Grail Shares New Data From Early Detection Assay Training in CCGA Study

CHICAGO (GenomeWeb) – Grail shared the first data from its efforts to develop a blood-based test for early cancer detection, focused mainly on assay specificity, at the annual meeting of the American Association for Cancer Research yesterday.

The company reported on a first analysis from its ongoing Circulating Cell-free Genome Atlas study, which has recruited over 11,000 individuals so far, and is expected to reach a planned 15,000 participants by the end of this year. In a training set of about 1,800 participants, Grail researchers were able to identify predictive signatures that segregated cancer cases with significant sensitivity while minimizing false positives in those without.

More specifically, the company reported that it detected a cancer-like signal in only five of more than 500 cancer-free enrollees, two of whom shortly went on to be diagnosed with a malignancy.

The CCGA is a longitudinal cohort study conducted at 141 sites with the goal of providing geographically diverse enrollment representative of the US population as a whole. The cohort is divided so that 70 percent of recruits have cancer and 30 percent are healthy controls.

Researchers collect blood samples and perform four different analyses — targeted sequencing, whole-genome sequencing, targeted and whole-genome bisulfite sequencing, and whole-transcriptome RNA seq. For participants with cancer, tissue samples are also analyzed when available. Grail also is using machine learning to mine the data from these assays in order to train a classifier that can distinguish cancer from non-cancer.

As the effort proceeds, Grail is following the group for up to five years, watching the non-cancer segment for future cancer diagnoses and the cancer cohort for data on treatment, recurrence, and mortality.

The new analysis presented at AACR reflects the CCGA's first stab at developing a classifier, using a group of 2,800 participants broken into a training set (1,792 individuals) and a test set (1,008 individuals). At the meeting, Grail Vice President of Research and Development Alexander Aravanis reported only on the training set, which was reduced to 1,399 individuals — 841 with cancer and 558 without. He said that the company plans to share data on performance of the test set at future meetings.

In the cancer group there were over 20 tumor types, with breast, lung, prostate, and colorectal cancers being the most frequent.

Reflective of Grail's efforts to recruit evenly across the US, Aravanis said that the distribution of cancer types and stages in the cases reported at AACR were largely consistent with known cancer incidence patterns and included early cancers detected by current screening programs or protocols and later-stage cancers diagnosed based on clinical presentation.

To train classifiers, samples from the training cohort were subject to three different analyses. Investigators performed whole-genome bisulfite sequencing, as well as whole-genome sequencing and targeted sequencing of cell-free DNA. In the latter two analyses, results were compared to a matched analysis of white blood cell DNA in order to control for background germline alterations and clonal hematopoiesis.

According to Aravanis, evidence of the value of matched white blood cell sequencing can be seen in the fact that non-tumor variants accounted for, on average, 78 percent of all variants detected in the non-cancer group and about 66 percent in the cancer group.

"This high background of cell-free mutational burden coming from the white blood cells is a very strong confounding biological signal that had to be corrected for both in our assays and in our informatics in order to maintain high specificity — particularly in the targeted sequencing assay," he said.

Aravanis highlighted a few examples from the whole-genome and whole-genome methylation data that illustrate the type of cancer-specific features that Grail's approach was designed to identify. These included copy number profiles derived from the whole-genome sequencing data and aberrant methylation patterns gleaned from the bisulfite sequencing approach.

Overall, the company found that the three classification approaches were highly correlated, with performance across the board being driven by the fraction of tumor DNA in the blood. But the methylation analysis looks so far to be the strongest distinguisher.

Across all three assays, the group only saw a "cancer-like" signal in five individuals — less than 1 percent of the 580 cancer-free subjects analyzed.

Underlining again the value of the company's germline-matching approach, while eight enrollees from the non-cancer group initially showed somatic copy number alterations in their cfDNA sequencing, four of those were found to be white blood cell matched, and so were able to be excluded as having a cancer-like signal.

As Grail has continued to follow the five individuals from the non-cancer group who had a cancer-like signal, two have since been confirmed to have cancer — one with stage III ovarian cancer and the other with stage II endometrial cancer.

"These are important examples where there are cancer signals in the non-cancer group at time of enrollment that appear to have anticipated their cancer diagnosis," Aravanis said.

In an email yesterday, Grail clarified that the study does not return results to participants, so neither of these cancer diagnoses were precipitated by its finding of a cancer-like signature.

Although specificity was a main focus of the Grail presentation and is deeply important in the context of cancer screening, the study also offered an early hint of what kind of sensitivity the company may be able to achieve.

In addition to the two signals detected in healthy individuals who were shortly diagnosed with a cancer, Aravanis also reported that the company saw strong genomic and metagenomic signals across several cancer types that aren't widely screened for and that have particularly dismal survival rates.

