Cyclomics: ultra-sensitive nanopore sequencing of cell free tumor DNA

Jeroen de Ridder took the stage for the evening session, introducing his efforts to create an ultra-sensitive test for cell-free DNA (cfDNA). This test is in development in order to address two very important challenges in cancer diagnostics. The first of these is treatment response monitoring; allowing data to influence decisions on whether to pursue treatment, switch to an alternative, escalate the scale of treatment or bring the levels down. The second challenge is that of recurrence monitoring – being able to get an early indication of disease recurrence would give the opportunity to restart treatment as soon as a recurrence is detected.

For these purposes, Jeroen describes current diagnostic methods, such as needle biopsies and MRI scans, aren’t fit for purpose, so liquid biopsies present the only reasonable avenue for solving these problems. This means then that Jeroen and the Cyclomics team have focussed their efforts on cell-free DNA. When cells die from apoptosis, they shed their DNA into the blood stream. This is not limited to healthy cells alone though, so cell-free tumour DNA contain mutations that can indicate the presence of that tumour.

At its most simple, Jeroen likens this challenge to an extremely complex game of “Where’s Waldo?”, as cfDNA molecules are extremely small, and only a small fraction of the total molecules present in a sample will actually contain the mutation. In one vial of blood then, there could be between only 10 and 1000 molecules with the mutation actually in. This really highlights the need for something very sensitive and fast to find these mutations.

So, Jeroen asked, how do we solve this very complex game of “Where’s Waldo?”. Introducing his young daughter holding a MinION, Jeroen explained that the solution also needs to be simple and ideally cost effective if it is to be the future of detection. This idea lead the Cyclomics team to begin developing an assay based on the MinION to fulfil that purpose, the method for which is called Cyclomics-seq.

Cyclomics-seq works on an enrichment strategy, capturing short molecules into circular molecules via the use of a backbone and rolling circle amplification, giving multiple copies of the original molecule in long stretches. These molecules can then be prepared for sequencing on the MinION, circumventing any random sequencing error by producing a per-molecule consensus sequence. Jeroen explained how this would create a sensitive, fast and flexible workflow that allows him to tell if a mutation is present in the patient molecule in an accurate fashion.

The bioinformatics part of the workflow consists of mapping with LAST-split and consensus calling with DAGCON followed by mutation detection by mapping with BWA-MEM and allele frequency detection with Sambamba.

In their initial results, the read length distribution showed a high number of reads over the 5 kb mark, indicating the majority of reads contained ten or more copies of the original insert of 400 bp. Some reads though, Jeroen went on to explain, had mapping gaps and didn’t align as well to the reference as the team would like. This meant that the assay underwent several rounds of development, and Jeroen demonstrated the results of this development with progressive graphs of the correlation between backbone and insert length. Data on the diagonal represented usable data, and the accumulation of points got stronger per iteration. Improvements came from several avenues, including new backbone designs and new enzymes between revisions.

To evaluate the performance of the assay, Jeroen and team mapped the median false positive rate versus the number of repeats in the molecule, finding that the false positive rate levelled off after 10 repeats. The false positive rate, although small, remained consistent at top end of the repeat number distribution, indicating that there were some errors still remaining that couldn’t be eliminated via consensus pile up.

Taking their developed assay to proof-of-concept stage, Jeroen and the Cyclomics team took pools which were 100% wild type, 100% mutant, and various dilution stages in between, sequenced them, and mapped the base called against the reference position at the known mutant point. At 100% wildtype, only a “handful” of reads displayed the wrong base, and the same was true of the 100% mutant pool. In the diluted samples the ratio of mutant to wildtype base diverges, meaning the mutant molecules are still distinguishable against a significant wildtype background.

The dilution series success gave Jeroen the confidence to move on to testing on clinical samples. Recurrence, he described, is a major problem in head and neck cancer where, even after a seemingly successful treatment, 50% of patients will experience a relapse. Head and neck cancers are also often extremely hard to image, meaning by the time a tumour is detected on an MRI scan it is often too late to tackle it effectively. Again, this points to liquid biopsy as the only reasonable way forward for early detection of recurrence.

