Finally, we have introduced digital PCR as an orthogonal validation method and to calibrate our platform.
In the next few slides, I want to demonstrate our platform's quantitative performance and some further validations that we performed. Here you see the results of a so-called plate homogeneity test, a functional validation of both the liquid handling robot and the qPCR instrument. We dispense 20 µl of the same PCR mix with template and evaluate homogeneity of signal across the plate. On the right side, you see extremely tight clustering of the amplification curves; and in the plate view on the left, you see a heatmap of the resulting Cq values. Note the scale between blue and red is only 0.1 cycle. There is no apparent spatial pattern. Please read our blog if you want to know more about this functional test, or if you wish to use our free web application to analyze your homogeneity data.
Another QC test is an RNA source plate volume test. RNA is eluted using 30 µl nuclease-free water, but recovery is somewhat variable. In this test, we want to evaluate the robustness of the Tecan EVO pipetting robot to handle different input volumes. We did 24 replicates of 30, 25, 20, and 15 µl input volume (of the same concentration) and did repeated aspiration of 6 µl. Please note the small Cq value range of the results, here from 22 to 23.4; the means are very similar and the standard deviations as well, independent of the input volume or repeat number of the aspiration. We conclude that with 13 µl in a 96-well plate, we can still accurate aspirate 6 µl.
And here’s the result from the recent EQA study in which we participated. All samples were correctly called, and we demonstrated excellent quantitative concordance, as shown on the left. Over a clinically relevant range, we deviate less than 50% of the expected concentration. Of note, from the almost 500 labs that participated, 15% could not detect SARS-CoV-2 RNA in the low positive sample with 200 copies/ml, equivalent to 20 copies into the RNA extraction or four copies into the qPCR. Also, almost 1% of the labs reported false positives in the no-virus control and 2-3% had false-positive results when testing other coronaviruses.
And finally, I demonstrate here the limit of detection of our platform. LOD is defined as the lowest quantity at which >95% of replicates are detected. We dispensed 32 replicates of 6, 3, 1 and 0 cDNA copies (digital PCR calibrated copies). First, we did not detect any signal in the negative controls, and detected all replicates with 6 copies per reaction. When using 3 copies, we lost signal in 5 wells. This means that our limit of detection is between 3 and 6 copies per reaction. Of note, when we calculated copies based on Poisson statistics of positives and negatives, the result and 95% confidence interval is very much in line with the digital PCR calibrated expected input.
At this moment, we have tested more than 180,000 patients. You see a binned histogram of more than 4,000 Cq values of positive cases, from around 130,000 tests conducted by the end of May. There is a monotonous increase from Cq value 10 to 33; and sharp decline at 37, aligning with our single molecule threshold. There is a striking difference in observed viral loads from Cq 7 to 37, equivalent to a billion-fold difference.
You can imagine that having very high titer patient samples in the extraction workflow may induce a risk of cross-contamination. So, we designed a method to detect possible cross-contamination by a very high positive sample adjacent to a negative sample. This so-called neighborhood analysis is built into our data analysis software to automate the detection of such spurious events. If cross-contamination would occur in practice, wells with more neighboring wells will have a higher chance of being cross-contaminated. Corner wells have 3 neighbors, edge wells have 5 neighbors, and regular (pink) wells have 9 neighbors. Considering this, we have weak statistical evidence that cross-contamination occurs, both in high-through and low-throughput labs.
If we have a suspicion of cross-contamination, which occurs luckily rarely, we queue such a sample automatically for repeat extraction.
To finish my webinar, I want to show what our future platform will look like. So far, I explained our platform 1.0. In 2.0 we will have - because of the modular approach and constant innovation - improvements built-in.
Despite its slow mutation rate, SARS-CoV-2 continues to accumulate mutations as it persists in the human population. It is important to evaluate PCR-based diagnostic assays as new SARS-CoV-2 genome sequences become available. Based on the COVID-19 Genome Analytics in Edge Bioinformatics from Los Alamos National Laboratory (USA), 99.5% of more than 50,000 SARS-CoV-2 genomes are detectable with the E gene assay, and 98.03 withe the N gene. Using the 2 assays at the same time does not only slightly increase analytical detection sensitivity, but also strain coverage to 99.99%.
Secondly, the Belgian institute for public health Sciensano estimates that, early next year, Belgium will face 40,000 daily consults of flu-like symptoms for a period of 100 days. Most of these will –hopefully- be free of SARS-CoV-2, but when doing a test to rule out COVID-19, it would prove valuable to determine if the patient is infected by influenza or respiratory syncytial virus. We are therefore developing a 7-target 4-color multiplex assay to co-detect SARS-CoV-2, influenza A and B, and RSV A and B.
Finally, a few words on sample pooling. You may have seen press releases, this one is in The New York Times, this one is in The Wall Street Journal. And even in Belgium last week, there was one in Le Soir where sample pooling is heralded as a miraculous solution to increase throughput and decrease costs. The idea is to combine multiple samples into a single extraction and qPCR. If the result of the pool is negative, it is assumed that all samples in the pool are negative. Of note, pooling is not difficult from an analytical perspective and it may indeed save cost and throughput, but thus far, no systematic and large scale simulation study has been performed using real life quantitative data to assess the impact on detection sensitivity. As you can imagine, by pooling, you dilute the target signal. And as such, low positive signals may be missed resulting in false negatives.
