Viral cancer detection tests

Abstract

Methods and systems of determining a cancer score, including: receiving a sample collected from a subject, in which the sample includes sequences of circulating tumor DNA (ctDNA); providing the sample to an analysis pipeline, in which the analysis pipeline detects and synthesizes a plurality of features; and determining, with the analysis pipeline, the cancer score based on the synthesized plurality of features.

Description

MEEI 2024-326-02Quarles 169511.00058VIRAL CANCER DETECTION TESTSCROSS-REFERENCE TO RELATED APPLICATIONS

[0001] The present application is based on and claims priority from U.S. Patent Application Ser. No. 63 / 648,472, filed on May 16, 2024, the entire disclosure of which is incorporated herein by reference.BACKGROUND

[0002] Liquid biopsies are a useful tool for cancer profiling and monitoring.Nevertheless, there exists a need to improve the sensitivity of methods and systems to detect cancer, particularly in settings with a low tumor burden such as screening or minimal residual disease. In particular, viral-related cancers, specifically HPV-related cancers, can benefit greatly from highly sensitive, viral related cancer detection systems and methods.SUMMARY

[0003] Disclosed herein are methods and systems for determining a cancer score. In various embodiments, the methods and systems may include one or more of the following.

[0004] In some embodiments, a method of determining a cancer score is provided. In some embodiments, the method includes: receiving a sample collected from a subject, in which the sample includes sequences of circulating tumor DNA (ctDNA); providing the sample to an analysis pipeline, in which the analysis pipeline detects and synthesizes a plurality of features; and determining, with the analysis pipeline, the cancer score based on the synthesized plurality of features. In some embodiments, the ctDNA includes viral circulating tumor DNA.

[0005] In some embodiments, a system for determining a cancer score is provided. In some embodiments, the system includes: a processor in communication with a memory, the memory having stored thereon a set of instructions, which, when executed by the processor, cause the processor to: receive a sample collected from a subject, in which the sample includes sequences of circulating tumor DNA (ctDNA); provide the sample to an analysis pipeline, inMEEI 2024-326-02Quarles 169511.00058which the analysis pipeline detects and synthesizes a plurality of features; and determine, with the analysis pipeline, the cancer score based on the synthesized plurality of features.

[0006] Methods and systems described herein can include a combination of one of more of the following embodiments.

[0007] In some embodiments, the ctDNA include viral ctDNA. In some embodiments, the ctDNA include circulating tumor HPV DNA (ctHPVDNA). The sample may include a combination of ctDNA and HPV antibodies of the subject.

[0008] In some embodiments, the plurality of features may include at least two of HPV reads as determined by unique molecular indices, HPV genome coverage, HPV genotype, HPV lineage and sublineage, normalized HPV genomes / human genomes, viral integration events, select human genes and gene mutations, ctHPVDNA fragment size features, prognostic viral single nucleotide polymorphisms, or prognostic human germline single nucleotide polymorphisms. The plurality of features may further include HPV antibodies, and / or patient demographics. Patient demographics may include sex, age, HIV status, or smoking status.

[0009] In some embodiments, the HPV genotype includes identifying the HPV genotype as one of more than 13 HPV genotypes. Viral integration events may include identifying viral genome breakpoints and human genome breakpoints. Gene mutations may include PIK3CA mutations.

[0010] In further embodiments, the analytics pipeline may include at least one trained network. The trained network may include a multi-label, multi-class network. The trained network may have an architecture that includes, or is based on, at least one of a random forest, extra tress, decision trees, gradient boosting, XGBoost, LightGBM, linear, logistic regression, ridge, lasso, elastic net, kernel-based, support vector machine, instance-based, k-nearest neighbor, probabilistic, or naive bayes. In some embodiments, a trained network is trained on sequencing data. In some embodiments, a trained network is trained on sequencing data and HPV antibody data. In some embodiments, a trained network is used to integrate sequencing data and HPV antibody data. A “first” network and a “second” network are not meant to limit order, it simply denotes that two separate trained networks are used. A first and a second network may each include multiple trained networks. A trained network may be trained on labeled subject dataMEEI 2024-326-02Quarles 169511.00058including data including control samples collected from subjects who have not been diagnosed with a cancer and samples collected from subjects who have been diagnosed with a cancer.

[0011] In some embodiments, the systems and / or methods may be used for at least one of screening, diagnosis, detecting molecular residual disease, prognosis, or surveillance. The method may be used longitudinally for more than one of screening, diagnosis, detecting molecular residual disease, or surveillance.

[0012] In embodiments in which the systems and / or methods are used for screening, the subject may not display any symptoms indicative of cancer at the time of screening.Alternatively, the subject may display symptoms indicative at the time of screening. The sensitivity of screening may be 25-fold more sensitive than dPCR. The sensitivity of screening may be 80-fold more sensitive than dPCR.

[0013] In embodiments in which the systems and / or methods are used for diagnosis, the sensitivity of the diagnosis may be greater than 98% at a specificity greater than 98%.

[0014] The systems and / or methods may be configured to detect a disease more than three and a half years before disease expression. The systems and / or methods may be configured to detect a disease more than 20 months before disease expression.

[0015] In some embodiments, the systems and / or methods may be used to detect molecular residual disease (MRD) of a subject. The systems and / or methods may be able to detect MRD more than 150 days before disease expression. The detection of the MRD may have a sensitivity greater than 50%. The detection of the MRD may have a sensitivity greater than 75%. The detection of the MRD after surgery may be 50% more sensitive than dPRC. The detection of the MRD may be 25% more sensitive than dPCR.

[0016] In some embodiments, a clinical decision may be made based on the results of using the systems and / or methods for screening, detection of MRD, or surveillance. In some embodiments, a patient sample may be a blood sample.MEEI 2024-326-02Quarles 169511.00058BRIEF DESCRIPTION OF THE DRAWINGS

[0017] FIGS. 1 A-1B show phases of tumor burden over time. FIG. 1 A shows tumor burden over the lifetime of subject, including clinical relapse. FIG. IB shows the ability for several methods to diagnose a patient at different levels of relative disease burden.

[0018] FIGS. 2A-2D show an overview of HPV-DeepSeek. FIG. 2A shows an overview of sample collection and preparation for HPV-DeepSeek. FIG. 2B shows training and validation of machine learning used in HPV-DeepSeek 2.0 across all HPV cancers. FIG. 2C shows an overview of a screening study in anal and oropharyngeal cancer. FIG. 2D shows an overview of a validation screening study in anal and oropharyngeal cancer. FIG. 2E shows a comparison of different machine learning architectures.

[0019] FIGS. 3A-3B show aspects of how a cancer score may be calculated. As shown in FIG. 3A, multiple features can be used to determine a final HPV cancer score. FIG. 3B shows an overview of duplex sequencing, which may be used to sequence viral DNA as part of determining features for calculating a cancer score.

[0020] FIGS. 4A-4F show diagnostic efficiency estimates of ctHPVDNAand HPV antibody detection approaches. FIG. 4A shows NNS estimates of a single-step screening approach based on sensitivity for HPV-DeepSeek, multiplex ddPCR and HPV Ab for HPV-associated oropharynx cancer (HPV+OPSCC) in men aged 55-74 based on disease incidence. FIG. 4B shows PPV estimates of a single-step screening approach based on sensitivity and specificity of the three approaches. FIG. 4C shows NNS estimates of a two-step screening approach with HPV Ab followed HPV-DeepSeek only in the population of HPV Ab screen positive men. FIG. 4D shows PPV estimates of a two-step screening approach. FIG. 4E shows a direct comparison of HPV-DeepSeek, dPCR, and Naveris NavDx test based on dPCR ability to detect screening blood samples collected from asymptomatic people who later developed HPV+ head and neck cancer. Green are samples screening positive, red are sample screening negative. HPV-DeepSeek is positive in all samples, extending to >70 months before cancer diagnosis. Naveris NavDx is negative in 4 / 7 samples with a maximum lead time of <40 months, more than a year less. FIG. 4F shows a schematic of the current post-screening test workup and treatment program, and a proposed treatment paradigm using HPV-DeepSeek.MEEI 2024-326-02Quarles 169511.00058

[0021] FIGS. 5A-5H show comparisons of blood-based ctHPVDNA and HPV Antibody detection approaches. FIG. 5A shows a schematic representation of HPV-DeepSeek workflow demonstrating two classification approaches: 1. Read / coverage-based classification and 2. Machine learning-based classification. FIG. 5B, top, shows a determination of HPV-DeepSeek threshold using serial dilutions ofHPV+HNSCC patient cfDNAinto control cfDNA, performed in triplicate. The circle size represents the quantity ofHPV+HNSCC patient cfDNA input. A threshold was set at 10 unique HPV genome reads and 10% HPV genome coverage. Samples below the test threshold are considered negative and are in the gray-shaded region. The white-shaded region contains positive samples. The x-axis represents the number of HPV reads, and the y-axis represents HPV genome coverage as a percent of the total genome. FIG. 5B, bottom, shows an HPV-DeepSeek diagnostic performance in 152 HPV+HNSCC cases and 152 population-level controls using read / coverage classification. HPV+HNSCC cases are green, controls are red. HPV genotypes are represented in different shapes. The 10 read, 10% genome coverage threshold is highlighted with dotted lines. FIG. 5C shows a schematic phylogenetic tree of the genotypic diversity among 152 HPV+ HNSCC cases detected by HPV-DeepSeek. Each branch represents a unique HPV genotype, with branch width proportional to the number of cases harboring that genotype. Sublineages are represented in the same format at the end of the HPV16 branch. FIG. 5D provides a pie chart showing the serology testing ofHPV+HNSCC cases (left) and controls (right) with the color corresponding to each category. FIG. 5E shows a Sankey plot showing a head-to-head comparison between three different diagnostic approaches in a cohort of 103 patients. The leftmost flow represents HPV Ab, the middle flow represents ddPCR, and the rightmost flow represents HPV-DeepSeek. Flow width is proportionate to the number of patients. FIG. 5F shows a scatter plot shows the Spearman correlation between ddPCR positive droplets (x-axis, log scale) and HPV-DeepSeek reads (y-axis, log scale) for 103 HPV+ HNSCC cases analyzed by both assays. Circles in the white region represent samples positive in both assays. Samples in red region are negative by ddPCR but positive by HPV-DeepSeek. Samples in violet region are negative in both assays. Shaded regions approximate the thresholds for HPV-DeepSeek positivity (>10 reads, red) and ddPCR positivity (>2 positive droplets, blue). FIG. 5G shows a comparison of NGS and ddPCR ctHPVDNA detection based on DNA input in dilutional experiments showing a minimum 25-fold increase in sensitivity by NGS.MEEI 2024-326-02Quarles 169511.00058FIG. 5H shows comparison of HPV-DeepSeek and ddPCR showing 80-fold increase in sensitivity by HPV-DeepSeek which translated to an increase in lead time of MRD detection of 6 months. Red line is HPV-DeepSeek and Blue line is ddPCR.

[0022] FIGS. 6A-6J show blood-based genomic prognostic feature annotation. FIG.6A shows a bar graph representing the total number of HPV mutation counts for an individual sample analyzed by both HPV-DeepSeek and NCI CHANGeS (gold standard test). Green represents the mutations detected in common between the two assays, blue represents mutations observed only in HPV-DeepSeek (plasma) and yellow only in NCI CHANGeS. Black dots denote the percentage of common mutation in each sample. FIG. 6B provides a Venn diagram showing the relationship between HPV-DeepSeek (plasma) and NCI CHANGeS for the total number of HPV mutations found in the 23 samples run on both assays. FIG. 6C provides a pie chart showing the distribution of high-risk SNPs in 134 patients with HPV16. FIG. 6D shows a schematic lollipop plot which displays the individual high-risk SNPs identified across the HPV16 cases. Lollipop height is number of patients with each mutation. FIG. 6E shows a schematic lollipop plot which represents the different PIK3CA gene mutations identified across 152 HPV+HNSCC cases. Lollipop height indicates the frequency of each given mutation.Asterisks denote hotspot APOB EC mutations. FIG. 6F shows breakpoint counts (y-axis), ordered by frequency in individual HPV+HNSCCs cases (x-axis). Color gradient denotes the breakpoint cluster size. FIG. 6G shows breakpoints (circles=231) identified in 52 / 152 HPV+HNSCC cases uniquely mapped to HPV genotypes (color coded). Within the viral genome, there was a statistically significant difference (P<0.05) between the expected and observed number of breakpoints for El, E2, E4, LI, L2 and URR. FIG. 6H shows breakpoints clustered within 500-kb windows (upstream and downstream) uniquely mapped to the human genome (x-axis, n = 231); y-axis denotes breakpoint counts. Color denotes the breakpoint cluster size. For the human genome, there was a statistically significant difference (P<0.05) between the expected and observed number of breakpoints for chromosome 9, 12, X, and Y. FIG. 61 provides a histogram showing the frequency of breakpoints in cancer vs. non-cancer genes showing a statistically increased number of mutations in cancer genes (p=0.004). FIG. 6J provides a coverage plot showing the alignment of 134 HPV16 HNSCC cases to identify preferential genome region loss.MEEI 2024-326-02Quarles 169511.00058The x-axis represents the HPV16 genome while the y-axis represents the number of cases that had reads spanning each genome coordinate.

[0023] FIGS. 7A-7D show Venn diagram showing the common and unique HPV SNPs of 23 patients for HPV-DeepSeek on paired tissue-plasma samples. (FIG. 7B) Bar graph showing the output of (FIG. 7A) for individual patients. Common mutations are represented in deep green, while the black dot on top of the bars indicates the percentage of mutations detected in both the assays, demonstrating that samples with a higher mutational load also showed higher concordance rates. (FIG. 7C) Venn diagram showing the results of the 23 tissue samples run across HPV-DeepSeek and NCI CHANGeS, (FIG. 7D). Bar graph showing the output of (FIG.7C) for individual patients.

[0024] FIGS. 8A-8C show a series of scatter plot graphs show a correlation between the number of unique HPV reads and the T stage (Tumor stage) (FIG. 8A), N stage (Node stage) (FIG. 8B), and Overall stage (FIG. 8C), respectively. The Spearman coefficient was used to calculate the correlation coefficient. The values are denoted on top of each individual graph.

[0025] FIG. 9 shows a relationship between high-risk features and cancer metrics. Left: Overall cancer stage in patients with <1 vs. >2 high-risk features (Wilcoxon test, p=0.018). Right: Log-transformed ctHPVDNA reads in patients with <1 vs. >2 high-risk features (Wilcoxon test, p=0.005). High-risk features include high-risk SNPs, PIK3C A mutation, and viral integration status. Multiple high-risk features correlated with higher cancer stage and increased ctHPVDNA reads.

[0026] FIGS. 10A-10E show HPV-DeepSeek fragmentomics analysis. FIG. 10A shows a comparison of the mean fragment length of 152 HPV+ HNSCC cases from ctHPVDNA (red) and non-HPV cfDNA (turquoise), and in 100 healthy controls from cfDNA (pink). The x-axis represents the length of cfDNA reads in base pair (bp), while the y-axis indicates the normalized frequency of read length. FIGS. 10B-10D provide violin plots showing differences in mean (FIG. 10B), standard deviation (FIG. 10C), and skewness (FIG. 10D) of average read lengths (p <0.001 for all comparisons). FIG. 6E provides bar plots showing total number of reads, total genome coverage, and HPV mutation count in 152 HPV+HNSCC cases. The mosaicMEEI 2024-326-02Quarles 169511.00058plot in the lower panel marks the cases that have high-risk SNPs, PIK3CA mutation, and HPV integration events. HPV genotypes are color-coded.

[0027] FIG. 11 shows a line graph to compare the fragmented patterns of cfDNA and ctHPVDNA across 152 HPV+HNSCC patients. The turquoise line represents the mean read length (in base pairs) of cfDNA, while the red line depicts the mean read length of ctHPVDNA from the same HPV+HNSCC patient. The grey line shows the ratio of ctHPVDNA to cfDNA read length for each sample. Each point on the x-axis corresponds to a unique patient sample. The y-axis on the left side quantifies the mean read length of cfDNA and ctHPVDNA, whereas the y-axis on the right provides the scale for the read length ratio.

[0028] FIGS. 12A-12C shows a bar plot showing the permutation importance of blood-based genomic features in predicting the T stage (FIG. 12A), N stage (FIG. 12B), and overall stage (FIG. 12C) using LogisticRegression classifier. The importance values on the x-axis represent the decrease in model performance after shuffling the feature, thereby highlighting its importance in prediction.

[0029] FIGS. 13A-13F show screening and early detection of HPV+OPSCC. FIG.13 A shows a histogram demonstrating blood sample collection timepoint in years prior to diagnosis for 28 HPV+OPSCC patients. FIG. 13B shows a histogram showing results of HPV-DeepSeek in 28 HPV+OPSCC patient samples demonstrating 22 / 28 (79%) samples screening positive. Samples in green are positive for HPV+OPSCC based on the pre-set cutoffs. Samples in red are negative. FIG. 13C shows HPV-DeepSeek results of 56 samples (28 HPV+OPSCC, green color; 28 age- and gender- matched controls, red color). The colors of green correspond to the time of collection, with light green representing samples collected closer to diagnosis and the dark green representing samples collected further from diagnosis. Lighter-green samples tend to have higher numbers of reads and higher coverage and darker-green samples tend to have lower numbers of reads and lower coverage supporting increasing ease of detection closer to time of diagnosis. FIG. 13D provides a histogram showing results of HPV-DeepSeek with machine learning demonstrating 27 / 28 (96%) samples screening positive. FIG. 13E provides a bar graph showing the comparison between HPV-DeepSeek and HPV-DeepSeek with machine learning. Data is divided in tertiles based on collection time point. The deep green color represents theMEEI 2024-326-02Quarles 169511.00058percentage of cases determined positive by HPV-DeepSeek. The light green color represents the cases detected with the addition of machine learning, showing improved performance with machine learning for samples further from the time of diagnosis. FIG. 13F provides histograms showing a comparison of HPV DeepSeek 1.0 (without machine learning) and HPV DeepSeek 2.0 (with machine learning) in as combined cohort of samples from two screening biobanks.

[0030] FIG. 14 shows logistic regression analysis predicting the sensitivity of HPV- DeepSeek and HPV Serology tests over time prior to diagnosis. Sensitivity is modeled as a function of years prior to diagnosis, with separate curves for HPV-DeepSeek (solid blue line) and HPV Serology (dashed green line). The x-axis represents years prior to diagnosis, while the y-axis displays predicted sensitivity. The model accounts for the interaction between test type and time. The figure shows the comparative sensitivity of both diagnostic tests.

[0031] FIG. 15 shows sliding window sensitivity analysis for HPV-DeepSeek and HPV Serology tests over time prior to diagnosis. Sensitivity is measured using the area under the curve (AUC) from receiver operating characteristic (ROC) analysis within 2-year sliding windows, stepped every 0.5 years. The solid blue curve represents the smoothed AUC for HPV-DeepSeek, while the dashed green curve represents HPV Serology. Spline interpolation is applied to smooth the AUC values across valid windows. The x-axis represents years prior to diagnosis, and the y-axis shows the corresponding AUC, reflecting the tests' sensitivity over time.

[0032] FIGS. 16A-16C show a comparative visualization of model feature importance. Bar plots, each corresponding to a different machine learning model: NaiveBayes (FIG. 16A), RandomF orest (FIG. 16B), and AdaBoost (FIG. 16C). Each plot illustrates the relative importance of five distinct features used by the respective models to perform classification tasks. The features are listed on the y-axis in descending order of importance. The x-axis represents the variable importance, indicating how much each feature contributes to the model's predictive power. This comparative analysis aids in understanding the behavior of different models and the significance of each feature in the classification process.

[0033] FIGS. 17A-17C show viral genome molecular fingerprinting and longitudinal cancer monitoring. FIG. 17A shows a phylogenetic tree of the 28 HPV+OPSCC samples basedMEEI 2024-326-02Quarles 169511.00058on the genotype, lineage and sublineage detected in plasma (left) and molecular fingerprinting heatmap (right) for samples sharing the same HPV16 sublineage. SNVs for each sample are shown in red, demonstrating that each virus is unique. FIG. 17B shows a pairwise correlation heatmap for the 15 samples that had FFPE tumor tissue blocks from diagnosis available, showing that plasma-tissue pairs correlate within a pair and not across pairs. The highest correlation of 1 is represented in red while no correlation or 0 is represented in white. FIG. 17C shows longitudinal monitoring of four HPV+OPSCC patients from screening time point to diagnosis, treatment and post-treatment monitoring. Vertical dotted line represents time of diagnosis.Patients 8 and 21 were treated with surgery. Patients 1 and 3 were treated with chemoradiotherapy. In patients 1, 8, and 21 ctHPVDNAis cleared following treatment and remained zero during monitoring. In patient 3, ctHPVDNA was detected 20 months before diagnosis. ctHPVDNA levels were monitored weekly during chemoradiotherapy treatment with decreasing levels, but no clearance. Following conclusion of treatment, ctHPVDNA levels began increasing. The patient was then found to have a second primary HPV malignancy (asterisk) for which they underwent surgery followed by chemoradiotherapy. ctHPVDNA cleared after this treatment but re-elevated, indicating recurrence, which was detected by cross- sectional imaging two months later.

