GRAIL, LLC was granted US12688905B2 on July 21, 2026, covering a method for spotting DNA from someone other than the patient inside a blood sample before that sample is scored for cancer. The patent, titled “Method and system of cancer detection using CpG-SNP contamination markers,” names inventors Onur Sakarya, Christopher Chang, Ajinkya Kokate and Samuel S. Gross. For a company whose commercial product is a multi-cancer blood test read out by a machine-learning classifier, the grant covers something less glamorous than the classifier itself and arguably closer to the revenue: the quality gate that decides whether a tube of blood gets scored at all.

The business problem the claims address is sample integrity at scale. A liquid-biopsy screening test runs on cell-free DNA circulating in plasma, and a screening business runs on volume — drawn in clinics, shipped, handled, and pooled through a sequencing lab. If fragments from another person enter a sample anywhere along that chain, a classifier trained to read faint methylation signal has no built-in way to know it is reading two people. The claimed method gives the pipeline a fingerprint it can check against. It targets single-nucleotide polymorphisms that sit on top of methylation sites — SNPs that either create a CpG site that was not there or delete one that was — and uses those positions as identity markers.

Mechanically, the claim looks for markers where the patient is homozygous, meaning both copies read the same. Fragments that disagree with the patient’s own homozygous reading at those positions did not come from the patient. Claim 1 then applies a threshold, in these terms:

determining if the test sample is contaminated based on a number of contamination cfDNA fragments being below a threshold— Method and system of cancer detection using CpG-SNP contamination markers, US12688905B2

What the claim actually says

That sentence is step (v) of claim 1, quoted as it issued. It is worth reading slowly, because the surrounding steps run the other direction. Step (vi) of the same claim reads: “in response to determining that the test sample is contaminated, excluding the sample from further analysis.” Read literally and in sequence, claim 1 finds a sample contaminated when the number of contamination fragments falls below a threshold, and then discards that sample. The patent’s own specification describes the same operation in the opposite terms — “In one or more embodiments, a sample with an above-threshold number of contamination fragments may be labeled as contaminated and withheld from further analyses.” — as does a later passage stating that above the threshold quantity, the system marks the sample as contaminated.

We are reporting the language as it issued and not characterizing what it means for the patent. The point for a business reader is narrower and factual: the operative independent claim and the written description use inverted phrasing for the same threshold test, and anyone assessing what US12688905B2 covers will be reading step (v) as it stands, not the description. The abstract, for its part, does not recite the threshold step at all. It stops at fragment labeling: “The system determines fragments having a haplotype that is different from the homozygous haplotype of the sample to be contamination fragments.” Anyone working from the abstract alone would not encounter the discrepancy.

The rest of claim 1 is more conventional and, commercially, is the part that ties the method to a product. The claim requires sequencing with probes designed to target the contamination markers — meaning the identity check is designed into the assay panel, not bolted on in software afterward — and it enumerates the specific polymorphism-and-dinucleotide combinations that count as additive or subtractive CpG-SNP sites. It closes by applying a classification model to the reads while excluding those flagged as contamination, and outputting a cancer prediction. Dependent claims narrow the marker set further: claim 3 requires a threshold population methylation frequency selected from the range of 70%-100%, claim 4 requires population haplotype frequencies within the range of 45%-55%, and claim 5 requires the haplotypes to be in Hardy-Weinberg equilibrium. Claim 17 restates the method as a system built around a sequencing device.

The record is classified under G16B 20/20, G16B 40/00 and G16H 50/20 — computational genomics for sequence variation, machine learning applied to bioinformatics, and clinical decision support. That placement is itself a signal about where the asserted value sits. This is not a chemistry patent on a library-prep reagent; it is coverage on the analytical pipeline, in the classes where diagnostic software competition is now fought.

Where it sits in the estate

The grant lands in a cluster of GRAIL records pointed at the same layer of the stack. US12646621B2, “Systems and methods for cancer condition determination using autoencoders,” issued June 2, 2026 and covers training an autoencoder on methylation patterns and using its reconstruction error as features for a supervised cancer-state model. US12580051B2, issued March 17, 2026, covers identifying methylation patterns that discriminate a cancer condition by building interval maps of fragment methylation patterns and scanning them against selection criteria. US12626780B2, issued May 12, 2026, covers selecting and managing data of high dimensionality — sequence reads tied to a disease condition. US12499972B2, issued December 16, 2025, covers a significance model for predicting noise in cfDNA read information and flagging false-positive variants.

Read together, those four plus the new grant sketch a stack rather than a scatter: noise modeling, contamination detection, feature identification, dimensionality handling, and classification, each covered separately. The company’s older grants sit further upstream in the wet lab — US10487358B2 (2019) on preparing sequencing libraries from mixed cfDNA containing double-stranded, damaged and single-stranded molecules, and US10144962B2 (2018) on differentially tagging RNA to sequence RNA and DNA from one sample. The shift in the intervening years, by the record, is from sample chemistry toward what happens to the reads afterward.

One housekeeping note that matters when tracing the portfolio: the company files under more than one name. The index attributes 31 records to “GRAIL, LLC,” with a further 27 spread across “GRAIL, Inc.,” “GRAIL, INC.,” “Grail, Inc.” and “Grail, LLC” — the older corporate form appears on grants through at least mid-2026, since patents issue under whatever name was recorded at assignment. Unrelated entities including Color Grail Research and Grail Gear LLC also surface on a name search and belong to nobody in this story. A name search is also an incomplete census in the other direction: a large share of records carry no assignee value at all, so any count of this kind is a floor, not a total.

What the new grant buys, in commercial terms, is coverage on a step that sits between the lab and the answer. Contamination screening is not a feature a screening test markets; it is the reason a result can be issued at all, and it is the kind of step a competing multi-cancer test would need in some form. GRAIL now holds an issued US claim on doing it with SNPs that toggle methylation sites, targeted by the assay’s own probes. The claim language on the threshold test is what it is, and any party reading the scope of that coverage will be reading step (v) exactly as printed above.