We used anonymous data from 1632 positive cases to simulate and compare 1D and 2D pooling strategies. This is work done by Jasper Verwilt, a doctoral fellow at Ghent University. It's clear that sensitivity is a function of prevalence and different pooling strategies. Each dot is the result of a simulation of 100,000 patients sampled from real quantitative data, whereby the color reflects the prevalence from 10%, which is high prevalence, down to 0.01%. On the Y-axis, you see the detection sensitivity in function of the various pooling strategies. You see that none of them reached 100%, of course, because you dilute the signal. In the preprint, Jasper shows that the choice of pooling method and pool size is an intricate decision with a prevalence-dependent efficiency/sensitivity trade-off. We believe that small 1D pools are a compromise between efficiency gain and loss of sensitivity, and may be useful for surveillance screening under low prevalence conditions.
With that, I want to come to my conclusions. We have set up a high throughput modular SARS-CoV-2 RT-qPCR detection platform. Primary patient sample handling is the major bottleneck, followed by RNA purification. The analytical measurements, meaning the qPCR, scale best. In tube virus inactivation is a key contributor to automation, and we have built customized solutions that are scalable, low cost, and ensure business continuity. High throughput and quality are thus not mutually exclusive.
To finish, I want to acknowledge several people namely the Belgian Task Force, the Reference Laboratory, especially Lies Laenen and Els Dequecker. Several companies have contributed with instruments such as Formulatrix, BASF Innovation Center Gent, Bio-Rad, Ghent University, Inbiose, CIMIT, Janssen, and finally the entire COVID-19 team at Biogazelle and especially the people here in alphabetical order, Jan Hellemans, Tom Maes, Pieter Mestdagh, Nele Nijs, Sophia Schollaert, Gaëlle Van Severen and Pieter Wytynck. With that, thank you for your attention and I'm happy to take questions.
Thank you very much for this very interesting presentation Jo. There are some questions in the question box so far, but you're welcome to write questions.
I will start with the first question and that's is pooling recommended on cases of higher positive populations.
Yes. It really depends. There are pooling options that are recommended for a high prevalence. Please hold on a few more days until we publish the manuscript as a pre-print online, but it really depends on the strategy and the prevalence. And for each, there is a solution which works always with a little bit loss of sensitivity. But I think it's not too difficult to get to 95%.
Okay. The next question is, are there any criteria that are being used to decide which samples get pooled, are high-risk samples? Let's say a retirement home that has confirmed case already?
For sure. I think the FDA has recently come with recommendations that any pooling strategy should have at least 95% sensitivity to be used in diagnostics. It is challenging, but it can be met. I think they recommend to pool only five. I think you can do a better pooling experiments by pooling more samples, increasing throughput, tremendously and reducing costs in so-called surveillance experiments, where you monitor the pandemic across populations or collectivities. And there, it's not unimaginable to pool eight to 12 samples or even more, but you have a little bit loss of sensitivity. But I think it brings me to the question, when is lower sensitivity compensated by detecting more cases. For instance, if you work under low prevalence conditions of let's say 0.1% as shown in my slide, if you would do 10,000 tests with a perfect assay, no pooling, 100% sensitivity, you would detect 10 cases. Suppose you apply pooling strategies which are lower in sensitivity, let's say 95%. If you could scale up to doing 100,000 tests, sacrificing a little bit on the sensitivity, you could still detect 85 cases, which is eight times more. So I think there is a place for pooling, probably not as a diagnostic tool, but really more for surveillance.
Okay. I do have another question here. Cost per test, which is a typical question.
Very difficult to comment on that. I think we can discuss that here for our platform, but it's not really relevant. The main driver is the throughput. We have noticed that you can arrive to an optimal scenario, but that often requires a lot of samples. So we are most cost efficient at 3,000 samples a day. As soon as we deviate from that, definitely go lower that mean, our cost per sample increases. The qPCR itself is a marginal cost. I think we're at about 1 Euro only for primers and probes and the qPCR extraction is a couple of euros. But of course you need to invest in instruments, you have depreciation, you have labor, et cetera. So it's a pretty complicated scheme that will differ from location to location. I'm happy to discuss more offline if it's really of interest.
Okay. There's another question. Is the software housed in servers on site or it's cloud based?
Yes, it's cloud based. So UgenTec has their software on the cloud and everything is done in the cloud in a secure environment, per industry standards. Extremely efficient, extremely helpful to use in such high throughput environment.
Okay. One more question here. Did you consider extraction free methods and RT-qPCR directly on the sample?
Yes, we did. We actually tested it, drew a few pre-prints, maybe also a few papers. Of course, that would be the Holy Grail, no extraction because extraction is expensive. It's time consuming. However, we tested, I think five or six different extraction-free methods and we were disappointed in terms of their sensitivity. Do they work? Yes, but you lose sensitivity. So for a diagnostic assay where you see that huge range of viral loads from extremely high positive to basically barely positive, we didn't want to take any risk and to maintain high diagnostic sensitivity, we did not pursue extra free methods. If you are sure that you work in a regime with high positive samples, you could definitely consider it. It increases throughput, decreases costs, but it comes invariably with reduced sensitivity.