[0034] FIG. 18 shows mutation profile of HPV16 genomes across different patients. Mutation analysis of HPV16 genomes from four patients (13, 17, 20, and 22). The top panel displays the distribution of mutations across the viral genome for each patient, represented by red blocks. The analysis was carried out without filtering threshold (AF> 0.5 & Depth> 2 reads) for the identification of all single nucleotide variations (SNVs), providing a complete molecular fingerprint for each viral sample. The bottom panel maps the HPV16 genome, with colored blocks representing different coding regions: E6, E7, El, E2, E4, E5, L2, and LI, and the untranslated region (URR). Arrows between the panels indicate the corresponding locations of mutations on the genome map. This unrestricted mutation plot aids in understanding the unique mutational landscape of HPV in each sample.

[0035] FIG. 19 shows analysis of single nucleotide variants (SNVs) in matched blood and tumor samples. Heatmap demonstrating the analysis of Single Nucleotide Variants (SNVs) in matched blood and tumor samples showing the HPV genome similarity betweenMEEI 2024-326-02Quarles 169511.00058plasma and tumor tissues. Each row corresponds to an individual patient sample, while each column represents a specific mutation. The presence of a mutation is indicated by red. The samples are labeled as P (Plasma) and T (Tumor) and the corresponding patient number. The bar graph on the right side represents the mutation count of each viral genome. The mutations shared between paired plasma and tumor are shaded in orange while the green bar represents the mutations that appear only in one sample.

[0036] FIG. 20 shows longitudinal ctHPVDNA monitoring. Changes in circulating tumor HPV DNA (ctHPVDNA) over time for seven patients. Each plot represents one patient's ctHPVDNA trajectory as a function of time, with the sample collection timepoint in months on the x-axis and the ctHPVDNA loads on the y-axis, plotted on a logarithmic scale. The vertical dashed lines represent the date of diagnosis. The data points are plotted as red dots, and the connecting lines indicate the trend over the observed period.

[0037] FIGS. 21 A-21C show a comparison of ctHPVDNA to other blood-based HPV cancer detection approaches. FIG. 21A shows a histogram representing the 28 HPV+OPSCC samples ordered by sample collection time in years prior to diagnosis. Below, features supportive of cancer detection. Samples positive for a given feature are in green, samples negative for a feature are in red, and untested samples are in beige. FIG. 2 IB shows a Venn diagram for HPV serology and ctHPVDNA results showing 20 / 26 patients had HPV oncoprotein antibodies detected of which 17 also had ctHPVDNA detected. FIG. 21C shows a bar plot indicating the number of samples with 0, 1, 2, or 3 cancer diagnosis-supporting features present. 25 / 28 cases had at least one cancer diagnosis supporting feature and 18 / 28 had at least two features.

[0038] FIG. 22 shows schematic representation of HPV-Human genome integration events in two patients. Schematic representation of HPV-human genome integration events detected in two patients using a threshold of MAPQ 40 and filtering for at least two supporting reads. The top bar represents the human genome with numbered chromosomes 1 through 22. The bottom line shows the HPV16 genome with numerical positions along its length from 1 to 7906. The blue line (LI) indicates an integration event between the E2 and E4 region of the HPVMEEI 2024-326-02Quarles 169511.00058genome and Chromosome 18 of the human genome. The pink line (L2) represents a second integration event between the E5 region of the HPV genome and Chromosome 19.

[0039] FIGS. 23A-23C show motivation for using HPV-DeepSeek to monitor molecular residual disease. FIG. 23 A shows a schematic of earlier detection of MRD using HPV-DeepSeek vs dPCR. FIG. 23B shows study design of MRD study. FIG. 23C shows scatter plots of HPV reads over the course of disease for cases with adjuvant therapy, cases without adjuvant therapy, and recurrent cases.

[0040] FIGS. 24A-24F show patient characteristics and ctHPVDNA detection at diagnosis. (FIG. 24A) HPV genotype detected. (FIG. 24B) Correlation between HPV-DeepSeek (y- axis) and ddPCR (x-axis) in 92 cases run on both assays prior to treatment. Red dots represent cases with recurrence at a later date and the shape of each dot represents an HPV genotype. The shaded area represents the HPV-DeepSeek (orange) and ddPCR (green) negative (FIG. 24C) Scatter plot showing the linear correlation between HPV DeepSeek and ddPCR. Blue dots represent individual cases, and the red line indicates the trend line. (FIG. 24D) Adjuvant therapy received. (FIG. 24E) Sites of recurrence for eight patients who experienced persistence or recurrence. (FIG. 24F) Dynamics of ctHPVDNA after surgery. Cases divided into four groups based on trends in postoperative ctHPVDNA reads: 1. Consistent positive, 2. Negative after adjuvant therapy, 3. Gradual decrease, 4. Rapid decrease). The line and shades represent the median ctHPVDNA level and interquartile range, respectively.

[0041] FIG. 25 shows average ctHPVDNA clearance by group.

[0042] FIGS. 26A-26D show validation of the prognostic value of HPV-DeepSeek. (FIGS. 26A and 26B) Kaplan-Meier curve of two-year disease-free survival (DFS) and overall survival (OS) based on ctHPVDNA status during the MRD-PS. The X-axis shows months after surgery, and the Y-axis shows the survival rate. (FIGS. 26C and 26D) DFS and OS is based on MRD status in the MRD-TC window. The X-axis shows the month after treatment completion. The red line represents ctHPVDNA positive, and the blue line represents negative. The hazard ratio and 95% confidence interval are shown in each graph. The Log-rank test calculated the P value. The number at risk every 6 months is displayed at the bottom of the graph.MEEI 2024-326-02Quarles 169511.00058

[0043] FIGS. 27A-27D show the prognostic value of single-time ctHPVDNA.(FIGS. 27A and 27B) Kaplan-Meier curve of two-year disease-free survival (DFS) and overall survival (OS) based on single-time ctHPVDNA status during the MRD-PS window. The X-axis shows months after surgery, and the Y-axis shows the survival rate. (FIG. 27C and 27D) DFS and OS based on single-time MRD detection in MRD-TC window. The X-axis shows the month after treatment completion. The hazard ratio and 95% confidence interval are shown in each graph. The Log-rank test calculated the P value. The number at risk every 6 months is displayed at the bottom of the graph.

[0044] FIGS. 28A-28B show the prognostic value of ctHPVDNA immediately after surgery. 2 -year DFS (FIG. 28A) and OS (FIG. 28B) based on ctHPVDNA in MRD timeframe immediately after surgery (MRD-early) using HPV-DeepSeek. The hazard ratio and 95% confidence interval are shown in each graph. The Log-rank test calculated the P value. The number at risk every 6 months is displayed at the bottom of the graph.

[0045] FIGS. 29A-29B show predicting recurrence using multivariate regression. (FIG. 29A) Forest plot showing multivariate regression analysis of ctDNA during the MRD window (PS), with adjusted hazard ratios (HR) and 95% confidence intervals (CI) for clinical factors like ctDNA status, age, sex, smoking, ENE, margins, pT, and pN stages. (FIG. 29B) Forest plot of multivariate regression for ctDNA during the MRD window (TC), presenting adjusted HRs and Cis for the same variables. Reference groups and significant HRs are highlighted to show their association with recurrence risk.

[0046] FIGS. 30A-30C show early detection of recurrent disease by detecting ctHPVDNA using HPV-DeepSeek and ddPCR. (FIG. 30A, FIG. 30B) Bar graph showing the sensitivity of HPV-DeepSeek and ddPCR assays for detecting residual disease in MRD status after surgery (MRD-PS) and treatment completion (MRD-TC). The y-axis shows sensitivity (%). (FIG. 30C) Bar graph showing the mean lead time to clinical diagnosis of recurrence after surgery by HPV-DeepSeek and ddPCR assays.

[0047] FIGS. 31 A-31G show longitudinal ctHPVDNA Monitoring of recurrent cases with significant improvement in sensitivity with HPV-DeepSeek over ddPCR. Line graphs showing the changes in ctHPVDNA levels over time for recurrent cases. Each plot representsMEEI 2024-326-02Quarles 169511.00058one patient's ctHPVDNA trajectory as a function of time, with the sample collection timepoint in days after surgery on the x-axis and the ctHPVDNA levels on the y-axis (logarithmic scale). The green line represents the standardized ctHPVDNA value (*copies / ml) for HPV-DeepSeek, and the blue line for ddPCR. The markers are color-coded for positive and negative time points.

[0048] FIGS. 32A-32B show longitudinal changes in ctHPVDNA levels and HPV SNPs in recurrent cases demonstrating the ability to differentiate recurrences and second primary tumors by leveraging viral WGS. Line graphs showing the changes in ctHPVDNA levels over time for a recurrent case (FIG. 32A) and 2nd primary case (FIG. 32B). Each plot represents one patient's ctHPVDNA trajectory as a function of time, with the sample collection timepoint in days after surgery on the x-axis and the ctHPVDNA levels on the y-axis (logarithmic scale). The green line represents the standardized ctHPVDNA value (*copies / ml) for HPV-DeepSeek, and the blue line for ddPCR. The markers are color-coded for positive and negative time points. The allele frequency (AF) change of SNPs is shown above graph. The vertical axis shows the position of each SNP, and the horizontal axis shows the time course. The dotted lines connect each time point with the AF at that time. Changes in allele frequency are used to determine secondary primary tumor or recurrent tumor.

[0049] FIGS. 33A-33D show predicting recurrence and extranodal extension using ctDNA with machine learning and multivariate regression. (FIG. 33 A) ROC curves for three machine learning models (Balanced Random Forest, Decision Trees, XGBoost) predicting extranodal extension from diagnostic blood samples. (FIG. 33B) Feature importance rankings of pre-diagnostic variables, including ctDNA metrics, derived from the machine learning models. (FIG. 33C) ROC curves for the same models evaluating performance in predicting cancer recurrence. (FIG. 33D) Feature importance plots highlighting the top predictors for recurrence in each machine learning model.

[0050] FIG. 34 shows a flow chart of a method to classify a subject sample.

[0051] FIG. 35 shows a schematic diagram of an exemplary computer system.MEEI 2024-326-02Quarles 169511.00058DETAILED DESCRIPTION

[0052] In accordance with some embodiments of the disclosed subject matter, mechanisms (which can include, for example, systems, computer readable media, and methods) for determining a cancer score are provided. In some embodiments, the subject sample is a sample of circulating tumor (ct)DNA, and the sample is classified based features detected in the sequences of the ctDNA. The classification can be used assist in diagnosis or prognostication of a subject, particularly a subject who may have a viral based cancer, particularly HPV related cancers.

[0053] All cells release short fragments of double- stranded DNA into the circulation called cell free (cf)DNA and when released by a cancer cell, circulating tumor (ct)DNA. Advancements in ctDNA analytic approaches have led to the implementation of ctDNA liquid biopsies for cancer molecular profiling and monitoring, including FDA-approved tests. ctDNA-based testing has several advantages over tissue-based approaches, including that it is minimally invasive, allows repeat sampling over time, and can be site agnostic, identifying cancers which may not be clinically apparent (below imaging resolution) or targetable (in regions which cannot be biopsied safely. However, use of ctDNA has been limited by ongoing challenges associated with sensitivity, particularly in settings with low tumor burden, such as screening. FIG. 1 A shows the evolution of tumor burden over time. FIG. IB shows the relationship between the level of disease burden and ability to diagnose a subject for several diagnostic methods described herein.

[0054] Most ctDNA approaches focus on detection of point mutations, which can have low allele fractions, making reliable detection challenging, especially without knowing the target(s) a priori. To overcome this limitation, currently, the most sensitive approaches are “tissue informed”, meaning the mutations are discerned in the tumor tissue from biopsy and then targeted in the blood, an approach that is invasive and is not feasible in the screening setting. The field of ctDNA-based early cancer detection, however, is rapidly progressing with numerous Multi-Cancer Early Detection (MCED) approaches now available in research and clinical settings. These approaches typically focus on pan-cancer detection utilizing methylation or fragmentomics. Thus, while these approaches have the advantage of detecting multiple cancerMEEI 2024-326-02Quarles 169511.00058types without needing to be customized to each patient, sensitivity is more limited than bespoke tissue-informed approaches. For example, the sensitivity of the Grail Galleri test is reported at 29%. While the future of cancer screening, detection, and monitoring is likely to be ctDNA based, current approaches lack the diagnostic accuracy needed for broad clinical application.

[0055] In addition to somatic ctDNA, virally-associated cancers release circulating tumor viral DNA. The most common virally driven cancer is HPV-associated cancer (circulating tumor HPV DNA (ctHPVDNA)). ctHPVDNA is more easily detected than mutated somatic ctDNA as ctHPVDNA has low homology to human cell free (cf)DNA and the HPV genome has conserved regions which can be targeted across patients, leading to improved sensitivity at lower cost. Singleplex and multiplex digital PCR (dPCR, which, in some embodiments, includes droplet digital PCR (ddPCR)) assays for detection of ctHPVDNA from the most common HPV genotype have been developed and applied. Work utilizing this assay, and others, has demonstrated that most, but not all patients with HPV+HNSCC have ctHPVDNA detectable in the blood at the time of presentation (Sensitivity 88-98%). When compared with standard tissue biopsy -based diagnosis, ctHPVDNA dPCR has improved Diagnostic Accuracy (Youden index 0.937 vs.0.707, p=0.0006), is 36% less expensive and has shorter time to diagnosis. In the treatment setting, it has been shown that ctHPVDNA is an accurate real-time biomarker of treatment response to both surgery and chemoradiotherapy and can detect recurrence earlier than standard of care imaging. An unrelated, commercial dPCR assay is commercially available (NAVERIS) and is covered by Medicare. Taken together, there is strong evidence to support ctHPVDNA as a tool for diagnosis, treatment monitoring, and recurrence detection.

[0056] While dPCR has high fidelity, sensitivity, and specificity at timepoints with gross cancer burden, such as at diagnosis, it has several critical limitations. First, dPCR is only able to detect a limited number of a priori DNA targets. While -90-95% of HPV+HNSCC are caused by four HPV genotypes, 5-10% of cases are caused by numerous other rare genotypes, which are missed by HPV dPCR panels even when multiplexing is utilized, inherently limiting the sensitivity. Second, HPV dPCR probes only target one to two specific ~150bp DNA fragment representative of the HPV genome of interest (the size of a fragment of ctDNA). Therefore, only -2% of the ctHPVDNA generated is captured (HPV genome is -8,000 bases), significantly limiting sensitivity in early detection settings when total ctHPVDNA is low. Similarly, theMEEI 2024-326-02Quarles 169511.00058inability of HPV dPCR to generate genome wide coverage means it cannot be used for interrogating prognostic features.

[0057] HPV antibodies can be used as biomarkers for identifying those at risk of developing HPV-related cancers. Individuals exposed to HPV develop Abs to the HPV capsid protein LI, which are common in the population. On the contrary, development of Abs to HPV oncoproteins, such as E6, are a biomarker for identifying those at highest risk for developing HPV+OPSCC26-28. HPV16 E6 Abs are present in >90% of HPV+OPSCC patients at diagnosis, are rare among individuals without cancer in the general population (specificity >99%) and develop up to 28 years prior to cancer diagnosis. At age 50, E6-seropositive men have a 50% cumulative risk of developing HPV+OPSCC and a five-year risk of 7.3%, which falls well within the range of other cancers, such as breast and lung, for which screening is currently recommended. However, the long and variable lag time between HPV16 E6 seroconversion and HPV+OPSCC development, and the non-linear nature of the biomarker, means E6 seropositivity has limited real-world clinical utility. Further, the performance characteristics of E6 Abs in anogenital HPV cancer are significantly worse, with only 29% of patients with anal cancer having E6 Abs at diagnosis, for example. Thus, while there is strong data to support HPVE6 Abs as a potential screening biomarker, more comprehensive combinatorial profiling approaches are needed to create a clinically useful test that could indicate with high specificity when an invasive cancer has developed.

[0058] Individuals exposed to HPV develop Abs to the HPV capsid protein LI, which are common in the population. On the contrary, development of Abs to HPV oncoproteins, such as E6, are a biomarker for identifying those at highest risk for developing HPV+OPSCC26-28. HPV16 E6 Abs are present in >90% of HPV+OPSCC patients at diagnosis, are rare among individuals without cancer in the general population (specificity >99%) and develop up to 28 years prior to cancer diagnosis. Thus, HPV antibodies can be used to inform cancer diagnosis, particularly when combined with other features.

[0059] Provided herein are methods of determining a cancer score based on receiving a sample including circulating tumor DNA (ctDNA) from a subject and providing the sample to an analysis pipeline, which detects and synthesizes a plurality of features. The ctDNAMEEI 2024-326-02Quarles 169511.00058may be viral related DNA. In some of the examples, the viral DNA is HPV-related DNA.However, the methods are not meant to be limited to HPV; any viral-related cancers can benefit from use of the methods described herein.

[0060] An optimized and validated multi-feature HPV whole genome sequencing liquid biopsy, referred to as HPV-DeepSeek, is described herein. FIGS. 2A-2D show a schematic overview of HPV-DeepSeek. HPV-DeepSeek can be used to overcome limitations in the scope and sensitivity of existing HPV liquid biopsy approaches. This approach uses custom next generation sequencing libraries to annotate different features in the blood. These features may include HPV viral load, HPV genome coverage, HPV genotype, lineage, and sublineage, normalized HPV genomes / human genomes per cell, viral integration events, select human genes and gene mutations, cell free DNA fragment size features, viral single nucleotide polymorphisms known to increase risk of cancer recurrence, and human germline SNPs known to increase risk of HPV cancers and risk of cancer recurrence. Some or all of the features can be used for cancer prognostication, and some or all of the features can be used for diagnosis. FIG. 3A shows how a combination of features can be used to calculate a cancer score, or a cancer classification and FIG. 5A shows in a schematic how the features are analyzed alone, and by machine learning.

[0061] Viral DNA may be sequenced using single strand consensus sequencing or duplex screening. Duplex screening suppresses noise and increases specificity. The analysis pipeline can be run with either sequencing method (see FIG. 3B for an overview of duplex sequencing).

[0062] Other features may include HPV antibodies. HPV antibodies may be collected and analyzed using HPV serology. In some embodiments, one analysis pipeline is used to analyze sequencing data, and another analysis pipeline is used to analyze a combination of sequencing data and HPV antibody data. Additional features include subject demographic information, such as age, sex, tumor stage, tumor size, tumor location, number of sexual partners, HIV status, or smoking status. The same or additional analysis pipelines may be used to incorporate subject demographic information. In some embodiments, the methods and systems described herein may be used for high risk patients, such as patients living with HIV.MEEI 2024-326-02Quarles 169511.00058

[0063] In some embodiments, the analysis pipeline includes one or more trained network. Different trained networks can be used based on whether or not HPV antibody data is included (e.g., one trained network is used without HPV antibody data, a separate trained network is used with HPV antibody data). In the case where HPV antibody data is used, the analysis pipeline and / or a trained network may be used to integrate the sequencing and HPV antibody information.

[0064] The HPV genotype can include identifying the detected ctHPVDNA as one of more than 13 HPV genotypes. In some embodiments, the ctHPVDNA is identifying as one of more than 40 HPV genotypes. In some embodiments, the HPV genotype is one of HPV genotypes 16, 18, 33, 35, or 45.

[0065] In some embodiments, the viral integration events include identifying viral genome breakpoints and human genome breakpoints. The select gene mutations may be mutations that are known to be associated with cancer, or mutations that are known to be deleterious to a subject. In some embodiments, the select gene mutations may be PIK3CA mutations.

[0066] In some embodiments, only a subset of the possible features is used for analysis. In some embodiments, a different number and subset of features can be used for diagnosis or prognostication. In some embodiments, the same number and subset of features are used for both diagnosis and prognostication. The number of unique HPV molecular reads, as determined by unique molecular indices and percentage of the HPV genome captures, is particularly important for sensitivity and specificity for diagnosis and detecting MRD. HPV mutation count is particularly important for prognostication.

[0067] Custom informatics pipelines were developed for annotating each of these features and machine learning was used to integrate the features to increase diagnostic and prognostic power through generation of a multi-feature HPV cancer score to both serve as an ultra-sensitive diagnostic biomarker of HPV-associated cancer and prognosticate cancer outcomes.

[0068] In some embodiments, machine learning approaches are incorporated into the analysis pipeline. Methods that include machine learning may be referred to as HPV-MEEI 2024-326-02Quarles 169511.00058DeepSeek 2.0. HPV-DeepSeek 1.0 refers to an analysis pipeline that does not include any machine learning approaches.

[0069] In some embodiments, the trained network is a multi-label, multi-class network. In some embodiments, the trained network includes more than one trained network. The architecture of the trained network may be tree-based (e.g., random forest, extra trees, decision trees), gradient boosting (e.g., XGBoost, LightGBM), linear (e.g., logistic regression, ridge, lasso, elastic net), kernel-based (e.g., support vector machine), instance-based (k-nearest neighbors), or probabilistic (Naive Bayes). FIG. 2D shows a comparison of different network architectures and the resulting sensitivity and accuracy.

[0070] In some embodiments, the network is trained on labeled subject data. The labeled subject data may include control sample, which are samples collected from subjects who have not been diagnosed with a cancer, and test samples, which are samples collected from subjects who have been diagnosed with a cancer. In some embodiments, the network is trained on only sequencing data. In some embodiments, the network is trained on sequencing data and antibody data. In some embodiments, the network is also trained on data annotated to include patient demographic information.