In terms of the target for the liquid biopsy, Jeroen explained that 72% of all cancer patients possess mutation in the tumour suppressor gene TP53, and if you look at head and neck cancer patients only this rises to 90%, making it a very logical choice to focus on. The first port of call was to work on addressing the false positive rate mentioned earlier in the more generalised proof of concept. Initial results on TP53 showed that one third of the reads contained no observable error over the target stretches, but rebasecalling with the high accuracy guppy version resulted in the number of regions with elevated error dropping significantly, opening up the possibility of including those regions in the final assay.

Addressing this further, the Cyclomics team has received some early access flow cells of the new pore, R10, which they also used to evaluate the false positive rates in TP53. The results, Jeroen said, looked “really quite promising”, as far more areas contained no observable errors, meaning more loci that could be included in the assay. There were still some problematic areas in the R10 data, but interestingly, when Jeroen mapped the erroneous areas for R10 against the tricky areas in R9.4, the error profiles did not overlap, meaning different areas were high accuracy. This throws up all sorts of interesting possibilities for use of the two pores – be that selecting the right flow cell for the target of the assay, or combining the two types of data, or another avenue.

Following the development and testing stages, the Cyclomics team set up for clinical trial with 30 early stage and 20 recurrence patients. The question they wanted to answer, Jeroen explained, was whether the Cyclomics protocol allowed the team to make better decisions on treatment regime than the normal monitoring process. The trial began late in 2018, so is mid-progress at the moment, allowing only data for the first two patients to be shared.

The first patient example Jeroen gave presented with a stage II tumour and had a mutation in TP53 at 60% prevalence. They had undergone radiation and chemotherapy, and although the MRI scans showed no visible difference in the first few weeks post-treatment, 9 months later the patient displayed residual disease at lymph nodes. The team obtained blood samples pre-treatment, and at week 2 and 5 post treatment. After including controls of healthy subjects, the preliminary results showed that the mutant levels in the blood dropped in post treatment, only to rise again by week 5, indicating there were some residual disease present very early in the process. This example gave what Jeroen described as a “fine-grain look at treatment response” that correlated with the eventual MRI visual phenotype.

The second patient example consisted of a presentation of a stage IV tumour, meaning data could only be obtained at the pretreatment phase. The aim for this data was to demonstrate that the relevant mutation could be clearly identified, and quickly, so Jeroen plotted the sequencing results as a function of sequencing time. This graph showed that within 5-7 hours, the team could see ample evidence for the presence of problematic mutations in the patient’s blood.

Based on these early examples, the Cyclomics team are optimistic that the process they develop can allow much more rapid interventions into treatment cycles by acting on mutation data on a quick turnaround.

To conclude his talk, Jeroen highlighted some of the methods for improvement he is pursuing right now. Referring back to the improvement basecalling with the new guppy basecaller made to the false positive rate, Jeroen explained that this gives him a clear conclusion – all the information must be in the signal, and it is the interpretation of that data that makes a crucial difference.

This lead the team to develop an approach using dynamic time warping on the backbone signal and insert signals, and applying deep learning to that data to assess whether a mutation is present directly from the signal, instead of passing via a basecalling step. The accuracy of detection across training epochs increased significantly, illustrating clearly that training and optimising for a particular target can massively increase algorithmic capability. The first results of using the algorithm were very promising, clearly showing the difference between wildtype and mutant reads at very high accuracy, and the hope is that these deep learning approaches will be utilised in the Cyclomics-seq pipeline very soon.

Finally, Jeroen described his overall vision – that Cyclomics-seq should reach patients – and explained that Cyclomics was now a newly formed spinout company who aim to market an ultra-flexible blood test for TP53. Subsequently, this could be expanded to additional fundamental gene targets, but the team are also actively pursing widening their approach to a whole-genome method that could incorporate base modification analysis too. Overall, the ambition is to be the first liquid biopsy company based on a third generation sequencing platform.