Estimating tumor purity from single samples

Inventors

Phillips, Nicholas • Harris, Jason

Assignees

Personalis Inc

Interested in licensing this patent?

MTEC can help explore whether this patent might be available for licensing for your application.

Publication Number

US-12512183-B2

Patent

Publication Date

2025-12-30

Expiration Date


Abstract

The disclosure provides methods for estimating tumor purity from tumor samples without use of matched-normal controls. A set of genomic regions are identified based on a nucleic acid sequence data that is aligned to a reference genome. Each genomic region of the set of genomic regions includes one or more nucleotide-sequence variants relative to a corresponding genomic region of the reference genome. A B-allele frequency distribution for the biological sample is determined based on a B-allele frequency determined for each genomic region of the set of genomic regions. The B-allele frequency distribution is processed using a trained machine-learning model to estimate a metric identifying tumor purity in the biological sample.

Core Innovation

The invention relates to a computer-implemented method of determining a likelihood of cancer in a subject and evaluating a therapeutic regimen for the subject. The method obtains a set of nucleic acid sequence data comprising whole genome sequencing data generated using a sequencer on a cfDNA sample of the subject, aligns the nucleic-acid sequencing data to a reference genome, and identifies genomic regions with nucleotide-sequence variants.

For the identified genomic regions, per-region B-allele frequency at heterozygous sites is computed to form a normalized BAF distribution. The BAF distribution is processed with trained machine-learning models to output a tumor-purity metric, and the method determines whether one or more genomic regions within the obtained nucleic acid sequence data comprise polymorphisms.

The trained machine-learning models include fully connected neural networks and one-dimensional or two-dimensional convolutional neural networks, and other regressors. Training datasets are created from nucleic acid sequence data derived from tumor cells diluted into normal cells, including VCF-level datasets and FASTQ-level datasets, using in silico tumor cell dilution into normal cells, and the outputs are used in downstream analysis and reporting.

Claims Coverage

The independent claim covers a computer-implemented method that obtains whole genome sequencing cfDNA data, uses one or more trained machine-learning models to determine whether genomic regions comprise polymorphisms based on a tumor-dilution training dataset, and generates a report to inform the likelihood of cancer. The coverage includes 7 inventive features.

Likelihood of cancer report from cfDNA whole genome sequencing

A computer-implemented method of determining a likelihood of cancer in a subject and evaluating a therapeutic regimen for the subject, comprising obtaining whole genome sequencing data generated using a sequencer on a cfDNA sample of the subject; determining whether one or more genomic regions within the obtained nucleic acid sequence data comprise polymorphisms using one or more machine learning models; and generating a report to inform the likelihood of cancer tumor based on the determination.

Machine-learning models trained on tumor cells diluted into normal cells

Using one or more machine-learning models trained on a training dataset generated from nucleic acid sequence data derived from one or more tumor cells diluted into normal cells, to determine whether one or more genomic regions comprise polymorphisms.

Genomic regions as polymorphism-containing regions relative to a reference genome

Identifying, from aligned nucleic acid sequence data, a set of genomic regions in which each region includes one or more candidate nucleotide-sequence variants relative to a corresponding genomic region of the reference genome.

Polymorphisms include specified variant types including CNVs

Identifying a set of genomic regions by detecting one or more of base changes, insertions, deletions, repeats, transversions, or copy number variants (CNVs).

Training uses nucleotide-sequence variant datasets with specified variant classes

Training on a nucleotide-sequence variant dataset where the nucleotide-sequence variants include at least one of a single nucleotide polymorphisms, one or more single nucleotide variations, insertion-deletion mutations, small insertions, small deletions, structural variant junctions, flanking sequences, and a combination thereof.

CfDNA sample isolated from plasma or serum

Providing that the subject’s cfDNA sample is isolated from one or more of plasma or serum.

Sequencing approaches for obtaining nucleic acid sequence data

Obtaining a set of nucleic acid sequence data using at least one sequencing approach selected from sequencing by synthesis, allele specific hybridization to a library of labeled oligonucleotide probes, and sequencing by synthesis using allele specific hybridization to a library of labeled clones followed by ligation.

Across the independent claim and its refinements, the core coverage is directed to generating a report informing cancer likelihood and evaluating a therapeutic regimen by using machine-learning models trained on tumor-cell-diluted-in-normal training data to determine whether genomic regions contain polymorphisms, where polymorphism-containing regions are defined relative to a reference genome and include specified variant types and training variant classes.

Stated Advantages

Documented Applications

Generating a report to inform the likelihood of cancer tumor.

Evaluating a therapeutic regimen for the subject.

Downstream analysis and reporting, including predicted somatic variants and diagnostic/prognostic markers.

Treatment recommendation, including use with diagnostic/prognostic markers and predicted somatic variants.

JOIN OUR MAILING LIST

Stay Connected with MTEC

Keep up with active and upcoming solicitations, MTEC news and other valuable information.