[0071] In head-to-head testing against ctHPVDNA dPCR, HPV-DeepSeek is 25-80-fold more sensitive, and has permitted, for the first time, accurate HPV cancer screening. HPV-DeepSeek 2.0 can be used to accurately define additional prognostic clinical features. Modeling was utilized to show specific populations that could use this test to screen for HPV cancers (see FIGS. 4A-4D). FIG. 4E shows a direct comparison of HPV-DeepSeek, ddPCR, and Naveris NavDx test to detect screening blood samples from asymptomatic people who later developed HPV+ head and neck cancer.

[0072] In some embodiments, the methods described herein can be used to diagnose a subject with a cancer. The diagnosis can indicate whether a patient has cancer or is likely to develop cancer in the future. The subject may be diagnosed with a specific type of cancer. The type of cancer may be HPV-related. HPV-related cancers include cervical cancer, anal cancer, oropharyngeal cancer, penile cancer, vaginal cancer, vulvar cancer, lacrimal sac cancer, conjunctival cancer, sinonasal cancer and nasopharyngeal cancer. In some exemplaryMEEI 2024-326-02Quarles 169511.00058embodiments, the patient is diagnosed as having head and neck cancer. The methods described herein are tissue agnostic. Therefore, the same test can be used to screen for any or all of these cancers.

[0073] The methods and systems described herein may be used for screening (e.g., asymptomatic testing for cancer), diagnosis (e.g., determining whether a patient has cancer), detecting molecular residual disease (MRD) (e.g., detecting MRD after treatment to determine if the treatment was successful), and surveillance (e.g., monitoring a patient for recurrence after treatment). The methods and systems may be used longitudinally (e.g., the same test is used for multiple phases). This is a significant advantage over other methods, which require specialized tests for different phases.

[0074] In some embodiments, the subject may not experience any symptoms indicative of having cancer at time of screening. In some embodiments, the subject may be experiencing symptoms indicative of having cancer (e g., the method is used to confirm a suspected diagnosis). For instance, symptoms for head and neck cancer include throat pain, difficulty swallowing, sensation of something stuck in one’s throat, bleeding, and / or an enlarged lymph node. HPV-DeepSeek enables screening that allows early detection with lead times greater than four years before cancer diagnosis. A screened cancer may be treated with immunotherapy or a personalized cancer vaccine.

[0075] In embodiments in which the methods and systems described herein are used to diagnose a subject, the diagnosis may have a sensitivity greater than 98% at a specificity greater than 98%. In some embodiments, HPV-DeepSeek is able to detect a disease more than three and a half years (~ 40 months or more) before disease expression (e.g., before a patient has symptoms of the cancer).

[0076] In some embodiments, the methods described herein can be used to predict molecular residual disease (MRD). MRD may persist after completion of cancer therapy. The subject then requires further monitoring or additional treatment. Repeated tests using the methods described herein can help identify if and when a subject needs to undergo additional treatment. HPV-DeepSeek offers greater sensitivity of MRD than existing approaches such as dPCR.MEEI 2024-326-02Quarles 169511.00058

[0077] The methods described herein are able to increase the average lead time to detect MRD. Lead time refers to the amount of time before exhibiting symptoms a patient can be diagnosed with cancer. Greater lead time to detecting MRD can help a patient receive early treatment and minimize risks. In some embodiments, the lead time is greater than one year, greater than two years, or greater than three years,

[0078] In some embodiments, the method may further include making a clinical decision based on the result of the cancer score. For instance, after screening, the clinical decision may be a treatment option, ordering of additional tests, or ordering changes in the frequency of visiting or testing. Treatment options for a screened cancer include but are not limited to immunotherapies, chemotherapies, personalized cancer vaccines, non-personalized vaccines or surgery. When detecting MRD or during surveillance, treatment options include but are not limited to therapeutic vaccines, immunotherapies, chemotherapies, surgery, or radiation.

[0079] The methods described herein can also assist in determining a prognosis for a subject. For instance, the methods described herein can help identify what stage of cancer a patient has using output from HPV-DeepSeek. Additionally, the methods can help determine personalized treatment for a subject (e.g., 1. the need for radiation or chemotherapy after surgery, 2. the dose of radiation to be delivered after surgery, 3. The dose of radiation to be used for definitive treatment, 4. The need for additional treatment after definitive treatment or adjuvant treatment, such as the addition of immunotherapy. FIG. 4F shows the current post-screening test work ups, and a proposed treatment paradigm made possible with HPV-DeepSeek.

[0080] A report can be provided to a user, where the report provides a cancer score or classification. A report can be used to aid a practitioner in determining the presence of a cancer in an asymptomatic person, a diagnosis of cancer, determining a prognosis, or identifying potential personalized therapeutic plans.

[0081] The methods and systems described herein can be used to generate a multifeature HPV cancer score from a blood sample. The cancer score may incorporate features from both the viral and human genomes, and the cancer score may be both diagnostic and prognostic. This single approach can be used from screening through diagnosis and treatment monitoring. This score permits significantly earlier detection of HPV-associated cancers than current liquidMEEI 2024-326-02Quarles 169511.00058biopsy approaches and standard of care, permitting earlier treatment. Earlier treatment can lead to better outcomes (e.g., better chance of survival) for a subject. The score also permits prognostication of cancer outcomes and can aid in the development of personalized care.

[0082] EXAMPLES

[0083] The following are non-limiting, illustrative examples to which the disclosed systems and methods may be applied. The examples should not be seen as limiting.

[0084] Example 1: Direct comparison of alternative blood-based approaches for early detection and diagnosis of HPV-associated head and neck cancers

[0085] Introduction

[0086] The incidence of human papillomavirus (HPV)-associated head and neck squamous cell carcinomas (HPV+HNSCC) is rising in the United States (US). Strategies for early detection could have meaningful public health benefits by diagnosing patients with subclinical disease, permitting the use of less morbid treatments for earlier stage disease and potentially improving survival as well. HPV+HNSCC can be detected with multiple blood-based analytes, both at and before clinical diagnosis. The two most extensively studied biomarkers are HPV early protein antibodies (HPV Ab) and circulating tumor HPV DNA (ctHPVDNA).

[0087] The first, HPV Ab has been shown to be a promising biomarker of HPV+HNSCC. HPV E6 Ab are present in -83% of HPV+HNSCC patients at diagnosis, -5% of HPV-independent HNSCC patients at diagnosis (-95% tumor specificity) and are rare among individuals without cancer in the general population (population specificity -99%).Combinatorial approaches including multiple early antibodies (El, E2, E5, E6, E7), and bacterially expressed E6 antigen-based approaches show improved sensitivity (-90%) with no change in specificity. The second, ctHPVDNA, is based on detection of short double-stranded HPV genome fragments in the circulation. ctHPVDNA is detectable in -89% of HPV+HNSCC patients at diagnosis with the most common detection approach, digital PCR (dPCR), and has a specificity of -97% in the general population. While dPCR ctHPVDNA detection has been shown to outperform standard of care diagnostic approaches, and is being used clinically, these approaches remain too insensitive for broad clinical deployment as stand-alone (tissue-biopsy-MEEI 2024-326-02Quarles 169511.00058free) diagnostic tests. Sensitivity has been found to increase with next generation sequencing (NGS)-based approaches (>91%) while maintaining specificity. Additionally, neither HPV Ab nor existing ctHPVDNA detection assays function as prognostic biomarkers at the time of diagnosis, further limiting their clinical utility. Known genomic prognostic features in HPV+HNSCC include the presence of PIK3CA mutations, viral genome integration events, and specific HPV16 single nucleotide polymorphisms (SNPs).

[0088] Critically, both HPV Ab and ctHPVDNA are also detectable prior to HPV+HNSCC diagnosis and thus are being studied as cancer screening biomarkers alone, and in combination. At present, there are no screening tests for HPV+HNSCC. HPV Ab can develop up to 28 years prior to cancer diagnosis. At age 50, HPV16 E6-seropositive men have a 50% cumulative risk of developing HPV+HNSCC and a five-year risk of 7.3%. However, the long and variable lead time between HPV Ab seroconversion and HPV+HNSCC development, and the non-linear nature of the biomarker, means HPV Abs have limited real-world clinical utility. ctHPVDNA is detectable is some patients prior to clinical cancer diagnosis, but sensitivity is limited with current dPCR -based approaches.

[0089] While HPV E6 Abs and ctHPVDNA are highly promising and biologically plausible diagnostic and screening biomarkers for ELPV+HNSCC, significantly more sensitive and comprehensive approaches are needed to generate a single or combinatorial assay with high diagnostic accuracy, prognostic capacity and thus clinical utility for screening and diagnosis. Further, head-to-head comparisons are needed to define operational characteristics of ctHPVDNA and HPV Ab alone, and together, as screening tools.

[0090] To address these limitations, we developed a multi-feature HPV whole genome sequencing (WGS) liquid biopsy, termed EIP V-Deep Seek, for improved low-level ctHPVDNA detection for use in HPV+HNSCC early detection. Here, we conducted a casecontrol cohort study to define performance characteristics of HPV-DeepSeek and compare this WGS-based approach, alone, and in combination, head-to-head with existing blood-based FLPV detection approaches (singleplex dPCR -based ctHPVDNA detection, multiplexed dPCR -based ctHPVDNA detection, multiplexed HPV Ab detection and clinical standard of care tissue biopsy) benched-marked to gold standard HPV+HNSCC tissue diagnosis. We then modeled theMEEI 2024-326-02Quarles 169511.00058operational feasibility of these approaches with their defined performance metrics as screening biomarkers for HPV+HNSCC. Lastly, we examined if machine learning could integrate bloodbased features to predict clinical prognostic features, such as disease stage.

[0091] Methods

[0092] Study design and enrollment: All participants provided written informed consent to a protocol approved by the Institutional Review Board at Dana Farber / Harvard Cancer Center. This study was conducted in compliance with the U.S. Common Rule. Cases: Patients >18 years old with a new or suspected diagnosis of HNSCC were prospectively identified upon presentation to the head and neck cancer clinics at Massachusetts Eye and Ear and Massachusetts General Hospital and consecutively enrolled. All patients underwent standard of care diagnostic workup. Controls: Controls were enrolled 1:1 to cases, and matched to age and sex. Controls were comprised of non-cancer “healthy” population controls and HPV-independent cancer controls, enrolled in a 2:1 fashion, to create a population mix similar to clinical care. Controls were enrolled from the same geographic regions as cases. All blood samples were assigned a study ID and blinded for further analysis.

[0093] Clinical diagnostics: All cancer cases and cancer control patients underwent standard diagnostic workup including practitioners’ choice tissue biopsy, HPV diagnostics (if indicated) per institutional standards and cross-sectional imaging. All cases suspected of being HPV+HNSCC underwent histomorphological assessment and a combination of pl6 immunohistochemistry (IHC), DNAPCR, or RNA chromogenic in situ hybridization (ISH). Cancer control patients who did not undergo clinical HPV testing had HPV testing post-hoc by the same anatomic pathology team to ensure completeness of data for analysis of diagnostic specificity. Diagnostic success of HPV+HNSCC was defined as achieving a histomorphological diagnosis of SCC and correct identification of HPV status (associated or independent) in the first diagnostic attempt (biopsy).

[0094] Blood and tissue collection and processing for HPV detection: 5-10 ml of blood was collected from all participants and processed for cell free DNA (cfDNA) as previously described8. Blood: cfDNA was extracted from l-4ml of frozen plasma as previously described for all cases and controls. Tissue: DNA was extracted from formalin-fixed, paraffin-embeddedMEEI 2024-326-02Quarles 169511.00058(FFPE) tissue blocks after tumor identification and microdissection and sheared prior to sequencing. 20ng of DNA either total cell free from plasma or total DNAfrom tissue was used as input for all downstream analyses.

[0095] ddPCR: Singleplex and multiplex ddPCR were performed, as previously described8. Multiplexed ddPCR was performed using probe mixing for HPV genotypes 16, 18, 33, 35, and 45 to create five distinct clusters for gating which were validated with dilutional experiments in comparison to singleplex ddPCR. Each genotype was targeted by one probe, which have previously been published, which differs from approaches which multiplex multiple probes per genotype. Thresholds for both singleplex and multiplex detection were preset and unchanged from prior reports.

[0096] HPV serology: Multiplexed Luminex-based HPV serology testing was carried out at the German Cancer Research Center (DKFZ) for HPV 16, 33, 35, and 45 using El, E2, E6 and E7 with pre-set thresholds. Serology was considered positive if above the pre-defined threshold and for the correct clinical genotype. Cross-reactivity (incorrect genotype) was considered negative, including for genotypes not included in the panel.

[0097] HPV-DeepSeek: A custom WGS hybrid capture library was designed for 43 HPV genotypes. Libraries were prepared following the KAPA HyperCap cfDNA Workflow vl .0 (Roche) with custom modifications. The steps include: 1. End repair & A-tailing; 2. KAPA Universal UMf adapter ligation; 3. Post-ligation cleanup; 4. Amplification using KAPA UDI primer mixes; 5. Cleanup using KAPA HyperPure beads; 6. Quantification by Qubit and quality check by Tapestation; 7. Preparation of 4-plex DNA sample library pools; 8. Use of customized KAPA HyperCap target enrichment probes for hybridization and amplification of the target libraries; 9. Cleanup of enriched target libraries using KAPA HyperCapture bead kit; 10.Quantification and quality check by Qubit and Tapestation; 11. Pooling and normalization of libraries; 12. Denaturation; and 13. Paired-end Illumina sequencing to ~500x HPV genome coverage per sample.

[0098] HPV-DeepSeek Informatics: Custom informatic pipelines were designed to process and analyze the sequencing data. The initial steps involved adapter trimming and quality filtering, followed by alignment to a custom reference genome encompassing human hg38 andMEEI 2024-326-02Quarles 169511.0005864 HPV reference genomes using BWA-MEM. After identifying Unique Molecular Indexes (UMI) and performing deduplication, HPV genotypes were defined by mapping to the HPV reference genome. The number of unique HPV-mapped UMIs in each sample was used to determine ctHPVDNAread counts. For HPV16 samples with > 50% genome coverage, positionwise base counts were obtained, and a consensus sequence was created by selecting the most prevalent nucleotide at each position. These consensus sequences were fed into the multialignment tool MUSCLE, along with multiple reference builds for HPV16 sublineages. The resulting alignment was fed into PhyML (run using TOPALi v2.5), assuming a transversion model, allowing for invariant sites and specifying a gamma distribution. The resulting phylogenetic tree enabled sublineage assignment of each sample by nearest distance to a sublineage reference genome. Once aligned to the appropriate reference genome, HPV SNPs were called using bcftools, filtered based on AF (> 0.5) and read depth (>4). Human singlenucleotide variants (SNVs) in PIK3CA were called using MuTect vl and annotated by SnpEff, with the cut-off value of AF>=0.04. Mutations calls from plasma samples were validated in matched tumor tissue when available. Viral-human integration events were identified by BWA-MEM. The breakpoints of these events were filtered to ensure chimeric reads were truly composed of a human-genome part and virus-genome part, and at least 4 high-quality reads supporting these events were required. All integration events were manually reviewed to ensure accuracy in IGV Individual breakpoints were then counted and reported as the number of integration events, per patient. We then evaluated whether breakpoints were clustered within a patient, using the methods described by Symer et al., with the addition of using DBSCAN (scikit-learn.org / L5 / modules / generated / sklearn.cluster.DBSCAN.html) to perform the clustering. Integration events in plasma samples were validated in matched tumor tissue when available, using HPV-DeepSeek. Cancer gene lists were obtained from the OncoKB™ Database. The complete list of genes used for annotation was obtained from the GENCODE database. For fragmentomic analyses, fragment length was measured using CollectlnsertSizeMetrics (Picard tools) and data was stratified based on alignment to human chromosomes 1-22 or HPV genomes, and patient of origin (HPV+HNSCC, cancer control, or healthy control). Calculations of skewness, kurtosis, and standard deviation were done using SciPy (version 1.11.4) and pandas (version 1.5.4). HPV-DeepSeek threshold determination: Optimal thresholds for sensitivity andMEEI 2024-326-02Quarles 169511.00058specificity were established using serial dilutions (Ing - O.OOOOlng) of HPV+HNSCC patient cfDNAinto a mixture of control patient cfDNArun in triplicate (FIG. 5G, Top, Table 1). points for a positive assay were determined to be >10 HPV reads and >10% HPV genome coverage and were utilized for all subsequent analyses.

[0099] Table 1. Total input in nanograms of DNAfrom HPV patient samples and control samples used to determine the detection limit for HPV-DeepSeek. Each row indicates the combination of patient sample DNA and control diluent used to achieve a total input of 20 ng per reaction. Control samples include a pure control diluent, and three different patient controls to establish baseline readings. The HPV patient samples are serially diluted with control diluent to demonstrate the assay's sensitivity across a range of HPV DNA concentrations from 1 ng to 0.00001 ng patient cfDNA input. All dilutions were in triplicate.

[0100] HPV-DeepSeek Threshold Determination: Optimal thresholds for sensitivity and specificity were established using serial dilutions (Ing - O.OOOOlng) of HPV+HNSCC patient cfDNAinto a mixture of control patient cfDNArun in triplicate (FIG. 5B, Table 1). Cut points for a positive assay were determined to be >10 HPV reads and >10% HPV genome coverage and were utilized for all subsequent analyses.

[0101] Tissue-based HPV WGS: DNAfrom 26 FFPE tissue samples underwent WGS using both HPV-DeepSeek and an amplicon-based HPV WGS developed by Mirabello and colleagues, referred to here as the NCI CHANGeS (Carcinogenic HPV All Next Generation Sequencing) assay and considered the reference gold standard for HPV genome annotation. ThisMEEI 2024-326-02Quarles 169511.00058assay uses Ion AmpliSeq panels designed to amplify the whole-genomes of the 13 high risk HPV types concurrently. WGS was performed for all of the HR-HPV types concurrently on the Ion PGM platform. Raw sequencing reads were quality-controlled and adaptor trimmed using the Torrent Suite Software (TMAP) and aligned to the HPV reference genomes using the Torrent Mapping Alignment Program. Variant calling was performed using the Torrent Variant Caller (TVC) and SnpEff was used for variant annotation. Three samples were excluded from the study due to low-quality reads. For viral SNP variant calling, AF thresholds were set at >0.5 for both 'NCI CHANGeS1and HPV-DeepSeek.

[0102] Modeling: Number needed to screen (NNS) and positive predictive value (PPV) were calculated to determine the diagnostic performance efficiency between ctHPVDNA-based approaches (ddPCR and WGS) in comparison to HPV Ab, following the method utilized by Lou et al. and as shown here.Sensitivity X incidencePPV = - - - Sensitivity X incidence) -I- ([1 — specificity] x [1 — incidence])

[0103] An incidence rate of 34.9 per 100,000 was used for HPV-associated oropharynx cancer in men aged between 55-74, according to Center for Disease Control statistics (gis.cdc.gov / Cancer / USCS / # / RiskF actors / ). For modeling a two-step approach, performance metrics of HPV Ab were first applied as above (step 1), followed by HPV-DeepSeek to only the population screening positive on step 1 (step 2).

[0104] Statistical analysis and hypothesis testing: The objective of this study was to compare the diagnostic accuracy of ctHPVDNA-based detection approaches (HPV WGS, multiplex ddPCR, singleplex ddPCR), and HPV Ab-based approaches to first attempt clinical biopsy, benchmarked to clinical gold standard diagnosis. Diagnostic accuracy was measured by Youden indexes as described by Chen et al. We tested the hypothesis that HPV WGS has improved diagnostic accuracy over other ctHPVDNA-based approaches, HPV Ab, combinatorial approaches (HPV WGS and Ab) and clinical standard of care. Statistical significance was assessed using a Z test to compare Youden indices across the same number of cases and controls,MEEI 2024-326-02Quarles 169511.00058with kappa statistics, variances, covariances, confidence intervals, and p-values calculated to determine the significance of differences. Multiplex ddPCR was used as the primary ddPCR approach for comparisons as this is the most commonly used approach clinically. Sensitivity for detection in early-stage disease was evaluated by Z test for comparisons of Youden indices across assays, while Fisher’s tests were employed for between-group comparisons. Wilcoxon rank sum tests were used for within-assay comparisons. The accuracy of blood-based annotation of known prognostic features, including PIK3C A mutations, integration events, genotype / sub-lineage assignment, and HPV16 high-risk SNPs, were validated using tissue sequencing with HPV-DeepSeek for both human and viral features and with NCI CHANGeS for viral features. Kendall r coefficient was used to assess the monotonic relationship between ctHPVDNA burden and clinical variables, while Cliffs Delta provided a measure of effect size (Table 2). To assess the relationship between ctHPVDNA levels and disease characteristics, we compared ctHPVDNA read counts across different groups within T stage, N stage, overall stage, and integration status. Bootstrap techniques were used to estimate 95% confidence intervals for the differences in median ctHPVDNA read counts between these groups. Spearman correlation coefficient was calculated to assess the relationship between ordinal clinical variables (e.g., T stage, N stage, and overall stage) and the number of unique reads, to enable an understanding of how tumor characteristics correlate. Correlation coefficients (Spearman and Kendall T) were interpreted as follows: less than ±0.10 as very weak, ±0.10-0.19 as weak, ±0.20-0.29 as moderate, and strong as ±0.30 or greater. Prior to multivariate analysis, a Box-Cox transformation was applied to normalize the outcome variable for further analysis. The Shapiro-Wilk test was then performed on the transformed data to verify statistical assumptions of normality.MEEI 2024-326-02Quarles 169511.00058&>

[0105] Table 2. Distribution and statistical analysis of 152 patients, organized by T stage, N stage, overall stage, and HPV integration status. Patients were classified by pathological staging, but in cases where surgery was not performed, clinical staging was used. For each category and subcategory, the mean number of reads and the standard deviation (SD) are shown. To quantify the effect size, we report proportion differences with 95% confidence intervals (CI), which describe the magnitude of difference between group proportions. The Kendall tau (r) coefficient with 95% CI measures the strength of the correlation between stage and the number of reads. Differences in medians with 95% CI are presented to show the median variation between categories. Eta-squared (q2) values with 95% CI provide a measure of effect size, except in the case of integration status where it is not applicable (N / A).

