Detecting somatic single nucleotide variants from cell-free nucleic acid with application to minimal residual disease monitoring

Inventors

ZHOU, Xianghong Jasmine • Li, Shuo • Li, Wenyuan

Assignees

University of California San Diego UCSD

Abstract

The present disclosure provides a probabilistic model for accurate and sensitive somatic single nucleotide variant (SNV) detection in cell-free nucleic acid samples comprising a set of sequence data. A joint genotype may be determined for each locus in the set of sequence data, and germline mutations may be intrinsically removed. A set of filtrations can be applied to eliminate low quality somatic variant calls. Further, a global tumor cell-free deoxyribonucleic acid (cfDNA) fraction and overlapping read mates can be considered, thereby enabling accurate SNV detection and variant allele frequency estimation from samples with low tumor cfDNA fraction. A sensitive early detection of minimal residual disease (MRD) is designed by using the probabilistic model and the machine learning model for distinguishing true variants from sequencing errors.

Core Innovation

The invention provides a method for treating a subject with minimal residual disease (MRD) by detecting a set of k truncal mutations before surgery and then determining whether MRD is present after surgery using a plasma cell-free deoxyribonucleic acid (cfDNA) sample. The method assays a pre-surgery blood sample, a resected tumor sample, or a biopsy tumor sample to detect the set of k truncal mutations, and assays post-surgery plasma cfDNA to detect a post-surgery mutation profile for the set of k mutations. MRD detection is performed based at least in part on the post-surgery mutation profile for the set of k truncal mutations.

The approach processes sequencing data from plasma cfDNA reads that align to genomic positions of the set of k truncal mutations by extracting a set of sequencing reads and creating a feature profile for each sequencing read. The feature profile is processed using a trained machine learning classifier that classifies each sequencing read as either having a true variant or having a sequencing error. The method determines an MRD predictive score indicative of a proportion, among the set of sequencing reads aligning to the genomic position of the set of k truncal mutations, that are classified as sequencing reads having true variants, and then detects the MRD based at least in part on the MRD predictive score.

The method further links MRD detection to treatment by administering cancer therapy responsive to detecting the MRD. The cancer therapy comprises one or more modalities selected from chemotherapy, radiation therapy, follow-up surgery, immunotherapy, cell therapy, proton therapy, or a combination thereof. This operational workflow connects truncal-mutation read-level classification to a predictive MRD score and uses that score to drive therapeutic administration.

Claims Coverage

The independent claim coverage includes 3 inventive features.

Across the independent claim, MRD is detected for treatment purposes using truncal mutations from pre-surgery samples, plasma cfDNA sequencing read extraction and per-read feature profiling, a trained machine learning classifier distinguishing true variants from sequencing errors, an MRD predictive score reflecting the proportion of reads classified as true variants, and administration of cancer therapy responsive to MRD detection.

Stated Advantages

Documented Applications

No documented applications found