According to the company's report, Grail could detect some cases among all stages of cancer in the cohort, but there was better sensitivity with increasing stage for all three assays. With specificity adjusted to 95 percent, bisulfite sequencing could pick up 65 percent of stage I-III cancers, and a full 95 percent of the stage IV cancers, Aravanis reported. Whole-genome ctDNA sequencing showed 61 percent sensitivity in earlier stages and 89 percent detection of stage IV cancers. And the targeted sequencing approach showed 50 percent detection and up to 80 percent in stage IV.

Aravanis also broke out data specifically for colorectal cancer during his presentation at the meeting, reporting that the company's analyses could pick up 69 percent of stage I and II CRC cases, increasing to 85 percent detection of stage III and IV tumors.

An audience member at the AACR presentation raised one important question about assessing sensitivity with the long-term follow-up of the trial still in its infancy — namely that the company hasn't followed the non-cancer group long enough yet to know that its non-cancer calls were true negatives and not false negatives.

Aravanis agreed, saying that the since the study is still only in the earliest phases of follow-up, these questions can't be answered yet, though they are going to be very important to track.

Freenome, too

Freenome, another company working toward commercializing early cancer screening tools, also presented data at the AACR meeting in two posters.

One of these reiterated a point that Freenome has stressed previously — that current sequencing costs and reimbursement levels make tests that rely only on broad and deep sequencing of circulating cell-free DNA impractical if not infeasible. As a result, integrating markers beyond DNA mutations and using machine learning, as Freenome has said it is doing, should help to overcome these challenges.

Freenome has promoted its embrace of an AI-supported, multipronged approach as a distinguisher from potential competitors like Grail, but based on the new data presented by Grail at the AACR meeting, both companies look to be proceeding with some respect for these caveats and for combinatorial, algorithmic development.

In an email, Aravanis reiterated that though it did not report on this at AACR, Grail is also doing broad sequencing of circulating RNA, which may surface additional signals of cancer — derived both from tumors and from other biological processes.

He also explained during the session that Grail's broad, deep sequencing approach is crucial for its assay development strategy, and appropriate for the discovery phase of the CCGA study. Doing this comprehensive analysis now, he argued, is what will hopefully allow the company to glean simpler more clinically feasible structures to bring to market.

In a second Freenome poster, investigators also shared some of the first data from their own ongoing analyses in a study of colorectal cancer patients.

The firm reported that it has implemented a system for the integrated analysis of multiple analytes, including cell-free DNA mutation detection and bisulfite sequencing similar to what Grail has also employed, plus microRNA analyses and detection of circulating proteins.

Researchers from Freenome studied blood samples from a small cohort that included 26 healthy individuals, 23 individuals with pre-malignant adenomas, and a similar number with colon cancers distributed from stage I to stage IV. Much like Grail, the company saw that genome-wide copy number variation could distinguish patients with a high tumor fraction — largely in later-stage patients, but also in some early cancers.

Signals of cancer could also be seen using the other analytes, including genome-wide methylation changes, elevated CEA and CYFRA 21-1 proteins, and aberrant miRNA profiles.

Interestingly, according to the Freenome authors, aberrant profiles among cell-free miRNA, cfDNA methylation, and circulating protein levels were more strongly associated with a high tumor fraction than with late-stage disease.

The team wrote that these findings suggest that some positive "early-stage" detection results might be better defined as "high tumor fraction" detection results, implying that successes demonstrated in detecting early-stage cancers could be limited to those with high tumor fractions.

But in the Grail data on colorectal cancer that Aravanis shared, although higher tumor fraction was clearly correlated with increased success for the classifiers, there was a wide distribution, with cases detected even at low tumor fractions in both the stage I/II and III/IV groups, as estimated by analysis of matched tumor tissue analyses.

Aravanis said at AACR that Grail plans to present further breakdowns of specific tumor types in its cohort at an upcoming medical meeting. Moving forward, the firm is focused on moving its analysis into a larger group of participants enriched for clinically important stage I and II cancers.

He also said that Grail looks at the data it shared this week as a starting point. The company believes that with increased sample sizes, its machine-learning approach is likely to only improve performance from here.

A genomic analysis of modern and ancient maize reveals a complicated domestication history, according to Reuters.

In PLOS this week: MYRF variant linked to congenital diaphragmatic hernia, analysis of the "dragon's blood" red resin produced by traditional medicine plants, and more.

CNBC reports that half of academic researchers leave after about five years.

Researchers have used genetic analysis to confirm a new type of salamander, the New York Times reports.

Jan
30
Sponsored by
Loop Genomics

This webinar will provide a comparison of several next-generation sequencing (NGS) approaches — including short-read 16S, whole-genome sequencing (WGS), and synthetic long-read sequencing technology — for use in microbiome research studies.

Jan
30
Sponsored by
Loop Genomics

This webinar will provide a comparison of several next-generation sequencing (NGS) approaches — including short-read 16S, whole-genome sequencing (WGS), and synthetic long-read sequencing technology — for use in microbiome research studies.