[0106] Machine learning: Data pre-processing involved binning categorical staging into grouped categories to create discrete target variables. To ensure consistent scaling across input variables, we normalized features using StandardScaler for applicable models. The dataset was then split into training (70%), validation (15%), and test sets (15%) using the train test split function. The training set was used to train the model, the validation set for hyperparameter tuning, while the withheld test set provided an evaluation of the final model's performance onMEEI 2024-326-02Quarles 169511.00058unseen data. We designed a multi-class machine learning approach to predict T, N, and overall stages. Using scikit-leam's MultiOutputClassifier, we trained a single multi-label, multi-class model to predict all three staging components. This framework applied interpretable model architectures, including Logistic Regression, Random Forest, and Gradient Boosting Machine. Model performance was assessed through accuracy, precision, recall, and Fl -score for each target variable. We applied a bootstrap method to calculate 95% confidence intervals for these metrics. To evaluate whether our models performed significantly better than naive guessing, we established a baseline classifier that predicts the most frequent class (or bin) for each target and conducted permutation tests for each target variable (naive classifier). This statistical test involved randomly shuffling the true labels of the target variable and comparing the model's accuracy against these permuted labels across 1,000 permutations. Analyses were performed using Python version 3.11.4, Jupyter notebook version 6.5.4, and sckit-learn version 1.3.0.

[0107] Results

[0108] Cohort demographics and clinical diagnostic workup.

[0109] 152 HPV+HNSCC patients were enrolled. 152 HPV+HNSCC patients were enrolled. The mean age was 62 years (range 36-84). The cohort was predominately male (88%; 134 / 152) with American Joint Committee on Cancer (AJCC) 8th edition stage I (77%; 117 / 152) oropharyngeal cancer (91%; 138 / 154) (Table 3). 152 population-level controls were enrolled to match the cases 1 : 1 with a distribution of healthy controls to HPV-independent HNSCC controls of 2: 1. Cancer cases underwent physicians’ choice clinical work up which led to fine-needle aspiration (FNA) of a lymph node in 53% of the cohort (81 / 152 cases) and primary tumor biopsy in 47% (71 / 152), as first diagnostic attempt. On the first diagnostic attempt, the diagnostic success rate was 68% (55 / 81) for FNA and 97% (69 / 71) for tissue biopsy, with an overall diagnostic success rate of 82% (124 / 152), meaning 18% of the cohort required repeat biopsy (Table 3). pl6 IHC was the most common HPV diagnostic test ordered on first diagnostic attempt with a diagnostic success rate of 64% (34 / 53) with FNA, 99% (68 / 69) with tissue and 84% (102 / 122) overall (Table 4, Table 5). A specific HPV genotype was assigned in 50% of cases (76 / 152).MEEI 2024-326-02Quarles 169511.00058

[0110] Table 3 : Patient demographics for Example 1. Clinical and demographic characteristics of HPV+HNSCC cohort. Diagnostic success determined by 1. Positive HPV testing and 2. Histopathology for squamous cell carcinoma. Diagnostic success rate is the positive diagnostic success results on first attempt tissue biopsy over total number of patients who had that biopsy type.MEEI 2024-326-02Quarles 169511.00058[0U1] Table 4. Composite total number of cases that underwent each biopsy type and HPV test type, obtained from all biopsy attempts, and composite success rates.

[0112] Table 5. The number of cases that underwent each biopsy type and HPV test type, subdivided by first, second, or third biopsy attempt, and diagnostic success rates.MEEI 2024-326-02Quarles 169511.00058

[0113] HPV-DeepSeek diagnostic accuracy and genotyping accuracy

[0114] cfDNA from all the cases and controls were run on HPV-DeepSeek (FIG. 5A). The mean number of ctHPVDNA reads for HPV+HNSCC cases was 48,997 (range 8-812,564) with a mean genome coverage of 90% (range 9-100%), resulting in a sensitivity of 98.7% and specificity of 98.7% using pre-specified thresholds (FIG. 5B and FIG. 5G).Sensitivity remained statistically unchanged when restricted to only oropharynx cancer (sensitivity 98.6%), early stage disease (AJCC 8 stage I sensitivity 98.2%), and when restricting to the very earliest cases (T1 with 0-1 lymph nodes <3cm sensitivity 94.1%) (Table 6). HPV genotype was assigned in 100% of cases (152 / 152) by HPV-DeepSeek. Eight genotypes were detected in the cohort, which matched clinical genotyping in 99% of cases (75 / 76) (FIG. 5C, Table 7). In the single case of discordant result between clinical tissue-based genotyping and HPV-DeepSeek, clinical PCR detected HPV 16 while two distinct genotypes were detected by HPV-DeepSeek, HPV16 with 4,990 reads and 100% genome coverage and HPV67 with 6,624 reads and 99.2% genome coverage, resulting in HPV67 being assigned as the dominant genotype due to a higher read count. HPV16 was the dominant genotype across the cohort (88%; 134 / 152) followed by HPV33 (3%; 5 / 152), 45 (3%; 4 / 152), 35 (2%; 3 / 152), 56 (2%; 3 / 152) and 69, 67, 18 (1 each) (FIG. 5C, Table 7). HPV16 sublineages were assigned in 98% of HPV16 cases (131 / 134), with HPV16 Al being the dominant sublineage (69%; 90 / 131) (FIG. 5C).MEEI 2024-326-02Quarles 169511.00058> > >> > > > > > > >> > >> >& < <

[0115] Table 6. Detailed comparison of performance of four different assays: HPV-DeepSeek, Singleplex ddPCR, Multiplex ddPCR, and HPV Ab, across patient subgroups (Oropharynx subsite only, AJCC 8 Stage 1, and Tl, 0-1 LN, <3cm). The table shows the percentage of positive results for each assay within each subgroup, accompanied by corresponding Fisher’s exact test p-values, to show the statistical significance of differences between compared groups.MEEI 2024-326-02Quarles 169511.00058

[0116] Table 7. Results of 304 (152 HPV+HNSCC and 152 controls) samples run on HPV-DeepSeek grouped by the genotype annotated by HPV-DeepSeek and status (true positive, false negative, false positive).

[0117] Head-to-head comparison of blood- and tissue-based diagnostic approaches

[0118] HPV-DeepSeek, HPV Ab, singleplex ddPCR, multiplex ddPCR and first attempt clinical biopsy were compared, benchmarked to clinical gold standard diagnosis (Table 8).Singleplex ddPCR sensitivity was 94.2% (98 / 104) and specificity was 98.6% (69 / 70). Multiplex ddPCR sensitivity was 90.6% (96 / 106) and specificity was 96.3% (77 / 80). HPV serology sensitivity was 86.4% (121 / 140) and specificity was 96.3% (4 / 108) (FIG. 5D). A combinatorial approach using both HPV-DeepSeek and HPV Ab, requiring both tests to be positive for a positive test, yielded a sensitivity of 87.4% (90 / 103) and specificity of 98.8% (80 / 81). In a head- to-head comparison of >100 overlapping cases for each comparison, HPV-DeepSeek demonstrated significantly improved diagnostic accuracy compared to ddPCR (HPV-DeepSeek Youden Index 0.990 vs multiplexed ddPCR Youden Index 0.904, P<0.001), HPV Ab (HPV-DeepSeek Youden Index 0.986 vs. HPV Ab Youden Index 0.827, P<0.001) and clinical workup first biopsy ((HPV-DeepSeek Youden Index 0.987 vs. clinical work up Youden Index 0.816, P<0.001) (FIG. 5E, Table 9). FIG. 5G shows the relative ctHPVDNA count using NGS andMEEI 2024-326-02Quarles 169511.00058ddPCR; NGS collected significantly more ctHPVDNAthan ddPCR. HPV-DeepSeek remained more sensitive and with resultant statistically higher diagnostic accuracy for early stage disease (Table 10). For example, HPV-DeepSeek sensitivity and diagnostic accuracy for AJCC 8 stage I disease was 98.7% and 0.987 vs. ddPCR, 89.9% and 0.861, p=0.024. HPV-DeepSeek sensitivity and diagnostic accuracy for T1 with 0-1 lymph nodes <3cm was 100% and 1.0 vs ddPCR 54.5% and 0.508 p=0.001). HPV-DeepSeek and ddPCR HPV read counts were highly correlated in positive samples (R=0.95, 95% CI 0.91-0.97) with HPV-DeepSeek correctly detecting HPV in samples that were ddPCR-negative due to low reads (FIG. 5F). FIG. 5H shows a comparison of HPV-DeepSeek and ddPCR, demonstrating an 80-fold increase in sensitivity by HPV-DeepSeek.< > >& <> >

[0119] Table 8: Blood-based approaches for early detection and diagnosis of HPV-associated head and neck cancers. Assay performance matrix of multiplex ddPCR (186 cases & controls), singleplex ddPCR (174 cases & controls), HPV Ab (248 cases & controls), HPV WGS (304 cases & controls), and a combination of HPV WGS and HPV Ab, (184 cases & controls). In the combinatorial approach, a sample is considered positive only if both HPV WGS and HPV Ab tests are positive. Assays were evaluated based on four metrics: sensitivity, specificity, positive predictive value, and negative predictive value. Youden Index was calculated for each assay, a composite measure of sensitivity and specificity. HPV WGS had the highest diagnostic accuracy. * Historical data, Hibbert et al., 2021.< & > <> <MEEI 2024-326-02Quarles 169511.00058

[0120] Table 9. Statistical comparison of Youden Indexes between HPV-DeepSeek, ddPCR, HPV Ab, and clinical workup. The confidence interval, the difference in Youden, z- statistic, and p-value are calculated.>>>&&&

[0121] Table 10. Quantitative comparison of the sensitivity, specificity, and Youden Index for HPV-DeepSeek against ddPCR and HPV Ab, in both AJCC 8 Stage 1 and Tl, 0-1 LN, <3 cm conditions.

[0122] Blood-based genomic prognostic feature annotationHigh-risk SNPs: 23 HPV+HNSCC cases underwent NCI CHANGeS HPV whole genome sequencing assay to generate gold-standard genome annotation from tumor tissue as well as plasma and tumor sequencing with HPV-DeepSeek. Viral genome mutation rates were varied across samples (FIGS. 6A-6J, FIGS. 7A-7D). There was high concordance for positive mutation calls between the assays: HPV-DeepSeek plasma vs. HPV-DeepSeek tumor: 0.902, HPV- DeepSeek plasma vs. NCI CHANGeS: 0.821, HPV-DeepSeek tumor vs. NCI CHANGeS: 0.873 (FIG. 6B, FIGS. 7A-7D). The concordance for both positive and negative mutation calls was equal to or nearly 100%: HPV-DeepSeek plasma vs. HPV-DeepSeek tumor: 1.000, HPV- DeepSeek plasma vs. NCI CHANGeS: 0.999, HPV-DeepSeek tumor vs. NCI CHANGeS: 0.999 (Table 11). Annotation of eight known high-risk HPV16 SNPs demonstrated 100% concordance across all assays (21 / 21 concordant calls). 16% of HPV16 cases (21 / 134) contained >1 high-risk SNP, consistent with prior tissue-based sequencing studies (FIG. 6C, FIG. 6D). PIK3CA: 9% of cases (14 / 152) contained high confidence PIK3C A mutations (FIG. 6E). Of the 14 PIK3CA mutations, five were known APOBEC-induced hotspot mutations. Viral integration events: 34%MEEI 2024-326-02Quarles 169511.00058(52 / 152) of cases contained >1 high confidence integration event (breakpoint). Of 231 total events, 65% (150 / 231) mapped to human genes. The number of integration events varied widely between patients (mean 4.4, range 1-40). As described previously by Symer et al., breakpoints within individual tumors were frequently clustered (FIG. 6F). Breakpoints were unevenly distributed across the viral genome and human genome. Normalized to length, El was the viral gene with the most breakpoints, and chr9 was the human chromosome with the most breakpoints (FIGS. 6G-6H, Tables 12A-12B). Breakpoint frequency was significantly higher in cancer-related genes compared to non-cancer genes (P<0.001) (FIG. 61, Table 12C). The HPV genome showed no preferential loss in any region, except for polyadenylation sites (FIG. 6J).> > > > > > > > > > > > > > > > > >

[0123] Table 11. Concordance rates evaluating agreement between HPV-DeepSeek variant calls for positive mutation detection (sensitivity) and overall mutation status detection (sensitivity and specificity). Sensitivity refers to the detection of true positive mutations in the HPV genome, while sensitivity and specificity assess the overall ability to determine both positive and negative mutation status correctly. HPV-DeepSeek results are compared against NCI CHANGeS, and plasma samples are compared against both NCI CHANGeS and tissue samples. The concordance rates range from 0 (no agreement) to 1 (perfect agreement).MEEI 2024-326-02Quarles 169511.00058<> & <

[0124] Table 12A. Distribution of insertional breakpoints within HPV16-positive genome. 212 virus-host breakpoints were identified and mapped to HPV16 genes. 'Gene' lists individual genes within the HPV16 genome, while 'Region' provides the nucleotide coordinate range for each gene obtained using the Papillomavirus Episteme (PaVE) database. 'Size' indicates the length of the gene in base pairs. 'Expected Breakpoints' shows the number of breakpoints anticipated across each gene if distributed randomly, based on gene size. 'Observed Breakpoints' records the actual number of breakpoints identified. The 'Ratio' column is the proportion of observed to expected breakpoints, and the 'p-value' column reports the statistical significance of the observed distribution, as determined by a two-tailed binomial test. The 'FDR-adjusted p-value' provides p-values adjusted for false discovery rate, offering a corrected measure of significance when multiple tests are conducted. The analysis follows the methodology of Symer et al. (2022).MEEI 2024-326-02Quarles 169511.00058> > > > > > > > > > > > >> > > > > > > > > > > > > > > > > > > > > >

[0125] Table 12B. Distribution of insertion breakpoints across human chromosomes.231 breakpoints were mapped. 'Size' indicates the length of the chromosome in base pairs.'Expected Breakpoints' shows the number of breakpoints anticipated across each chromosome if they were distributed randomly, based on gene size. 'Observed Breakpoints' records the actual number of breakpoints identified. The 'Ratio' column is the proportion of observed to expected breakpoints, and the 'p-value' column reports the statistical significance of the observed distribution, as determined by a two-tailed binomial test. The 'FDR-adjusted p-value' provides p-values adjusted for false discovery rate, offering a corrected measure of significance when multiple tests are conducted. The analysis follows the methodology of Symer et al. (2022).MEEI 2024-326-02Quarles 169511.00058P

[0126] Table 12C. The table shows the distribution of breakpoints in cancer / non-cancer genes between expected and observed. Expected values were calculated based on the total gene count from GENCODE (27,176 genes), multiplied by the number of samples with HPV integration events (52), as per the methodology of Symer et al. (2022). Cancer-associated genes were identified using the OncoKB database. In the 'Expected' row, the number of cancer and non-cancer genes is listed alongside the fraction they represent in the total gene count. The 'Observed at HPV Breakpoints' row reports the actual number of observed breakpoints in both categories and the fraction of breakpoints in cancer genes relative to all genes. The 'Ratio' column compares the observed fraction in cancer genes to the expected fraction. The p-value, calculated using Fisher's exact test, provides statistical significance for the breakpoints in cancer-associated genes.

[0127] Blood-based clinical prognostic feature annotation

[0128] As reported previously, levels of ctHPVDNA at diagnosis correlated positively with T stage (P<0.001), N stage (P<0.001), and overall stage (P=0.05) (FIGS. 8A-8C). Multivariate regression analysis identified N stage as the strongest predictor of ctHPVDNA levels (Table 13). Cases with multiple high-risk features detected from the blood (high-risk SNPs, PIK3C A variants, integration events) had a higher overall stage (P<0.01) and higher ctHPVDNA read count (P<0.01) (FIG. 9). Fragmentomic features were analyzed. ctHPVDNA demonstrated a significantly shorter mean fragment size (148 bp) than total cfDNA from HPV+HNSCC patients (167 bp) and healthy controls (169 bp) (P<0.001) (FIGS. 10A-10E). The mean read length ratio between ctHPVDNA and cfDNA was consistently lower across HPV+HNSCC cases (FIG. 11). ctHPVDNA also exhibited broader distribution of fragment sizes (P<0.001) and increased skewness (P<0.001) (FIGS. 10C-10D). Next, to evaluate the potential ofMEEI 2024-326-02Quarles 169511.00058integrating all HPV-DeepSeek features (FIG. 10E) to predict clinical prognostic features such as disease stage from the blood, multiclass classifiers were trained and tested using all blood-based HPV-DeepSeek outputs with an emphasis on predicting nodal stage, a crucial clinical feature in determining treatment approach. Several models, including decision trees, random forests, support vector machines, and ensemble methods were judged based on their ability to handle both numerical and categorical data, for their interpretability, as well as robustness to overfitting. Logistic Regression demonstrated the best accuracy in the held-out test set, achieving 0.87 (95% CI: 0.74-1.00) for N stage, 0.87 (95% CI: 0.70-1.00) for overall stage and 0.70 (95% CI: 0.52-0.87) for T stage (FIGS. 12A-12C, Table 14). Classification accuracy was significantly higher than a baseline guessing strategy for T stage (P=0.020), N stage (P=0.001), and overall stage (P=0.006) (Table 15).>>>

[0129] Table 13. Multivariate analysis using a Generalized Linear Model (GLM) with a Gamma distribution and a log link function. The study aims to examine the influence of predictors on the dependent variable, the number of reads. 'Characteristic' outlines the variables of predictors (T stage, N stage, overall stage, and integration status). The unadjusted and adjusted risk ratios with 95% confidence intervals are presented, comparing the reference group to higher stages and integration status. The '(P>|t|)' columns provide the p-values for testing the nullMEEI 2024-326-02Quarles 169511.00058hypothesis that each coefficient equals zero in the unadjusted and adjusted models. The adjusted model accounts for potential confounders and indicates the independent effect of each variable on the outcome.&& > ><> < <& <>

[0130] Table 14. Accuracy of multiclass classifiers in held out test set. The rows correspond to model architectures. The columns refer to the model predictions for T stage, N stage, and overall stage. For each predicted variable, the table provides the baseline performance (the accuracy of the classifier predicting the majority class), the model performance (the accuracy achieved by the classifier), and the p-value, from the permutation test showing a statistically significant difference between a baseline strategy and the model performance.&

[0131] Table 15. Performance of Logistic Regression evaluated on the prediction of T, N, and overall stages. The metrics reported include the weighted Accuracy, Precision, Recall,MEEI 2024-326-02Quarles 169511.00058and Fl -Score, as well as the permutation test p-values, which assess the statistical significance of the model's accuracy compared to a naive classifier. Naive classifier values are provided for reference.

[0132] Modeling HPV+HNSCC early detection

[0133] We modeled screening efficiency of HPV-associated oropharynx cancer, the majority subset of HPV+HNSCC in men using diagnostic accuracy of ctHPVDNAbased approaches (HPV-DeepSeek, multiplex ddPCR) and HPV Ab to estimate the NSS and PPV. Using an incidence rate of 34.9 per 100,000 for HPV-associated oropharynx cancer in men ages 55-74, HPV-DeepSeek yielded the lowest NNS (2,903 men) and highest PPV (2.6) compared to multiplex ddPCR (3,163 and 0.8), and HPV Ab (3,316 and 0.8) (FIGS. 4A-4B, Table 16). We then modeled a two-step screening approach, as has been proposed previously, in which first a low cost, long lead time risk biomarker was applied (HPV Ab) followed by the highest accuracy diagnostic biomarker (HPV-DeepSeek) in only patients who were HPV Ab positive. Using this two-step approach, NNS in the second step was 96 and PPV was 44.7 (FIGS. 4C-4D, Table 17).

[0134] Table 16. One-step hypothetical screening program with results for each test evaluated.MEEI 2024-326-02Quarles 169511.00058

[0135] Table 17. Two-step hypothetical screening program with HPV Ab followed by HPV-DeepSeek in participants positive on HPV Ab screen.

[0136] Discussion

[0137] We conducted a head-to-head comparison of multiple blood-based HPV+HNSCC ctHPVDNAand HPV Ab early detection approaches in a case-control cohort. We found that HPV WGS-based detection alone demonstrated the highest sensitivity, specificity, and diagnostic accuracy compared to dPCR-based ctHPVDNA detection, HPV Ab-based detection and combinatorial approaches and thus the lowest number needed to screen and highest positive predictive value when screening efficiency was modeled. Integration of blood-based features from HPV WGS detection with machine learning resulted in accurate clinical staging. Taken together, these results highlight the potential utility of an HPV WGS liquid biopsy approach for screening early detection and diagnosis of HPV+HNSCC.

