NEW YORK – Researchers from the University Medical Center Utrecht in the Netherlands and their collaborators have developed a new workflow that promises to rapidly classify central nervous system (CNS) tumors during surgery by combining ultra-fast nanopore sequencing with machine learning algorithms.
Described in a proof-of-concept study published in Nature on Wednesday, the method illustrates nanopore sequencing’s potential for informing surgical decision-making, paving the way for the technology’s routine deployment in the operating room.
"The first-line treatment for CNS tumors is surgery," said Jeroen de Ridder, a principal investigator at UMC Utrecht and one of the corresponding authors of the study. "But how aggressively that surgical procedure happens depends on the tumor type."
While for certain CNS tumors, especially those that are currently incurable, it might be better to prioritize preventing neurological damage and comorbidity caused by the surgery, de Ridder explained, for some other types, it is absolutely essential to clear out as many tumor cells as possible to minimize the patient’s risk of recurrence and the need for a second surgery.
However, the exact tumor type is hard to determine before surgery due to the sample's inaccessibility, de Ridder said. Traditionally, clinicians rely on preoperative imaging and rapid histological analysis, where frozen tumor sections are assessed under the microscope during the procedure, to help determine the cancer type. This is not always conclusive and can be even incorrect, he pointed out.
In recent years, genome-wide DNA methylation profiles have become a powerful tool to help distinguish tumor subtypes. But current methods for methylation analysis normally involve microarrays or bisulfite short-reading sequencing, which have a turnaround time of at least several days, making these technologies not very useful for guiding decision-making during surgery, de Ridder said.
To address these bottlenecks, de Ridder and his team turned to nanopore sequencing, which enables the direct measurement of methylated cytosines in real time.
According to de Ridder, one challenge of using nanopore sequencing intraoperatively is that, given the time constraints, only "extremely sparse" methylation signals can be generated in a short amount of sequencing time, usually at most 30 minutes. In addition, it is unknown ahead of time which part of the genome will be sequenced to generate the methylation signals.
"I always compare it metaphorically to, ‘I show you an image, and give you only one percent of the pixels of the image, and I don't tell you which pixels I will give you, but you still need to recognize what is in the image'," he explained.
To cope, the UMC researchers developed a deep-learning tool named Sturgeon, a patient-agnostic classifier neural network that is optimized to classify CNS tumors with sparse data.
Given that existing reference databases for CNS tumor methylation profiles are mostly built upon microarray-based assays, de Ridder said, the team also had to generate simulated nanopore sequencing data to help train the Sturgeon algorithm.
"One of the innovative aspects of this study is that we used simulation to bridge this gap, so that we could still learn from the reference microarray data and then apply it effectively to the latest nanopore data," he noted.
Overall, Sturgeon was trained on 36.8 million simulated nanopore runs and validated on a further 4.2 million simulated nanopore runs, according to the study authors.
To test the performance of the Sturgeon, the UMC researchers applied the algorithm to sparse nanopore sequencing data from 50 retrospective CNS tumor samples. The tool was able to correctly classify 45 out of the 50 samples based on data equivalent to 20 minutes to 40 minutes of sequencing.
In a second step, the researchers tested Sturgeon and the rapid nanopore sequencing workflow in real intraoperative scenarios. In 25 surgeries where the method was deployed, the workflow achieved 18 correct diagnoses within 90 minutes, which is the desired timeframe for guiding surgical decision-making. Meanwhile, the other seven cases did not meet the required confidence threshold to reach a conclusion, the authors noted.
"What we have now uniquely enabled is to still have the richness of the molecular data and the resulting accuracy in terms of being able to distinguish between different tumor types, and so fast that it is compatible with a surgical procedure," de Ridder said.
"This is a great first paper," said Stephen Kingsmore, CEO and president of Rady Children's Institute for Genomic Medicine in San Diego, who was not involved in the study.
A surrogate for molecular profiling, methylation profiles can be used for both guiding surgical resection and appropriate treatment after the surgery, Kingsmore said. "The problem has been that, although people have built good classifier tools to actually distinguish the cancers and their outcomes, up until now, there has been no way to do this quickly."
According to Kingsmore, Rady Children's started to offer a comprehensive molecular tumor board analysis, which includes deep tumor genome and RNA sequencing, germline genome sequencing, and methylation microarrays, for patients with CNS tumors seven years ago. The goal for that program, he said, is to achieve a turnaround time of 21 days, which is normally the time between surgery and the start of chemotherapy.
Kingsmore praised the study authors’ achievement of turning a methylation profile analysis assay into a sub-two-hour workflow that can be applied during surgery, making "a huge, huge difference" for surgical decision-making.
While he said the researchers showed "a pretty good performance" of their method, he also thinks it is important to see other groups recapitulate it in the future to demonstrate its generalizability as well as to help improve the machine learning model.
In addition, he noted that the current study does not address outcomes. "It's not enough just to classify tumors," he said, pointing out the need for longitudinal follow-up to demonstrate the method’s positive impact on patients' outcomes, such as improved five-year survival or certain neurologic milestones.
"This is a paradigm shift," said Danny Miller, a pediatric geneticist at the University of Washington who was also not involved in the study. "You can imagine such a device in the OR, where the surgeon gives the sample to a scrub tech who then loads the device, and 30 to 40 minutes later, they get some type of answer."
Miller’s team, itself at the forefront of ultra-rapid nanopore sequencing, previously used the technology to help elucidate a newborn's genetic risk for a specific disorder in as little as three hours after birth.
He noted that "the story you get out does not totally reflect all the effort that goes in" when it comes to rapid nanopore sequencing. For example, there is a need for coordination and logistical planning in order to obtain results at such a fast speed.
Despite the method’s promise, Miller, echoing Kingsmore's points, also thinks the question moving forward is how this workflow can be standardized and made reproducible in order to meet the "higher bar" for routine clinical use.
Moreover, he said that given Sturgeon was trained on microarray data, which only covers 450,000 CpG sites, the study also highlights the need to build a long-read-based methylation reference database, where all 44 million CpG sites in the genome are taken into consideration. That, in turn, may lead to a "more refined" classification of the tumors, potentially helping with diagnostic yield, he noted.
Commenting on ability to reproduce the workflow in other settings, de Ridder said the approach only requires "a very modest" lab setup and a laptop with a decent GPU, making "the threshold to start using this extremely low." While his team has not done a formal cost analysis, he estimated that the material cost for each sample is in the range of €500 to €700 ($530 to $740).
While de Ridder and other co-authors have applied for a patent for the technology used in the study, he said the Sturgeon tool is available to researchers through GitHub.
"We hope that by filing a patent, there is at least an opportunity for an industry partner to pick this up and really take the next steps to make sure that this really reaches the clinic," de Ridder said. "We also of course hope that Oxford Nanopore [is] interested in maybe using this as an app on their platform. That would be amazing if they are able to bring this forward."
Oxford Nanopore did not provide comments on any potential commercialization interests it may have for the Sturgeon classifier.
Moving forward, de Ridder said one goal for his team is to continue updating Sturgeon and retrain the algorithms with up-to-date methylation reference data.
In addition, he said the group will collaborate with clinicians towards a bigger validation study, preferably with many different medical centers throughout Europe, to demonstrate the method’s clinical performance and its benefits for patient outcomes.
Beyond that, de Ridder said, the lab is also looking to apply the method more broadly to other cancers that could potentially benefit from a more rapid methylation-based diagnosis.
"I think there are ample opportunities there to make a difference," he said.