[0138] The incidence of HPV+HNSCC is rising in the US and thus strategies for early detection could have meaningful public health benefits. Both ctHPVDNA detection and HPV Ab detection have shown promise as potential screening biomarkers for HPV+HNSCC. However, at present, both have critical limitations preventing clinical utility and implementation. Primarily, ctHPVDNA detection approaches lack adequate sensitivity for low-level ctHPVDNA detection, such as that seen in screening while HPV Ab have a long and variable (up to >25 years) lead time between seroconversion and HPV+HNSCC development, and are non-linear,MEEI 2024-326-02Quarles 169511.00058leading to limited real-world clinical utility when used alone. Combinatorial ctHPVDNA ddPCR and HPV Ab approaches have been evaluated in small cohorts but have not been found to improve performance at cancer diagnosis. Such limitations could be addressed through significantly more sensitive ctHPVDNA detection approaches alone, or in combination with HPV Ab. Here, we found in head-to-head comparison that HPV WGS was significantly more sensitive than dPCR-based ctHPVDNA detection, and that this sensitivity increase was maintained in the earliest stages of disease, where population-level screening is aimed. Because of the extremely high diagnostic accuracy of HPV WGS, combining this approach with HPV Ab did not lead to statistically improved performance at diagnosis.

[0139] Beyond improved diagnostic accuracy, an additional advantage of a multifeature HPV WGS approach is its potential to also function as a prognostic biomarker. In this study we found HPV-DeepSeek accurately annotated known genomic prognostic features from the blood, including integration events, PIK3C A mutations, viral sublineage and high-risk viral SNPs. This suggests that such an approach could not only serve as a screening biomarker but also as an integrated prognostic biomarker for personalized treatment. We further explored if a completely blood-based approach could be used to not only generate genomic prognostic information but predict clinical stage. Integrating HPV-DeepSeek output with machine learning led to overall stage predication accuracy of 0.87, suggesting that with further refinement, accurate stage prediction from only blood-based features may be feasible. Similar to prior to studies, ctHPVDNA detection outperformed standard of care clinical diagnosis. Further, HPV-DeepSeek resulted in an accurate HPV genotype assignment 50% more frequently than standard of care. When taken together, it is feasible to imagine that a fully blood-based approach could be used as a longitudinal integrated diagnostic and prognostic biomarker, functioning as a screening detection biomarker, confirmatory diagnostic biomarker, and prognostic biomarker, without the need for traditional standard of care tissue biopsy. Such an approach would require rigorous prospective study and validation.

[0140] The high diagnostic accuracy of HPV WGS in this study, even at the earliest stages of disease, suggests screening early detection may be feasible, with the goal of detecting HPV+HNSCC prior to the onset of regional lymph node metastases- the most common presenting symptom at present. Shifting diagnosis to earlier stages of disease could lead to lessMEEI 2024-326-02Quarles 169511.00058treatment morbidity by permitting targeted, single modality treatments (surgery or radiation alone) as opposed to the multi-modality treatments typically used at present time (surgery followed by radiation or radiation with chemotherapy), simultaneously improving quality of life and decreasing time toxicity, patient and payor costs, and potentially mortality as well. Prior studies examining HPV+HNSCC screening utilizing HPV Ab alone have suggested screening paradigms in which men would be screened at 50 years old3. We chose to model men ages 55-75 due to the mean age of HPV+HNSCC increasing across time with the most rapidly increasing demographic being men 65-741,31. Utilizing a framework described by Lou et al., we modeled NSS and PPV demonstrating lower NNS and higher PPV for HPV WGS than HPV Ab and dPCR -based ctHPVDNA detection. To put in these numbers in context, a NNS of 2,903 men and a PPV of 2.6 compare favorably to proposed NNS and PPV values for EBV-associated nasopharyngeal carcinoma screening, a similar oncogenic viral cancer type endemic in southern China, in which circulating tumor EBV DNA screening is being studied. There, NNSs of 5,656-2,262 and PPVs of 1.0-2.4% for incidence rates of 20 and 50 / 100,000 respectively have been proposed as reasonable24. However, these NSS values are elevated above existing cancer types for which screening is recommended in the US (Breast, Colon, Lung, Cervical). Thus, we then modeled a two-step screening approach, as has been proposed previously3, in which first a low cost, long lead time risk biomarker was applied (HPV Ab) followed by a highly accurate diagnostic cancer biomarker (HPV WGS) in only patients who were positive on first screen. This yielded a NSS of 96 in the second step, on par or lower than existing US cancer screening approaches. These NNS and PPV estimates based on incidence rates are conservative, as a highly accurate screening approach would be expected to detect not only symptomatic incident cases but asymptomatic cancers as well. High specificity is critical for cancer screening to minimize harm to benefit ratio by avoiding unnecessary clinical interventions and their resultant psychosocial and cost impacts from false positives. Individual test specificity of HPV WGS and HPV WGS + HPV Ab was adequately high for use in high incidence populations but would likely not be sufficient for low incidence population use, such as the general population (specificity 98.7-98.8%). Future studies are needed to define, evaluate and model alternative assay cut points which favor specificity (such as sensitivity at 99.5% specificity), as would be needed for population level screening and in settings with ultra-low tumor burden (asymptomaticMEEI 2024-326-02Quarles 169511.00058cases years prior to clinical detection), as well as two-step screening approaches which require both tests to be positive before a more intensive intervention and workup is undertaken.Additionally, further modeling is needed to develop post-blood-based detection clinical work-up algorithms to ensure benefit over harm in concert with cost considerations. Lastly, the high NNS, even with an ultrasensitive test, raises questions regarding the epidemiologic practicality of screening for HPV+HNSCC and suggests early detection may need to focus on specific high risk groups where the incidence of HPV+HNSCC is higher.

[0141] Multiple limitations should be noted. Cases and controls were both recruited from hospital clinics. Thus, controls may not reflect the general population (selection bias), although controls recruited from hospital clinics are likely to have increased cancer risk, leading to a bias favoring decreased specificity. Specificity estimates may change when applied in population-based cohorts, which will be a necessary next step. Cases were individuals who were either symptomatic from their cancer or had an incidentally found lesion on cross-sectional imaging or physical exam. Thus, these cases may differ in tumor size (larger) and stage (higher) from a screening setting, leading to over-estimation of sensitivity. Taken together, because this is a EDRN phase 2 biomarker study and not a phase 3 or 4 study, meaning the cases and controls are not from population based studies, the screening modeling results should be interpreted with caution and may not be generalizable to the population. While the comparison assays used in this study have performance metrics similar to published literature, performance of alternative dPCR and HPV Ab assays may exceed these, which could theoretically alter findings of the study. For example, multi-probe dPCR assays may have improved sensitivity compared to single probe dPCR assays, such as those used here. Other groups have examined HPV WGS vs dPCR in cervical cancer, finding improved sensitivity at the time of diagnosis. However, they also found no improvement in sensitivity in the minimal residual disease setting after treatment with chemoradiotherapy. The HPV Ab assays used here only target 4 HPV genotypes, inherently decreasing the sensitivity. Lastly, integration events detected in the plasma were validated in tissue by the same methodology, however, additional validation with gold standard long-read WGS of the tumor tissue will be required for additional external validation. The number of integration events detected in plasma was lower than expected from the tissue-based literature, suggesting we may be under-detecting integration events.MEEI 2024-326-02Quarles 169511.00058

[0142] In conclusion, we conducted the first head-to-head direct comparison of blood-based HPV+HNSCC ctHPVDNAand HPV Ab early detection approaches finding that HPV WGS-based ctHPVDNA detection alone demonstrated the highest sensitivity, specificity, and diagnostic accuracy and thus the lowest number needed to screen and highest positive predictive value when screening efficiency was modeled. These results highlight the promise of HPV WGS liquid biopsy for screening early detection of HPV+HNSCC and the need for modeling and cost-effectiveness studies to evaluate and guide HPV+HNSCC screening.

[0143] References for Example 1

[0144] 1. Tota, J. E. et al. Evolution of the Oropharynx Cancer Epidemic in the United States: Moderation of Increasing Incidence in Younger Individuals and Shift in the Burden to Older Individuals. J Clin Oncol 37, 1538-1546 (2019).

[0145] 2. Hibbert, J., Halec, G., Baaken, D., Waterboer, T. & Brenner, N. Sensitivity and Specificity of Human Papillomavirus (HPV) 16 Early Antigen Serology for HPV-Driven Oropharyngeal Cancer: A Systematic Literature Review and Meta-Analysis. Cancers (Basel) 13, 3010 (2021).

[0146] 3. Robbins, H. A. et al. Absolute Risk of Oropharyngeal Cancer After an HPV16-E6 Serology Test and Potential Implications for Screening: Results From the Human Papillomavirus Cancer Cohort Consortium. Journal of Clinical Oncology 40, 3613-3622 (2022).

[0147] 4. Brenner, N. et al. Characterization of human papillomavirus (HPV) 16 E6 seropositive individuals without HPV-associated malignancies after 10 years of follow-up in the UK Biobank. EBioMedicine 62, 103123 (2020).

[0148] 5 Kreimer, A. R. et al. Timing of HPV16-E6 antibody seroconversion before OPSCC: findings from the HPVC3 consortium. Annals of Oncology 30, 1335-1343 (2019).

[0149] 6. Lang Kuhs, K. A. et al. Human Papillomavirus 16 E6 Antibodies in Individuals without Diagnosed Cancer: A Pooled Analysis. Cancer Epidemiology, Biomarkers & Prevention 24, 683-689 (2015).

[0150] 7. Naegele, S., Ruiz-Torres, D. A., Zhao, Y, Goss, D. & Faden, D. L.Comparing the Diagnostic Performance of Quantitative PCR, Digital Droplet PCR, and Next-MEEI 2024-326-02Quarles 169511.00058Generation Sequencing Liquid Biopsies for Human Papillomavirus-Associated Cancers. The Journal of Molecular Diagnostics 26, 179-190 (2024).

[0151] 8. Siravegna, G. et al. Cell-Free HPV DNA Provides an Accurate and Rapid Diagnosis of HPV- Associated Head and Neck Cancer. Clinical Cancer Research 28, 719-727 (2022).

[0152] 9. Lang Kuhs, K. A., Brenner, J. C., Holsinger, F. C. & Rettig, E. M.Circulating Tumor HPV DNA for Surveillance of HPV-Positive Oropharyngeal Squamous Cell Carcinoma. JAMA Oncol 9, 1716 (2023).

[0153] 10. Berger, B. M. et al. Detection of Occult Recurrence Using Circulating Tumor Tissue Modified Viral HPV DNA among Patients Treated for HPV-Driven Oropharyngeal Carcinoma. Clinical Cancer Research 28, 4292-4301 (2022).

[0154] 11 Hanna, G. J. et al. Negative Predictive Value of Circulating Tumor Tissue Modified Viral (TTMV)-HPV DNA for HPV-driven Oropharyngeal Cancer Surveillance. Clinical Cancer Research 29, 4306-4313 (2023).

[0155] 12. Rettig, E. M. et al. Relationship of HPV16 E6 seropositivity with circulating tumor tissue modified HPV16 DNA before head and neck cancer diagnosis. Oral Oncol 141, 106417 (2023).

[0156] 13. Leung, E. et al. HPV Sequencing Facilitates Ultrasensitive Detection of HPV Circulating Tumor DNA. Clinical Cancer Research 27, 5857-5868 (2021).

[0157] 14. Beaty, B. T. et al. PIK3CA Mutation in HPV-Associated OPSCC Patients Receiving Deintensified Chemoradiation. JNCL Journal of the National Cancer Institute 112, 855-858 (2020).

[0158] 15. Koneva, L. A. et al. HPV Integration in HNSCC Correlates with Survival Outcomes, Immune Response Signatures, and Candidate Drivers. Mol Cancer Res 16, 90-102 (2018).

[0159] 16. Lang Kuhs, K. A. et al. Genetic variation within the human papillomavirus type 16 genome is associated with oropharyngeal cancer prognosis. Annals of Oncology 33, 638-648 (2022).MEEI 2024-326-02Quarles 169511.00058

[0160] 17. Kreimer, A. R. et al. Kinetics of the Human Papillomavirus Type 16 E6 Antibody Response Prior to Oropharyngeal Cancer. JNCI: Journal of the National Cancer Institute 109, (2017).

[0161] 18. Kreimer, A. R. et al. Evaluation of Human Papillomavirus Antibodies and Risk of Subsequent Head and Neck Cancer. Journal of Clinical Oncology 31, 2708-2715 (2013).

[0162] 19. Rettig, E. M. et al. Detection of circulating tumor human papillomavirus DNA before diagnosis of HPV-positive head and neck cancer. Int J Cancer 151, 1081-1085 (2022).

[0163] 20. Waterboer, T. et al. Multiplex human papillomavirus serology based on in situ-purified glutathione s-transferase fusion proteins. Clin Chem 51, 1845-53 (2005).

[0164] 21. Pinheiro, M. et al. Association of HPV35 with cervical carcinogenesis among women of African ancestry: Evidence of viral-host interaction with implications for disease intervention. Int J Cancer 147, 2677-2686 (2020).

[0165] 22. Pinheiro, M. et al. Phylogenomic Analysis of Human Papillomavirus Type 31 and Cervical Carcinogenesis: A Study of 2093 Viral Genomes. Viruses 13, (2021).

[0166] 23. Cullen, M. et al. Deep sequencing of HPV16 genomes: Anew high-throughput tool for exploring the carcinogenicity and natural history of HPV16 infection. Papillomavirus Res 1, 3-11 (2015).

[0167] 24. Lou, P.-J. et al. Performance and Operational Feasibility of Epstein-Barr Virus-Based Screening for Detection of Nasopharyngeal Carcinoma: Direct Comparison of Two Alternative Approaches. Journal of Clinical Oncology 41, 4257-4266 (2023).

[0168] 25. Chen, E, Xue, Y, Tan, M. T. & Chen, P. Efficient statistical tests to compare Youden index: accounting for contingency correlation. Stat Med 34, 1560-76 (2015).MEEI 2024-326-02Quarles 169511.00058

[0169] 26. Gunning, A. et al. Analytical Validation of NavDx, a cfDNA-Based Fragmentomic Profiling Assay for HPV-Driven Cancers. Diagnostics 13, 725 (2023).

[0170] 27. Rettig, E. M. et al. Association of Pretreatment Circulating Tumor Tissue-Modified Viral HPV DNA With Clinicopathologic Factors in HPV-Positive Oropharyngeal Cancer. JAMA Otolaryngology-Head & Neck Surgery 148, 1120 (2022).

[0171] 28. Symer, D. E. et al. Diverse tumorigenic consequences of human papillomavirus integration in primary oropharyngeal cancers. Genome Res 32, 55-70 (2022).

[0172] 29. Kentnowski, M. et al. Determinants of the level of circulatingtumor HPV16 DNA in patients with HPV-associated oropharyngeal cancer at the time of diagnosis. Sci Rep 13, 21226 (2023).

[0173] 30. Lewis, J. S. et al. Assessing the feasibility of a multimodal liquid biopsy for the diagnosis of HPV-associated oropharyngeal squamous cell carcinoma. Am J Clin Pathol 161, 570-578 (2024).

[0174] 31 Windon, M. J. et al. Increasing prevalence of human papillomavirus-positive oropharyngeal cancers among older adults. Cancer 124, 2993-2999 (2018).

[0175] 32. Pepe, M. S. et al. Phases of Biomarker Development for Early Detection of Cancer. JNCI Journal of the National Cancer Institute 93, 1054-1061 (2001).

[0176] 33. Bhambhani, C. et al. Development of a high-performance multiprobe droplet digital PCR assay for high-sensitivity detection of human papillomavirus circulating tumor DNA from plasma. Oral Oncol 143, 106436 (2023).

[0177] 34. Han, K. et al. Clinical Validation of Human Papilloma Virus Circulating Tumor DNA for Early Detection of Residual Disease After Chemoradiation in Cervical Cancer. Journal of Clinical Oncology 42, 431-440 (2024).Example 2: Circulating tumor HPV DNA whole genome sequencing enables HPV-associated oropharynx cancer early detection

[0178] IntroductionMEEI 2024-326-02Quarles 169511.00058

[0179] Circulating tumor (ct)DNA-based early detection approaches have the potential to drastically improve cancer screening, particularly for cancers that currently do not have screening approaches. At present, only four cancer types have population-level screening tests recommended, meaning that most diagnoses are in cancer types for which there is no screening and thus tend to be later-stage disease, when the cancer has reached a size to become symptomatic. Because diagnosis of later-stage cancers results in decreased survival, increased costs, and increased treatment morbidity, detection of asymptomatic, early-stage cancers creates a unique opportunity for simultaneous improvement in all three of these critical domains.

[0180] HPV-associated oropharyngeal squamous cell carcinoma (HPV+OPSCC) is the most common HPV-associated malignancy in the United States, yet unlike cervical cancer, it lacks a screening test. Oropharyngeal infection with HPV typically occurs >20 years prior to clinical cancer diagnosis and genomic models suggest HPV+OPSCC begins developing 10-15 years prior to diagnosis. Many patients with HPV+OPSCC develop HPV early oncoprotein-directed antibodies (HPV Ab) (sensitivity 83-90%, specificity 95-99% at the time of diagnosis) which can be detected >10 years before diagnosis, supporting the presumed long development times of HPV+OPSCC 12-19. HPV+OPSCCs develop inside the deep crypts of the palatine and lingual tonsils and thus are frequently asymptomatic, even at the time of clinical diagnosis. The most common presenting symptom for HPV+OPSCC is a palpable neck mass, as HPV+OPSCCs metastasize to cervical lymph nodes in nearly all cases, by the time of diagnosis. Taken together, there is a substantial time window in which screening early detection of HPV+OPSCC may be feasible and enable detection of earlier-stage, non-metastatic cancers. While HPV Ab have been proposed as a potential screening biomarker for HPV+OPSCC, the long (up to 28 years) and variable lead time from seroconversion to cancer development limits their clinical utility.

[0181] HPV+OPSCCs release fragments of the viral cancer genome, termed circulating tumor HPV DNA (ctHPVDNA), which is detectable at diagnosis using droplet digital PCR-based approaches in -90% of patients, at -97% specificity, and is more reliable than mutation- or methylation-based ctDNA detection. dPCR-based detection of ctHPVDNA is now being used clinically at diagnosis and in treatment monitoring for HPV+OPSCC. A small proof-of-principle study has suggested that ctHPVDNA may be detectable in asymptomatic individuals years prior to clinical diagnosis with HPV+OPSCC, but that existing dPCR-based approaches areMEEI 2024-326-02Quarles 169511.00058not sensitive enough for reliable ctHPVDNA detection in most cases. Importantly, while transient oropharyngeal infection with HPV is common, it does not lead to detectable ctHPVDNA.

[0182] We recently developed a multi-feature HPV whole genome sequencing (WGS) liquid biopsy, termed HPV-DeepSeek, for improved low-level ctHPVDNA detection for use in HPV+OPSCC early detection (Bryan et al.). We tested the performance of HPV-DeepSeek head-to-head with existing blood-based HPV detection approaches (HPV Ab and ctHPVDNA dPCR) in a cohort of 153 HPV+OPSCC cases at cancer diagnosis and 153 paired controls, finding significantly improved sensitivity, specificity and diagnostic accuracy, which was maintained even the earliest stages of disease, suggesting accurate blood-based screening early detection may be feasible (EDRN phase 2)(Bryan et al.). Here, in a EDRN phase 3 study, we utilize this significantly more sensitive assay to test the hypothesis that ctHPVDNA can be accurately detected years prior to clinical diagnosis with HPV+OPSCC in a unique case-control cohort. Additionally, we examine how machine learning can be applied to further improve early detection sensitivity.

[0183] Methods

[0184] Participants and study design. All participants provided written informed consent to protocols approved by the Institutional Review Boards at Dana Farber / Harvard Cancer Center and MassGeneralBrigham. This study was conducted in compliance with the U.S. Common Rule. Screening cohort: The MassGeneralBrigham biobank has prospectively collected blood samples and data from 140,000 patients presenting for care in the healthcare system based in Boston, Massachusetts. Biobank clinical data was queried by biobank staff to identify patients who: 1. were diagnosed with HPV+OPSCC >1 year after blood sample contribution, 2. had >lml of plasma stored, and 3. had no history of an HPV-associated malignancy at the time of blood sample contribution. Medical record review was then undertaken to confirm HPV+OPSCC diagnosis and for extraction of clinicodemographic data, yielding a cohort of 28 HPV+OPSCC patients with l-4ml of plasma. 20 / 28 HPV+OPSCC patients were eventually diagnosed within MassGeneralBrigham permitting retrieval of fresh-frozen paraffin-embedded (FFPE) tumor tissue blocks from the time of diagnosis in 15 patients. 1:1 age-, sex-, and plasma volume-MEEI 2024-326-02Quarles 169511.00058matched controls without a history of HPV cancer were identified in the same biobank by random selection by biobank staff. Cases and controls were then coded and transferred for testing.

[0185] Independent machine learning training and testing cohort. A cohort of 153 HVP+OPSCC cases with blood samples from the time of cancer diagnosis and 153 general population controls were utilized to train and test machine learning models. Cases consisted of untreated HPV+OPSCC patients presenting to Mass Eye and Ear at the time of cancer diagnosis. HPV+OPSCC diagnosis consisted of histopathologic confirmation of SCC, pl6 immunohistochemistry, and direct HPV testing with RNA in-situ hybridization or DNA polymerase chain reaction (PCR). Controls consisted of 100 patients with no history of cancer presenting for non-cancer care at Mass Eye and Ear and 53 patients with a history of HPV-independent cancers, presenting for care at Mass Eye and Ear.

[0186] Blood and tissue collection and processing for DNA. Cell free (cf) DNA was extracted from l-4ml of frozen plasma as previously described for all cases and controls20. DNA was extracted from FFPE blocks after tumor identification and microdissection and sheared prior to sequencing. 20-50ng of either total cell free from plasma or total DNA from tissue was used as input for all downstream analyses.

[0187] HPV-DeepSeek processing and analysis. All samples were run and analyzed on HPV-DeepSeek as previously described (FIG. 2A). Cut points for a positive assay were >10 HPV reads and >10% HPV genome coverage as previously determined. Viral clonal mutations were called using bcftools, filtered based on allele fraction (>0.5) and read depth (>4) and then manually reviewed in Integrated Genomics Viewer (IGV) by three researchers independently (LA, SH, DD) to remove low-evidence mutations and artifacts. For cross-sample contamination analysis, samples were considered unique if they harbored a pattern of variants not present in any other sample. For tumor-plasma sample concordance analysis, the mutation patterns of plasma and tumor samples from the same patient were compared, to ensure the same single nucleotide variant (SNV) pattern.

[0188] Machine learning and statistical analysis. A binary machine learning task was employed to classify samples into HPV+OPSCC or no HPV+OPSCC. Before modeling, standardMEEI 2024-326-02Quarles 169511.00058data preprocessing techniques were implemented. Min-max normalization was performed on numeric features. We allocated 80% of the independent machine learning training and testing cohort for training and held out 20% for testing. We trained several interpretable model architectures, including Random Forest, AdaBoost, and Naive Bayes. Model performance metrics included accuracy, precision, recall, macro Fl -score, specificity, negative predictive value (NPV), and the area under the receiver operating characteristic curve (AUC-ROC).Hyperparameter tuning was conducted using 10-fold cross-validation on the training set for each model. We computed feature importance for the interpretable tree-based models to identify influential predictors. Once the models were tuned, they were finalized, and bootstrap resampling was conducted on the held-out test set to generate empirical 95% confidence intervals for the models' performance metrics. AUC-ROC curves were generated for the cross-validation and test sets using Stratified K-Fold cross-validation with five splits. To interpret the models, we utilized the permutation importance method. This method was applied to calculate and normalize the feature importance scores for all models. We visualized the feature importance to evaluate the impact of each feature on the models’ predictive performance. The machine learning models had equivalent performance with overlapping confidence intervals (Table 18). Naive Bayes was chosen as a representative model.Lower Ci andomFore cura 0.99464516 0.98387 andomFore ecisi 0.99418448 0.984 andomFore recal 0.99390323 0.98387 andomFore f1 0.99498247 0.98386 andomFore auc 0.99446774 0.98387 AdaBoost cura 0.9966129 0.98387 AdaBoost ecisi 0.99626563 0.984 AdaBoost recal 0.99632258 0.98387 AdaBoost f1 0.99651522 0.98386 AdaBoost auc 0.99629032 0.98387 NaiveBayes cura 0.98901613 0.98387 NaiveBayes ecisi 0.98932813 0.984 NaiveBayes recal 0.98875806 0.98387 NaiveBayes f1 0.98923914 0.98386NaiveBayesauc0.989096770.98387MEEI 2024-326-02Quarles 169511.00058

[0189] Table 18. Comparative performance metrics of machine learning models on test set. This table presents the performance evaluation of three bootstrapped machine learning models - NaiveBayes, RandomForest, and AdaBoost - using five metrics on the test dataset. The metrics used for assessment are accuracy, precision, recall, fl score, and area under the receiver operating characteristic curve (AUC). Each model’s performance is quantified by a score, alongside the lower and upper bounds of the 95% confidence interval (CI) for each metric. The parameters for bootstrap resampling were 1,000 repetitions from a sample of 50% of the original dataset.

[0190] HPV serology. Multiplexed Luminex -based HPV serology testing was carried out at the German Cancer Research Center (DKFZ) for HPV 16, 33, 35, and 45 using El, E2, E6 and E7 with pre-set thresholds, as previously described.

[0191] Results

[0192] Accurate early detection of HPV+OPSCC

[0193] cfDNA from 28 patients who contributed a blood sample to the MassGeneralBrigham biobank 1.3-10.8 years prior to diagnosis with HPV+OPSCC and 1:1 age-, sex- and plasma volume-matched controls underwent library preparation and molecular barcoding with HPV-DeepSeek and were loaded into a single multiplexed next generation sequencing run (total n=56) (FIG. 2C, FIG. 13 A, Table 19). Using prespecified cut-offs, 22 / 28 of cases tested positive for ctHPVDN A (overall sensitivity 79%) (FIG. 13B). ctHPVDNA results were negative in all controls (0 / 28 controls, 100% specificity). The maximum lead time from a positive sample to clinical cancer diagnosis was 7.8 years. The diagnostic accuracy was highest within four years of cancer diagnosis (Youden Index 1.0) and decreased further from diagnosis (FIG. 14, FIG. 15). Similarly, samples that were collected closest to diagnosis generally were the furthest from the test threshold and samples collected further from diagnosis were closer to the test threshold, or negative, as expected, suggesting increasing release of ctHPVDNA as the cancer develops across time (FIG. 13C).MEEI 2024-326-02Quarles 169511.00058

[0194] Table 19: Patient demographics for Example 2.

[0195] Machine learning increases sensitivity of HPV+OPSCC early detection

[0196] Utilizing HPV-DeepSeek output from an independent cohort of 153 HPV+OPSCC patients and 153 population controls, we trained and tested a machine learning model for classifying HPV+OPSCC vs no HPV+OPSCC. The bootstrapped AUC of this Naive Bayes model for detecting HPV+OPSCC at diagnosis was 0.99 (95% CI 0.98-1.0) (FIG. 2B). We next applied the locked classifier to HPV-DeepSeek output from the 28 HPV+OPSCC prediagnostic cases and 28 matched controls, finding improved classification accuracy (prespecified ctHPVDNA cutoff sensitivity 0.79; Naive Bayes sensitivity 0.96, 95% CI 0.96-0.96) (FIG. 13D, FIGS. 16A-16C, Table 20). The maximum lead time from a positive screeningMEEI 2024-326-02Quarles 169511.00058sample to clinical cancer diagnosis was increased to 10.3 years. The detection rate was 100% within 10 years of diagnosis. The only case not identified as HPV+OPSCC was that with the longest lead time, nearly 11 years. Machine learning improved the classification of the samples further from the time of diagnosis (FIG. 13E). FIG. 13F shows a comparison of HPV-DeepSeek 1.0 (left) and HPV-DeepSeek 2.0 (right). HPV-DeepSeek 2.0 improves the sensitivity for screening for HPV+ OPSCC over HPV DeepSeek 1.0.IterationRandomForestAdaBoostNaiveBavesModelMean 0.929 0.893 0.964 Model95Cl886-0.9842-0.94934-0.99

[0197] Table 20. Iterative Performance and Confidence Intervals of Machine Learning Models on Validation Set. Analysis of the performance of three machine learning models - NaiveBayes, RandomForest, and AdaBoost on the validation dataset. Performance is evaluated over five iterations, with each iteration's result expressed as a fraction of successful predictions out of 28 cases. The upper table shows the raw performance data for each iteration, while the lower provides summary metrics. This includes the mean accuracy of the model across all iterations and the 95% confidence interval (CI) for the model accuracy.MEEI 2024-326-02Quarles 169511.00058<< <> > >>&>& > ><

[0198] Table 21. HPV Multiplexed Serology Assay. This table presents the findings of a multiplex serology assay for HPV, conducted on a sample set of 28 cases. The serological responses are quantified as median fluorescence intensities (MFIs) for multiple HPV genotypes and their respective antigens. The assay targets antibodies against multiple epitopes of HPV16, including E6, E7, El, and E2, as well as other high-risk HPV types 33, 35, and 45, using their E6 and E7 antigens. A case is classified as HPV16 antibody positive if the MFI for HPV16 E6 alone is above the cut-off, or if at least three out of four antigens (El, E2, E6, E7) related to HPV 16 are above their respective cut-offs. In this instance, the E6 cut-off is specifically adjusted to 484 MFI. For non-HPV16 types (HPV33, HPV35, and HPV45), positivity requires both E6 and E7 MFIs to exceed their respective cut-off values. Some cases exhibited positivity for multiple HPV genotypes, which could indicate cross-reactivity in the assay. Where applicable, type specific E6 ratios were calculated for HPV16 and HPV33. Finding HPV16 E6MEEI 2024-326-02Quarles 169511.00058MFIs were considerably higher than the HPV33 E6 MFIs, these cases were classified as HPV16 positive by serology. Assay results are arranged in rows for each patient, with the patient ID listed in the first column, and subsequent columns indicating the MFI readings for each antigen. The row at the bottom of the table specifies the cut-off MFI values used to determine seropositivity.

[0199] Viral genome molecular fingerprinting

[0200] We utilized viral genome molecular fingerprinting between and within cases to validate our findings. 6 / 22 HPV-DeepSeek positive cases displayed unique HPV genotypes or sub-lineages (FIG. 17A). As expected from epidemiologic studies, some viral genomes shared the same HPV 16 sub-lineages, with 12 / 22 cases belonging to the HPV16 Al sub-lineage, 3 / 22 being D3 sublineage and 2 / 22 being A2 sublineage. Within shared sub-lineage samples, we evaluated SNV patterns in each viral genome demonstrating that 13 / 17 viruses displayed unique SNV fingerprints. Four HPV genomes lacked any SNVs, precluding assignment as unique viruses, meaning 18 / 22 cases were confirmed to contain unique viral genomes (FIG. 17A, FIG.18).

[0201] 15 / 22 HPV-DeepSeek positive cases were treated for their eventual cancer within the MassGeneralBrigham healthcare system and had FFPE tumor tissue blocks which could be retrieved from biopsy at the time of diagnosis. DNA from tumor tissue was subject to HPV-DeepSeek and viral genome fingerprinting was performed to verify the same viral genome in both pre-diagnostic plasma and diagnostic tumor tissue. All plasma-tissue pairs shared the same genome, with matching only to the pair-mate, and to no other viral genome, confirming that the same virus was present from screening detection to diagnosis (FIG. 17B, FIG. 19).

[0202] Longitudinal monitoring from early detection through treatment

[0203] 7 / 28 patients had been enrolled into prospectively liquid biopsy studies at the time of cancer diagnosis unrelated to their pre-diagnostic biobank sample collection (FIG. 20). Patient 1. ctHPVDNA was detected 16 months before diagnosis. ctHPVDNA levels were monitored weekly during chemoradiotherapy (CRT) treatment with clearance at the end ofMEEI 2024-326-02Quarles 169511.00058treatment and remaining zero during monitoring. Patient 3. ctHPVDNA was detected 20 months before diagnosis. ctHPVDNA levels were monitored weekly during CRT treatment with decreasing levels but no clearance. Following conclusion of treatment, ctHPVDNA levels began increasing. The patient was then found to have a second primary HPV malignancy for which they underwent surgery followed by CRT. ctHPVDNA cleared after this treatment but reelevated, indicating recurrence, which was detected by cross-sectional imaging two months later. Patient 8. ctHPVDNA was detected 27 months before diagnosis. ctHPVDNA rapidly cleared following surgery and remained zero during monitoring. Patient 21. ctHPVDNA was not detectable 79 months before diagnosis but was detected at the time of clinical diagnosis. The patient was treated with surgery. ctHPVDNA cleared rapidly with surgery and remained zero during monitoring. Patient 23. ctHPVDNA was detected 91 months before diagnosis. A sample at diagnosis was not available. ctHPVDNA was negative after treatment with surgery and CRT but subsequent monitoring showed an increase, indicating a recurrence. Patient 27. ctHPVDNA was not detectable 125 months before diagnosis, but was detected at the time of diagnosis. Patient 28. ctHPVDNA was not detected 130 months before diagnosis but was detected at diagnosis. The patient underwent weekly ctHPVDNA monitoring during CRT with clearance at the end of treatment and remained zero during monitoring. This figure emphasizes the potential of an integrated screening, diagnosis, and monitoring biomarker.

[0204] Four of these patients had samples collected at the time of clinical cancer diagnosis and into treatment monitoring affording a continuous global view from cancer screening detection through post-treatment monitoring (FIG. 17C). In one of these patients, ctHPVDNA persisted across treatments, predicting recurrence.

[0205] Comparison to other blood-based HPV cancer detection approaches

[0206] Plasma-based HPV Ab, viral integration events and PIK3CA gene mutations were assessed in parallel to ctHPVDNA (FIG. 21 A). 26 / 28 HPV+OPSCC cases were tested for HPV Ab with 20 / 26 having detectable Ab (overall sensitivity 77%). 17 / 26 (65%) of cases had both ctHPVDNA and HPV Ab (FIG. 21B, Table 21). ctHPVDNA was more likely to be detected closer to diagnosis compared to HPV Ab (Diagnostic accuracy <4 years of cancer diagnosis: HPV-DeepSeek Youden Index 1.0 vs HPV Ab Youden Index 0.82; p-value 0.004) (FIG. 14, FIG.MEEI 2024-326-02Quarles 169511.0005815). Viral genome integration into the human genome, a hallmark of HPV cancers was detected by HPV-DeepSeek 2 / 28 patients (FIG. 17A, FIG. 22). Consistent with existing reports demonstrating the challenges of detecting mutated genes in ctDNA for cancer early detection, no pathogenic PIK3CA mutations were detected in the 28 cases. 18 / 28 cases had at least two cancer diagnosis-supporting features detected in the plasma (FIG. 21C).

[0207] Discussion

[0208] Utilizing a novel multi-feature HPV WGS liquid biopsy in a case-control cohort study of 56 participants, we found accurate blood-based early detection of HPV+OPSCC with 79% overall sensitivity at 100% specificity and a maximum lead time from molecular detection to clinical cancer diagnosis of 7.8 years. The addition of a locked machine learning model trained and tested on an independent cohort of 306 cases and controls led to an increased overall sensitivity of 96% and maximum lead time of 10.3 years. These results highlight the enormous potential of ctDNA detection approaches for cancers that currently lack screening tests and specifically highlight the potential for blood-based HPV+OPSCC cancer screening. The findings of this study demonstrate some of the most significant lead times from ctDNA early detection to clinical cancer diagnosis reported for any cancer type, fitting with the presumed natural history of HPV+OPSCC, which is believed to develop over a time course of multiple decades.

[0209] We ensured the robustness and validity of these results through viral molecular fingerprinting, demonstrating that each case harbored a unique viral genome, and that the same viral genome was present in both the screening early detection blood sample and tissue sample from the time of cancer diagnosis. We further compared ctHPVDNA WGS-based detection to HPV Ab, an existing candidate screening biomarker, finding similar overall sensitivity, but significantly improved diagnostic accuracy within four years of cancer diagnosis for ctHPVDNA, addressing a key limitation of HPV Ab-based screening approaches- the long lead time from Ab development to cancer development. ctHPVDNA-based detection holds additional advantages over HPV Ab-based detection including that it is quantitative and dynamic, and thus can be trended across time, and is a disease biomarker, as opposed to a risk biomarker. However, HPV Ab detection is more cost- effective, a significant consideration for cancerMEEI 2024-326-02Quarles 169511.00058screening approaches. These findings, in combination with the overall high sensitivity and specificity of ctHPVDNA detection, suggest that a single, or combinatorial, blood-based screening early detection approach for HPV+OPSCC is feasible, with the goal of shifting diagnosis to earlier stages of disease and thus decreasing treatment-related morbidity and mortality. Necessary next steps prior to a prospective EDRN phase 4 study include defining the appropriate populations for screening, conducting detailed modeling studies to evaluate potential morbidity and mortality benefits and harms, and developing post-blood-based early detection workup algorithms.

[0210] In conclusion, using plasma samples from asymptomatic individuals who went on to develop HPV+OPSCC and healthy population-level control samples, paired with a novel ultrasensitive multi-feature HPV WGS liquid biopsy, we demonstrate sensitive and specific blood-based early detection of HPV+OPSCC extending to a decade prior to clinical cancer diagnosis. These results raise the opportunity for a blood-based screening test for the most common HPV-associated cancer in the US.

[0211] References for Example 2

[0212] 1. U.S. Preventive Services Task Force, Grade A and B, Category Cancer. (uspreventiveservicestaskforce.org / uspstf / topic_search_results?topic_status=All&grades%5B%5 D=A&grades%5B%5D=B&category%5B%5D=l 5&searchterm). .

[0213] 2. Sung, H. et al. Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA Cancer J Clin 71, 209-249 (2021).

[0214] 3. Hubbell, E., Clarke, C. A., Aravanis, A. M. & Berg, C. D. Modeled Reductions in Late-stage Cancer with a Multi-Cancer Early Detection Test. Cancer Epidemiology, Biomarkers & Prevention 30, 460-468 (2021).

[0215] 4. Ahlquist, D. A. Universal cancer screening: revolutionary, rational, and realizable. NPJ Precis Oncol 2, 23 (2018).

[0216] 5. WHO. Guide to cancer early diagnosis. World Health Organization; 2017.MEEI 2024-326-02Quarles 169511.00058

[0217] 6. Siegel, R. L., Miller, K. D., Wagle, N. S. & Jemal, A. Cancer statistics, 2023. CA Cancer J Clin 73, 17-48 (2023).

[0218] 7. Howlader N NA, Krapcho M, Miller D, Bishop K, Kosary CL, Yu M, Ruhl J, Tatalovich Z, Mariotto A, Lewis DR, Chen HS, Feuer EJ, Cronin KA (eds). SEER Cancer Statistics Review, 1975-2014. National Cancer Institute. Bethesda, MD.(https: / / seer.cancer.gov / csr / 1975 2014 / , based on November 2016 SEER data submission, posted to the SEER web site, April 2017.).

[0219] 8. Clarke, C. A. et al. Projected Reductions in Absolute Cancer-Related Deaths from Diagnosing Cancers Before Metastasis, 2006-2015. Cancer Epidemiology, Biomarkers & Prevention 29, 895-902 (2020).

[0220] 9. Chaturvedi, A. K. et al. Human papillomavirus and rising oropharyngeal cancer incidence in the United States. Journal of Clinical Oncology 29, 4294-4301 (2011).

[0221] 10. Zhang, Y, Fakhry, C. & D’Souza, G. Projected Association of Human Papillomavirus Vaccination With Oropharynx Cancer Incidence in the US, 2020-2045. JAMA Oncol 7, e212907 (2021).

[0222] 11. Leshchiner, I. et al. Inferring early genetic progression in cancers with unobtainable premalignant disease. Nat Cancer 4, 550-563 (2023).

[0223] 12. Busch, C.-J. et al. Detection of stage I HPV-driven oropharyngeal cancer in asymptomatic individuals in the Hamburg City Health Study using HPV16 E6 serology - A proof-of-concept study. EClinicalMedicine 53, 101659 (2022).

[0224] 13. Holzinger, D. et al. Sensitivity and specificity of antibodies against HPV16 E6 and other early proteins for the detection of HPV16-driven oropharyngeal squamous cell carcinoma. Int J Cancer 140, 2748-2757 (2017).

[0225] 14. Kreimer, A. R. et al. Timing of HPV16-E6 antibody seroconversion before OPSCC: findings from the HPVC3 consortium. Annals of Oncology 30, 1335-1343 (2019).MEEI 2024-326-02Quarles 169511.00058

[0226] 15. Waterboer, T. et al. Early Detection of Human Papillomavirus- Driven Oropharyngeal Cancer Using Serology From the Study of Prevention of Anal Cancer. JAMA Oncol 6, 1806 (2020).

[0227] 16. Hibbert, J., Halec, G., Baaken, D., Waterboer, T. & Brenner, N. Sensitivity and Specificity of Human Papillomavirus (HPV) 16 Early Antigen Serology for HPV-Driven Oropharyngeal Cancer: A Systematic Literature Review and Meta-Analysis. Cancers (Basel) 13, 3010 (2021).

[0228] 17. Robbins, H. A. et al. Absolute Risk of Oropharyngeal Cancer After an HPV16-E6 Serology Test and Potential Implications for Screening: Results From the Human Papillomavirus Cancer Cohort Consortium. Journal of Clinical Oncology 40, 3613-3622 (2022).

[0229] 18. Brenner, N. et al. Characterization of human papillomavirus (HPV) 16 E6 seropositive individuals without HPV-associated malignancies after 10 years of follow-up in the UK Biobank. EBioMedicine 62, 103123 (2020).

[0230] 19. Lang Kuhs, K. A. et al. Human Papillomavirus 16 E6 Antibodies in Individuals without Diagnosed Cancer: A Pooled Analysis. Cancer Epidemiology, Biomarkers & Prevention 24, 683-689 (2015).

[0231] 20. Siravegna, G. et al. Cell-Free HPV DNA Provides an Accurate and Rapid Diagnosis of HPV-Associated Head and Neck Cancer. Clinical Cancer Research 28, 719— 727 (2022).

[0232] 21. Damerla, R. R. et al. Detection of Early Human Papillomavirus- Associated Cancers by Liquid Biopsy. JCO Precis Oncol 3, (2019).

[0233] 22. Rettig, E. M. et al. Prognostic Implication of Persistent Human Papillomavirus Type 16 DNA Detection in Oral Rinses for Human Papillomavirus-Related Oropharyngeal Carcinoma. JAMA Oncol 1, 907 (2015).

[0234] 23. Rettig, E. M. et al. Detection of circulating tumor human papillomavirus DNA before diagnosis of HPV-positive head and neck cancer. Int J Cancer 151, 1081-1085 (2022).MEEI 2024-326-02Quarles 169511.00058

[0235] 24. Rettig, E. M. et al. Relationship of HPV16 E6 seropositivity with circulating tumor tissue modified HPV16 DNA before head and neck cancer diagnosis. Oral Oncol 141, 106417 (2023).

[0236] 25. Tewari, S. R. et al. Association of Plasma Circulating Tumor HPV DNA With HPV-Related Oropharynx Cancer. JAMA Otolaryngology-Head & Neck Surgery 148, 488 (2022).

[0237] 26. Pepe, M. S. et al. Phases of Biomarker Development for Early Detection of Cancer. JNCI Journal of the National Cancer Institute 93, 1054-1061 (2001).

[0238] 27. Waterboer, T. et al. Multiplex human papillomavirus serology based on in situ-purified glutathione s-transferase fusion proteins. Clin Chem 51, 1845-53 (2005).Example 3: Clinical validation of an HPV whole genome sequencing assay for circulating tumor HPV DNA for molecular residual disease detection in HPV-associated head and neck cancer patients treated with surgery

[0239] Introduction

[0240] Human papillomavirus associated head and neck squamous cell carcinoma (HPV+HNSC) is the most common HPV-associated cancer in the United States, surpassing HPV-associated cervical cancer. Surgery is a common approach for primary treatment of early-stage HPV+HNSCC. Many patients who undergo surgery receive adjuvant radiation or chemoradiation therapy to treat potential residual disease, which is currently predicted based on clinicopathologic risk factors including positive margins, extranodal extension (ENE), multiple positive nodes, vascular invasion, and lymphatic invasion. However, there are limitations in predicting residual disease based on the use of these features alone - the use of clinicopathologic risk factors for prediction is non-standardized and has poor individualized predictive and prognostic capacity. Currently, there are no established biomarkers to predict residual disease.

[0241] Circulating tumor DNA (ctDNA) is an emerging minimally invasive and reliable prognostic biomarker, for detecting molecular residual disease (MRD) and predicting recurrence in multiple solid cancers. Prospective trials in cancers such as colorectal have demonstrated not only strong DFS prognostic capacity but also OS. Previous studies haveMEEI 2024-326-02Quarles 169511.00058demonstrated that HPV+HNSCC release circulating tumor HPV DNA (ctHPVDNA) into the blood where it serves as an accurate real-time biomarker of disease status after surgery. In patients without pathological risk factors, ctHPVDNA rapidly cleared after surgery in the patients. On the other hand, in patients with residual disease, ctHPVDNA remained elevated after surgery, which also correlated with an increased recurrence rate. However, these studies showed the patient with microscopic levels of residual disease with ctHPVDNA negative by dPCR suggesting significantly more sensitive assays are necessary.

[0242] Currently, the most common approaches for detecting ctHPVDNA is digital PCR (dPCR). However, dPCR has several limitations in the context of residual disease monitoring. First, dPCR has relatively low accuracy in MRD settings, where the presence of minimal amounts of ctHPVDNA requires highly sensitive detection methods. The limited sensitivity of dPCR can lead to false-negative results. Second, dPCR is unable to detect the full HPV genome, as it typically targets specific viral genes, such as the E7 oncogene. This limitation makes dPCR ineffective in scenarios where the targeted gene is absent or mutated, resulting in an incomplete assessment of the viral genomic landscape. Furthermore, dPCR assays are typically designed to detect a set number of HPV genotypes, which can be a limitation given multiple high-risk HPV types associated with HNSCC. This hinders the application of dPCR in cases involving less common HPV variants. Next-generation sequencing (NGS) is a newer platform, which offers a highly sensitive approach for detecting and characterizing ctHPVDNA13. NGS has higher analytical sensitivity compared to dPCR, including the ability to detect low-frequency variants and minimal ctHPVDNA levels, which is important for the early identification of residual disease. Unlike dPCR, NGS allows for the annotation of the entire HPV genome, such as tumor-specific mutations, structural variants, and viral integration sites. This provides a more comprehensive molecular profile.

[0243] Data on failure of de-escalation trials highlight the limitations of non-biomarker-driven strategies in managing HPV+HNSCC. For example, de-escalation approaches relying solely on clinicopathologic factors have often failed to adequately stratify patients, resulting in either under-treatment of high-risk individuals or overtreatment of those who may not require adjuvant therapy.MEEI 2024-326-02Quarles 169511.00058

[0244] Our group has developed a custom whole HPV genome hybrid-capture-based next-generation sequencing assay, termed HPV-DeepSeekl4. HPV-DeepSeek has 98.7% sensitivity and 98.7% specificity for the diagnosis of HPV+HNC and is significantly more sensitive than dPCR. HPV-DeepSeek was developed for use in low ctDNA settings such as screening early detection and MRD. Here, we conducted a prospective observational cohort study of 103 HPV+HNC patients treated with curative-intent surgery to test the primary hypothesis that ctHPVDNAMRD detection post-surgery predicts inferior disease-free survival (DFS) and overall survival (OS). We also evaluated the secondary hypothesis that ctHPVDNA detection during surveillance correlates with worse outcomes. We further compared the prognostic performance of HPV-DeepSeek to dPCR, assessed the value of ctHPVDNAMRD relative to traditional clinicopathologic risk factors, and examined lead times from molecular detection to clinical recurrence. This study aims to validate the utility of HPV-DeepSeek in guiding treatment decisions and improving outcomes for HPV+HNSCC patients.

[0245] FIG. 23 A shows a schematic of how HPV-DeepSeek can lead to detecting MRD before clinical relapse. FIG. 23B shows a schematic illustrating the prediction that patients with MRD after treatment prediction have lower survival rates than patients without MRD. FIG.23C shows HPV reads over the course of treatment. Recurrent cases (red) are increased after surgery compared to cases with and without adjuvant therapy.

[0246] Methods

[0247] All patients provided written informed. The protocol was approved by the Dana Farber / Harvard Cancer Center institutional review board. The study is registered on ClinicalTrials.gov (NCT06730412) and was conducted in accordance with the Declaration of Helsinki. All participants provided written informed consent This study was conducted in compliance with the U.S. Common Rule.

[0248] Study design and participants: We conducted a prospective observational cohort study which enrolled 103 AJCC 8 Stage I-III HPV+HNC patients at Massachusetts Eye and Ear between August 2020 and March 2024. Key eligibility criteria include the following: 1) age >18 years, 2) newly diagnosed, untreated, histologically confirmed HPV+HNC, 4) scheduled for curative intent resection as primary treatment, 5) Eastern Cooperative Oncology GroupMEEI 2024-326-02Quarles 169511.00058(ECOG) performance status of 0-2. Blood samples for ctHPVDNA analysis were collected before surgery, in the MRD windows, and in the surveillance window.

[0249] HPV-associated cancers were detected by histomorphology consistent with squamous cell carcinoma on tissue biopsy and pl 6 immunohistochemistry with or without HPV PCR, or HPV RNA in situ hybridization. Adjuvant therapy after surgery was determined by a multi-disciplinary tumor board based on clinicopathologic risk factors. Clinical post-treatment follow-up examinations occurred at 2-to-4-month intervals per NCCN surveillance guidelines and included physical exam and endoscopy. Cross-sectional imaging was obtained at three months and one year post-treatment completion, as well as at the additional discretion of the treating physician. Clinical information was extracted from the Electronic Patient Care Record (EPIC) and input into a study specific electronic data capture system. Three independent reviewers extracted data from surgical pathology reports to ensure accuracy and conflicts were manually re-reviewed to ensure accuracy and consistency.

[0250] Three MRD windows were analyzed: 1) MRD immediately post-surgery (MRD-early) was defined as post-operative days 1-3, 2) MRD post-surgery (MRD-PS) was defined as 4 days-6 months after surgery and before the start of adjuvant therapy, 3) MRD posttreatment completion (MRD-TC) was defined as 4 days-6 months after surgery for patients treated with surgery alone and <6 months after radiation completion for patients treated with adjuvant therapy. The potential clinical application of analysis in these windows include: 1) for MRD-early, immediate prognostication and personalized adjuvant treatment selection, 2) for MRD-PS, presumed optimal prognostication and personalized adjuvant treatment selection timing and 3) for MRD-PT prognostication and personalized post-adjuvant treatment and surveillance. MRD-PS and MRD-PT windows are overlapping for patients treated only surgically. POD 4 was chosen as the earliest time point following surgery for the MRD-PS window as a separate analysis was conducted examining POD 1-3 vs POD 4-14 values in patients treated with surgery in a lead in cohort demonstrating clearance occurred in some patients at the later time points with this more sensitive assay, contrary to earlier data our group had generated examining POD 1 as a landmark timepoint with ddPCR-based analysis. The surveillance window was defined as the time from the end of the MRD-TC window to last followMEEI 2024-326-02Quarles 169511.00058up, recurrence or death. The relevant clinical application is the early detection of recurrence, which could trigger non-routine imaging and / or salvage treatment initiation

[0251] ctHPVDNA testing: A clinically validated multi-feature HPV whole genome sequencing assay was used for detection and quantification of ctHPVDNA in blood samples (HPV-DeepSeek). HPV-DeepSeek uses hybrid capture next generation sequencing to detect and quantitate 43 HPV genotypes, mutations in Pik3CA, viral integration events, high risk prognostic viral SNPs and multiple fragment size features. Detailed HPV-DeepSeek methodology can be found in Bryan et al. In brief, total cfDNA was extracted from 4ml of whole blood as previously described9. 15-20ng of total cell-free DNA was used as input for downstream assays. Samples were considered positive using a pre-defined threshold (10 unique HPV reads and 10% HPV genome coverage), as described previously. ctHPVDNA results were reported as negative or if positive, then by the genotype detected and the number of unique HPV reads. A clinically validated ddPCR assay targeting specific regions of the E7 gene from five high-risk HPV genotypes: 16, 18, 33, 35, and 45 was used for head-to-head comparisons to HPV-DeepSeek. Samples were considered positive using a pre-defined threshold (500 FGFR1 events total and > 2 HPV reads), as described previously.

[0252] Statistical Analysis: The primary endpoints were DFS and OS. The hypothesis tested was that patients with MRD after surgery (MRD-PS window) compared to patients without MRD would have inferior 2-year DFS and OS. The secondary endpoint was 2-year DFS and OS after treatment completion (MRD-TC window) in patients with and without MRD. DFS was defined as the period from the date of surgery or the completion of adjuvant therapy to the date of relapse diagnosis. OS was defined as the time from the date of surgery or the completion of adjuvant therapy to the date of death from any cause. The Kaplan-Meier method was used to estimate the survival distribution. Differences between groups were tested using the log-rank test. When multiple samples were taken within a designated window, any positive same was regarded as MRD positive. Disease recurrence required confirmatory tissue biopsy.

[0253] The exploratory endpoints included: (1) 2-year DFS) and OS) in patients with and without MRD in the MRD-early window, (2) detection of molecular recurrenceMEEI 2024-326-02Quarles 169511.00058compared to clinical detection of recurrence, (3) sensitivity and lead times of HPV-DeepSeek compared to ddPCR, and (4) identification of key features driving recurrence risk. Sensitivity and lead times between HPV-DeepSeek and ddPCR were compared using a McNemar’s test. Prognostic factors associated with DFS were evaluated using a multivariable Cox proportional hazard model. The ability to predict recurrence and clinicopathologic risk factors using HPV-DeepSeek outputs at the diagnostic (pre-treatment) time point was assessed through machine learning models.

[0254] To predict recurrence and clinicopathologic risk factors, we implemented a machine learning pipeline focused on binary classification tasks. Preprocessing included encoding categorical variables through custom-defined mappings and one-hot encoding. Features used in the models included genotype, high-risk SNPs, PIK3C A mutation status, HPV integration status, sequencing reads, HPV mutation count, coverage metrics, and fragment length statistics. Data were standardized using StandardScaler, and class imbalance was addressed with SMOTE to ensure balanced representation of recurrence outcomes in the training data. Multiple classifiers were trained and evaluated including Decision Trees, Balanced Random Forest, Extra Trees, and XGBoost. Hyperparameter optimization was performed using RandomizedSearchCV with 3-fold stratified cross-validation to identify the best-performing configurations. Model evaluation focused on metrics including accuracy, precision, recall, Fl -score, and AUG, with bootstrapped 95% confidence intervals calculated for each metric. Model performance was evaluated through feature importance analyses, using both permutation importance and intrinsic importance measures (e.g., feature importances for tree-based models and coefficients for linear models). Confusion matrices and ROC curves were generated to compare the predictive performance of selected models. All analyses were conducted using Python v3.11.4 in Jupyter Notebook v6.5.4, with key libraries including scikit-learn vl.3.0, imbalanced-learn vO.10.1, and XGBoost vl.7.5.

[0255] Results

[0256] Patient cohort and characteristics

[0257] 103 patients were enrolled over a 44-month period starting 8 / 18 / 2020 and ending 3 / 14 / 2024. A total of 560 samples were collected and processed (mean of 5.6MEEI 2024-326-02Quarles 169511.00058samples / patient). The mean age was 62 years (range 36-84). The cohort was predominately male (89%: 92 / 103), AJCC8 stage I (89%: 92 / 103) and oropharyngeal subsite (93%: 96 / 103), with 2 unknown primary, 2 nasal cavity and 2 oral cavity case included (Table 22). HPV16 was the dominant genotype (88%: 89 / 101), with 5 additional genotypes detected (FIG. 24A). All patients underwent curative intent resection as primary treatment.<>MEEI 2024-326-02Quarles 169511.00058MEEI 2024-326-02Quarles 169511.00058>MEEI 2024-326-02Quarles 169511.00058

[0258] Table 22: pre-operative and post-operative patient demographics for Example

[0259] ctHPVDNA detection with HPV-DeepSeek at clinical diagnosis

[0260] The mean pre-operative ctHPVDNA level was 32,617 (range 8-495,758), with a mean HPV genome coverage of 85.3% (range 8.5%-100%). The sensitivity of HPV-DeepSeek was 98.1% (101 / 103). 92 cases had both HPV-DeepSeek and ddPCR performed with sensitivities of 99% (91 / 92) and 89% (82 / 92), respectively. HPV-DeepSeek positive, ddPCR negative samples had low ctHPVDNA levels or genotypes outside the ddPCR panel, leading to significantly improved sensitivity in HPV-DeepSeek vs ddPCR (p=0.004) (FIG. 24B). There was a linear correlation between ddPCR and HPV-DeepSeek reads (Pearson correlation coefficient (r)=0.86, P<0.001) (FIG. 24C). MRD cohort and characteristics

[0261] MRD cohort and characteristics

[0262] HPV-DeepSeek negative cases at enrollment were excluded from subsequent analysis (n=2). One additional patient was excluded from the survival analysis due to follow up <6 months leaving 100 patients for survival analysis. 73% (74 / 101) of patients received adjuvantMEEI 2024-326-02Quarles 169511.00058therapy after surgery: 50% (50 / 101) received radiation therapy and 24% (24 / 101) received chemoradiotherapy (FIG. 24D). Seven patients experienced a recurrence, and one patient was found to have a second primary during the surveillance period. 4 / 8 were distant metastases, 3 / 8 were locoregional recurrences and 1 / 8 was a second primary HPV+HNSCC (FIG. 24E). There were six deaths. The mean, median and range follow up was 791 days, 796 days, 138-1376 days.

[0263] ctHPVDNA clearance dynamics following surgery and adjuvant treatment

[0264] Patients were grouped into four categories based on clearance kinetics: (1) those who became MRD negative in the MRD-early window 37 / 101 (37%), (2) those who became MRD negative in the MRD-PS window 32 / 101 (32%), (3) those who became MRD negative after adjuvant treatment 12 / 101 (12%), and (4) those who remained MRD positive following adjuvant treatment 6 / 101 (6%) (FIG. 24F). We evaluated how baseline ctHPVDNA values related to rate of clearance. The rapid clearance group (cleared by the MRD-early window) had significantly lower baseline ctHPVDNA values than those who cleared more slowly (p = 0.0037) and those who cleared only after adjuvant treatment (p = 0.0001). The slope of decline was not significantly different between the rapid and slow clearance groups (p = 0.075) but was significantly steeper compared to those who cleared only after adjuvant treatment (p < 0.05), (Tables 23A-23C, FIG. 25). Taken together, these findings indicate that initial ctHPVDNA values influence the time to clearance and the slope of clearance is also associated with likelihood of residual disease and thus likelihood of requiring adjuvant therapy. Association of ctHPVDNA status in MRD windows with recurrence and survival.

[0265] Table 23 A. Slope Statistics by Treatment Group. Table summarizes the dynamics of ctHPVDNA clearance rates across treatment groups, reporting the Median, Mean, Standard Deviation, and Sample Count of slope values. It includes Kruskal-Wallis H-test results (H-statistic and p-value) to assess group differences.MEEI 2024-326-02Quarles 169511.00058> <<

[0266] Table 23B. ctHPVDNA Read Counts by Treatment Transitions: Table shows ctHPVDNA read count distributions for treatment transitions, detailing the Median, Mean, Standard Deviation, and Sample Count for each group. Kruskal-Wallis test results evaluate differences in read counts across groups.>>

[0267] Table 23C. Percentage Change and Slope Analysis: Table highlights percentage changes and slope values of ctHPVDNA clearance during treatment transitions, with Change (%), Slope, and Sample Count for each group. Kruskal-Wallis test results determine the statistical significance of differences among groups.

[0268] Association of ctHPVDNA status in MRD windows with recurrence and survival

[0269] ctHPVDNA status was available for 75 patients during the MRD-PS window, the primary outcome assessment. 17 / 75 patients (23%) were MRD positive with six having a recurrence. 58 / 75 patients (77%) were MRD negative with one having a recurrence (HR: 25.0,MEEI 2024-326-02Quarles 169511.0005895% CI: 3.0-208.3, p<0.001), demonstrating a 2-year DFS of 60% (95% CI: 31-80) vs 100% respectively (HR: 25.0, 95%CI 3-208.3, P<0.001) (FIG. 26A). The association of MRD positivity with a significantly increased risk for recurrence was observed for both the earliest stages of disease (T0-1N0-1M0) DFS (80% (95%CI: 20-97) vs 100%, P -0.046) and later stages (55% (95%CI: 23-78) vs 100%, HR: 17.7, 95% CI: 2.0-153.9, P < 0.001) as well as when restricted to only oropharynx cancer (DFS 55% (95%CI: 26-77) vs 100%, HR: 27.7, 95% CI: 3.3-231.3, P < 0.001). MRD positivity was also found to be significantly associated with worse OS compared to MRD negative patients (HR: 13.6, 95% CI: 1.5-121.4, P = 0.002), demonstrating a 2-year OS of 73% (95% CI: 43-89) vs 98% (95% CI: 88-100) respectively (FIG.26B). The association of MRD positivity with increased risk of death was observed for the later stages (60% (95%C1: 24-83) vs 100%, P = 0.001) as well as when restricted to only oropharynx cancer OS (70% (95%CI: 38-88) vs 98% (95%CI: 88-100), HR: 14.8, 95% CI: 1.7-132.7, P = 0.001). There was no association with OS at the earliest stages of disease (T0-1N0-1M0) (100% vs 98% (95%CI:85-100), P=0.65). Analysis of only the earliest timepoint in the window, showed slightly lower HRs, but results were largely unchanged (FIG. 27A-27D), suggesting multitimepoint sampling within a window increases diagnostic accuracy (Table 24).

[0270] Table 24. Survival analysis using MRD status by HPV-DeepSeek. Results were categorized based on two approaches within the MRD window: (1) 'Single' — using only the earliest ctHPV DNA result, and (2) 'Multiple' — using the highest ctHPV DNA level from multiple results within the window. DFS, OS, and their respective HRs are shown.MEEI 2024-326-02Quarles 169511.00058

[0271] ctHPVDNA status was available for 66 patients during the MRD-TC window, the secondary outcome assessment. 6 / 66 (9%) were MRD positive with six (100%) having a recurrence. 60 / 66 (91%) were MRD negative with 0% (0 / 60) having a recurrence, demonstrating a 2-year DFS of 0% vs 100% respectively (P<0.001) (FIG. 26C). The association of MRD positivity with a significantly increased risk for recurrence was observed for both the earliest stages of disease (T0-1N0-1M0) DFS (0% vs 100%, P < 0.001) and later stages (0% vs 100%, P < 0.001) as well as when restricted to only oropharynx cancer (DFS (0% vs 100%, P < 0.001). MRD positivity was also found to be significantly associated with worse OS compared to MRD negative patients (P < 0.001), demonstrating a 2-year OS of 50% (95% CI: 11-80) vs 100% respectively (FIG. 26D). The association of MRD positivity with a significantly increased risk for death was observed for the later stages (OS 40% (95%CI: 5-75) vs 100%, P < 0.001) as well as when restricted to only oropharynx cancer OS (50% (95%CI: 11-80) vs 100%, P < 0.001). In the earliest stages of disease (T0-1N0-1M0), there were no deaths regardless of MRD status.

[0272] ctHPVDNA status was available for 89 patients during the MRD-early window, an exploratory outcome. 48 / 89 (54%) were MRD positive with seven having a recurrence. 41 / 89 (46%) were MRD negative with one having a recurrence, demonstrating a 2-year DFS of 87% (95% CI: 73-94) vs 96% (95% CI: 76-99) respectively (HR: 6.2, 95% CI: 0.8-50.7, p=0.052) (FIG. 28A). MRD positivity was also found to be not significantly associated with worse OS compared to MRD negative patients (HR: 3.9, 95% CI: 0.5-33.7, P = 0.18), demonstrating a 2-year OS of 88% (95% CI: 74-95) vs 97% (95% CI: 81-100) respectively (FIG.28B). These results suggest that the MRD-early window is not a reliable predictor of DFS or OS.

[0273] We conducted a multivariate analysis to evaluate the prognostic significance of ctDNA status within the MRD-PS and MRD-TC windows. In the MRD-PS window, MRD+ was strongly associated with worse DFS, with an adjusted hazard ratio (HR) of 10.76 (95% CI: 1.46-79.10, P = 0.02), compared to MRD- patients (FIG. 29A). Traditional clinicopathologic risk factors, including sex, age, smoking status, extranodal extension (ENE), margins, pT stage, and pN stage, were not significantly associated with DFS in this window. In the MRD-TC window, MRD+ demonstrated an even stronger association with worse DFS, with an HR of 272.74 (95% CI: 4.75-15,668.32, P = 0.007) (FIG. 29B). Similar to the MRD-PS window, traditional riskMEEI 2024-326-02Quarles 169511.00058factors such as ENE, positive margins, and smoking status did not achieve statistical significance.

[0274] MRD status and benefit from adjuvant therapy

[0275] We examined the impact of adjuvant treatment on MRD+ and MRD- patients in the MRD-early and MRD-PS window, after adjusting for stage. In the MRD-early window, 54% (48 / 89) of patients were MRD-positive, of whom 83% (40 / 48) received adjuvant treatment, and 46% (41 / 89) of patients were MRD-negative, of whom 63% (26 / 41) received adjuvant treatment. MRD-positive patients did not demonstrate a statistically significant benefit from adjuvant treatment (adjusted HR: 1.38, 95% CI: 0.16-11.67, P = 0.77; recurrence rate: 15% (6 / 40) for the adjuvant therapy group versus 13% (1 / 8) for no-adjuvant therapy). No benefit was observed for MRD-negative patients (P = 0.996; clinical recurrence rate: 0% (0 / 26) for adjuvant therapy versus 7% (1 / 15) for no adjuvant therapy).

[0276] In the MRD-PS window, 23% (17 / 75) of patients were MRD positive, of whom 88% (15 / 17) received adjuvant treatment and 77% (58 / 75) of patients were MRD negative, of whom 64% (37 / 58) received adjuvant treatment. MRD-positive patients benefited from adjuvant treatment (adjusted HR: 0.162, 95% CI: 0.026-1.009, P = 0.051; recurrence rate: 27% (4 / 15) for the adjuvant therapy group versus 100% (2 / 2) for the no-adjuvant therapy group). No benefit was observed for MRD-negative patients (P = 0.997; clinical recurrence rate: 3% (1 / 37) for the adjuvant therapy group versus 0% (0 / 21) for the no-adjuvant therapy group).

[0277] We assessed if ctHPVDNA clearance following adjuvant treatment (MRD- TC and Surveillance windows) was predictive of efficacy as well as outcomes. Of the 17 patients who were MRD positive in the MRD-PS window, patients who cleared ctHPVDNA after adjuvant treatment had superior DFS and OS (DFS: 100% vs 0%, P<0.001; OS: 100% vs 50% (95%CI: 11.1-80.4), P=0.039). HPV-DeepSeek versus ddPCR for detecting MRD and recurrence

[0278] HPV-DeepSeek versus ddPCR for detecting MRD and recurrenceWe compared the sensitivity and leads times from molecular detection of recurrence to clinical detection of recurrence of HPV-DeepSeek and current standard of care HPV liquid biopsyMEEI 2024-326-02Quarles 169511.00058(ddPCR). To evaluate sensitivity, we examined patients with biopsy confirmed recurrence (7) and utilized MRD status in the MRD-PS window. HPV-DeepSeek Sensitivity was 83% (5 / 6) vs 33% (2 / 6) with ddPCR (P = 0.08) (FIG. 30A). As shown in FIG. 30A, detection of MRD after surgery with HPV-DeepSeek may be 50% more sensitive than detection with dPCR. When restricting to only the first timepoint, if multiple timepoints were samples in the window, Sensitivity was 67% (4 / 6) vs 33% (2 / 6), respectively. Utilizing the MRD-TC window, HPV-DeepSeek sensitivity was 100% (5 / 5) vs 60% (3 / 5) with ddPCR (P = 0.114) (FIG. 30B). As shown in FIG. 30B, detection of MRD with HPV-DeepSeek after treatment completion may be 24% more sensitive than detection with dPCR. When restricting to only the first timepoint, Sensitivity was 60% (3 / 5) vs 40% (2 / 5), respectively. The mean time to biopsy proven recurrence in the cohort was 447 days. To measure lead time, we utilized the first positive sample beginning in the MRD-PS window. The mean lead time was 253 days (range 240-533) by HPV-DeekSeek and 52 days (range 0-421) by ddPCR (FIG. 30C, Table 25).Table 25: Lead time for ctHPVDNA detection in recurrent cases by HPV-DeepSeek and ddPCR assay.MEEI 2024-326-02Quarles 169511.00058

[0279] Viral genome fingerprinting for differentiating primary tumor recurrence from second primary and metastatic tumors

[0280] We performed viral genome fingerprinting across all patients with longitudinal positive samples after treatment completion (n=7). Three patients had locoregional recurrences, three patients had distant recurrences, and one patient had a second primary HPV-associated cancer (HPV+ sinonasal SCC) (FIGS. 31A-31G). Five patients had identical viral genome fingerprints across all samples within a patient. Two patients had the emergence of new viral genome fingerprints. The first patient had a 6 SNP fingerprint present at diagnosis which was replaced by a new fingerprint by POD 125. An admixture of these fingerprints was seen in the post-operative window (FIG. 32A). This patient was clinically diagnosed with liver metastases on POD 232, supporting the fingerprint evolution represented the regression of the primary tumor with treatment and the emergence of the metastatic lesion, differentiated by the unique viral SNPs within tumor arising from a viral subclone. In a second patient, a four SNP fingerprint was present at diagnosis which was replaced by a new fingerprint by POD 32. An admixture of these fingerprints was seen in the post-operative window (FIG. 32B). A second primary cancer was diagnosed in the nasal cavity on POD 112, supporting the new unique viral genome. This was validated using tumor tissue sequencing from both the first and second primary cancer.

[0281] Application of machine learning model to predict need for adjuvant therapy and risk of recurrence prior to treatment initiation

[0282] We evaluated if machine learning using pre-treatment HPV-DeepSeek outputs could predict high-risk pathological features and recurrence. We assessed multiple models for their ability to predict recurrence and the key high-risk pathological features of extranodal extension (ENE). Balanced Random Forest was selected for assessment because it effectively addresses class imbalance by balancing the dataset through under-sampling the majority class, ensuring fair representation of minority outcomes. It consistently outperformed other models, such as Decision Trees and XGBoost, in terms of AUC, recall, and precision during cross-validation. The model achieved AUCs of 0.798, 95% CI: [0.560, 0.982]) for ENE (FIG. 33A) and 0.875 (95% CI: [0.667, 1.000]) for recurrence (FIG. 33C). Interestingly featureMEEI 2024-326-02Quarles 169511.00058importance analyses identified that for ENE, fragmentomic features factored in heavily (FIG. 33B) while for recurrence, high-risk SNPs, integration status, and genotype, all known prognostic features, were more influential (FIG. 33D). These findings highlight the predictive utility of both genomic and fragmentomic features above base HPV-DeepSeek features used to classify MRD+ vs - samples. Feature selection and importance analysis identified the most influential predictors for each outcome.

[0283] References for Example 3

[0284] 1) Centers for Disease Control and Prevention. Cancers Associated with Human Papillomavirus. Centers for Disease Control and Prevention, U.S. Department of Health and Human Services; 2024.

[0285] 2) Martel Cd, Georges D, Bray F et al. Global burden of cancer attributable to infections in 2018: a worldwide incidence analysis. Lancet Glob Health. 2020; 8: el80-el90.

[0286] 3 ) Cracchiolo JR, Baxi SS, Morris LG et al. Increase in primary surgical treatment of T1 and T2 oropharyngeal squamous cell carcinoma and rates of adverse pathologic features: National Cancer Data Base. Cancer. 2016; 122: 1523-32.

[0287] 4) Zhan KY, Puram SV, Li MM et al. National treatment trends in human papillomavirus-positive oropharyngeal squamous cell carcinoma. Cancer. 2020; 126: 1295-1305.

[0288] 5) Head and Neck Cancers (version 4.2024) National Comprehensive Cancer Network (NCCN). Accessed October 6, 2024. ccn.org / professionals / physician_gls / pdf / head-and-neck.pdf.

[0289] 6) Kotani D, Oki E, Nakamura Y et al. Molecular residual disease and efficacy of adjuvant chemotherapy in patients with colorectal cancer. Nat Med. 2023; 29 : 127-134.

[0290] 7) Han K, Zou J, Zhao Z, et al. Clinical Validation of Human Papilloma Virus Circulating Tumor DNA for Early Detection of Residual Disease After Chemoradiation in Cervical Cancer. J Clin Oncol. 2024; 42 :431-440.MEEI 2024-326-02Quarles 169511.00058

[0291] 8) Kurtz DM, Scherer F, Jin MC et al. Circulating Tumor DNA Measurements As Early Outcome Predictors in Diffuse Large B-Cell Lymphoma. J Clin Oncol.2018 ;36 :2845-2853.

[0292] 9) Chera BS, Kumar S, Shen C et al. Plasma Circulating Tumor HPV DNA for the Surveillance of Cancer Recurrence in HPV-Associated Oropharyngeal Cancer. J Clin Oncol. 2020; 38: 1050-1058.

[0293] 10) O’Boyle CJ, Siravegna G, Varmeh S et al. Cell-free human papillomavirus DNA kinetics after surgery for human papillomavirus-associated oropharyngeal cancer. Cancer. 2022; 128 :2193-2204.

[0294] 11) Naegele S, Das D, Hirayama S et al. Circulating Tumor HPV DNA in Patients With Stage I and II HPV-Associated Head and Neck Cancer After Surgery. JAMA Otolaryngol Head Neck Surg. 2024 ;150: 521-523.

[0295] 12) Paolini F, Campo F, Locca O et al. It is time to improve the diagnostic workup of oropharyngeal cancer with circulating tumor HPV DNA: Systematic review and metaanalysis. Head Neck. 2023 ; 45: 2945-2954.

[0296] 13) Naegele S, Ruiz-Torres DA, Zhao Y et al. Comparing the Diagnostic Performance of Quantitative PCR, Digital Droplet PCR, and Next-Generation Sequencing Liquid Biopsies for Human Papillomavirus-Associated Cancers. J Mol Diagn. 2024; 26: 179-190

[0297] 14) Das D, Hirayama S, Aye L et al. Circulating tumor HPV DNA whole genome sequencing enables HPV-associated oropharynx cancer early detection. Submitted.

[0298] 15) Bryan ME, Aye L, Das D et al. Direct comparison of alternative bloodbased approaches for early detection and diagnosis of HPV-associated head and neck cancers. Submitted

[0299] 16) Singh R, Song S. Clinical outcomes following observation, postoperative radiation therapy, or post-operative chemoradiation for HPV-associated oropharyngeal squamous cell carcinomas. Oral Oncol. 2023; 146: 106493.

[0300] 17) Rosenberg A, Vokes EE. Optimizing Treatment De-Escalation in Head and Neck Cancer: Current and Future Perspectives. Oncologist. 2021; 26: 40-48.MEEI 2024-326-02Quarles 169511.00058

[0301] 18) Che L, Cohen M, Hatzoglou et al. Early Disease Recurrence Following Post-operative HPV ctDNA Directed Active Surveillance in Oropharyngeal Carcinoma -Outcomes of a Prospective Pilot Study. Int. J Radiat Oncol Biol Phys. 2024; 118: e91.

[0302] 19) Campo F, Locca O, Paolini F et al. The landscape of circulating tumor HPVDNAand TTMV-HPVDNA for surveillance of HPV-oropharyngeal carcinoma: systematic review and meta-analysis. J Exp Clin Cancer Res. 2024; 43: 215.

[0303] 20) Widman AJ, Shah M, Frydendahl A et al. Ultrasensitive plasma-based monitoring of tumor burden using machine-learning-guided signal enrichment. Nat Med. 2024; 30: 1655-1666.

[0304] Example 4: Illustrative Embodiments of Methods and Systems Described Herein

[0305] FIG. 34 shows an example process 3400 to determine a cancer score. At step 3402, a subject sample is received. The subject sample may include ctDNA. In some embodiments, the subject sample include sequences of ctHPVDNA. At step 3404, the sample is provided to an analysis pipeline. The analysis pipeline may detect and synthesize a plurality of features. At step 3406, a cancer score may be determined based on the synthesized plurality of features, subject sample may be classified by the trained network. The classification may be based on the annotated and integrated features.

[0306] In FIG. 35, an example 3500 of a system (e.g., a data processing system) for characterizing a protein in accordance with some embodiments of the disclosed subject matter is shown.

[0307] In some embodiments, computing device 3504 and / or server 3516 can be any suitable computing device or combination of devices, such as a desktop computer, a laptop computer, a smartphone, a tablet computer, a wearable computer, a server computer, a virtual machine being executed by a physical computing device, etc. As described herein, system 3500 can present information about the characterized protein to a user (e.g., a researcher and / or a physician).MEEI 2024-326-02Quarles 169511.00058

[0308] In some embodiments, communication network 3502 can be any suitable communication network or combination of communication networks. In some embodiments, communication network 3502 can be any suitable communication network or combination of communication networks. For example, communication network 3502 can include a Wi-Fi network (which can include one or more wireless routers, one or more switches, etc.), a peer-to-peer network (e.g., a Bluetooth network), a cellular network (e.g., a 4G network, a 5G network, etc., complying with any suitable standard, such as CDMA, GSM, LTE, LTE Advanced, WiMAX, etc.), a wired network, etc. In some embodiments, communication network 3502 can be a local area network, a wide area network, a public network (e.g., the Internet), a private or semi-private network (e.g., a corporate or university intranet), any other suitable type of network, or any suitable combination of networks. Communications links shown in FIG. 35 can each be any suitable communications link or combination of communications links, such as wired links, fiber optic links, Wi-Fi links, Bluetooth links, cellular links, etc.

[0309] FIG. 35 additionally shows an example of hardware that can be used to implement computing device 3504 and server 3516 in accordance with some embodiments of the disclosed subject matter. In some embodiments, computing device 3504 can be used to execute one or more set of instructions to identify a behavioral catalog. In other embodiments, computing device 3504 can be used to identify therapeutic interventions. In still other embodiments, computing device 3504 can be used to identify a configuration of parameter of a gene regulatory network to perform a desired function.

[0310] As shown in FIG. 35, computing device 3504 can include one or more hardware processor 3506, one or more displays 3508, one or more inputs 3510, one or more communications 3512, and / or memory 3514. In some embodiments, processor 3506 can be any suitable hardware processor or combination of processors, such as central processing unit, a graphics processing unit, etc. In some embodiments, display 3508 can include any suitable display devices, such as a computer monitor, a touchscreen, a television, etc. In some embodiments, inputs 3510 can include any suitable input device and / or sensors that can be used to receive user input, such as a keyboard, a mouse, a touchscreen, a microphone, etc.MEEI 2024-326-02Quarles 169511.00058

[0311] In some embodiments, communication systems 3512 can include any suitable hardware, firmware, and / or software for communicating information over communication network 3502 and / or any other suitable communication networks. For example, communications systems 3512 can include one or more transceivers, one or more communication chips and / or chip sets, etc. In a more particular example, communications systems 3512 can include hardware, firmware and / or software that can be used to establish a Wi-Fi connection, a Bluetooth connection, a cellular connection, an Ethernet connection, etc.

[0312] In some embodiments, memory 3514 can include any suitable storage device or devices that can be used to store instructions, values, etc., that can be used, for example, by processor 3506 to present content using display 3508, to communicate with server 3516 via communications system(s) 3512, etc.

[0313] Memory 3514 can include any suitable volatile memory, non-volatile memory, storage, or any suitable combination thereof. For example, memory 3514 can include RAM, ROM, EEPROM, one or more flash drives, one or more hard disks, one or more solid state drives, one or more optical drives, etc. In some embodiments, memory 3514 can have encoded thereon a computer program for controlling operation of computing device 3504. In such embodiments, processor 3506 can execute at least a portion of the computer program to present content (e.g., images, user interfaces, graphics, tables, etc.), receive content from server 3516, transmit information to server 3516, etc.

[0314] In some embodiments, server 3516 can include a processor 3518, a display 3520, one or more inputs 3522, one or more communications systems 3524, and / or memory 3526. In some embodiments, processor 3518 can be any suitable hardware processor or combination of processors, such as a central processing unit, a graphics processing unit, etc. In some embodiments, display 3520 can include any suitable display devices, such as a computer monitor, a touchscreen, a television, etc. In some embodiments, inputs 3522 can include any suitable input devices and / or sensors that can be used to receive user input, such as a keyboard, a mouse, a touchscreen, a microphone, etc.

[0315] In some embodiments, communications systems 3524 can include any suitable hardware, firmware, and / or software for communicating information overMEEI 2024-326-02Quarles 169511.00058communication network 3502 and / or any other suitable communication networks. For example, communications systems 3524 can include one or more transceivers, one or more communication chips and / or chip sets, etc. In a more particular example, communications systems 3524 can include hardware, firmware and / or software that can be used to establish a WiFi connection, a Bluetooth connection, a cellular connection, an Ethernet connection, etc.

[0316] In some embodiments, memory 3526 can include any suitable storage device or devices that can be used to store instructions, values, etc., that can be used, for example, by processor 3518 to present content using display 3520, to communicate with one or more computing devices 3504, etc. Memory 3526 can include any suitable volatile memory, nonvolatile memory, storage, or any suitable combination thereof. For example, memory 3526 can include RAM, ROM, EEPROM, one or more flash drives, one or more hard disks, one or more solid state drives, one or more optical drives, etc. In some embodiments, memory 3526 can have encoded thereon a server program for controlling operation of server 3516. In such embodiments, processor 3518 can execute at least a portion of the server program to transmit information and / or content (e g., results of a tissue identification and / or classification, a user interface, etc.) to one or more computing devices 3504, receive information and / or content from one or more computing devices 3504, receive instructions from one or more devices (e.g., a personal computer, a laptop computer, a tablet computer, a smartphone, etc.), etc.

[0317] In some embodiments, any suitable computer readable media can be used for storing instructions for performing the functions and / or processes described herein. For example, in some embodiments, computer readable media can be transitory or non-transitory. For example, non-transitory computer readable media can include media such as magnetic media (such as hard disks, floppy disks, etc.), optical media (such as compact discs, digital video discs, Blu-ray discs, etc ), semiconductor media (such as RAM, Flash memory, electrically programmable read only memory (EPROM), electrically erasable programmable read only memory (EEPROM), etc.), any suitable media that is not fleeting or devoid of any semblance of permanence during transmission, and / or any suitable tangible media. As another example, transitory computer readable media can include signals on networks, in wires, conductors, optical fibers, circuits, or any suitable media that is fleeting and devoid of any semblance of permanence during transmission, and / or any suitable intangible media.MEEI 2024-326-02Quarles 169511.00058

[0318] A number of references to patent and non-patent documents are made throughout the publication, each of which is herein incorporated by reference in its entirety.

[0319] Thus, while the invention has been described above in connection with particular embodiments and examples, the invention is not necessarily so limited, and that numerous other embodiments, examples, uses, modifications and departures from the embodiments, examples and uses are intended to be encompassed by the claims attached hereto.

Claims

MEEI 2024-326-02Quarles 169511.00058CLAIMSWhat is claimed is:

1. A method of determining a cancer score, comprising:receiving a sample collected from a subject, wherein the sample comprises sequences of circulating tumor DNA (ctDNA);providing the sample to an analysis pipeline, wherein the analysis pipeline detects and synthesizes a plurality of features;determining, with the analysis pipeline, the cancer score based on the synthesized plurality of features.

2. The method of claim 1, wherein the ctDNA comprises viral ctDNA.

3. The method of claim 1 or 2, wherein the ctDNA comprises circulating tumor HPV DNA (ctHPVDNA).

4. The method of any one of the preceding claims, wherein the sample further comprises HPV antibodies of the subject.

5. The method of any one of the preceding claims, wherein the plurality of features comprises at least two of HPV reads as determined by unique molecular indices, HPV genome coverage, HPV genotype, HPV lineage and sublineage, normalized HPV genomes / human genomes, viral integration events, select human genes and gene mutations, ctHPVDNA fragment size features, prognostic viral single nucleotide polymorphisms, or prognostic human germline single nucleotide polymorphisms.

6. The method of any one of the preceding claims, wherein the plurality of features further comprises HPV antibodies.MEEI 2024-326-02Quarles 169511.000587. The method of any one of the preceding claims, wherein the plurality of features further comprises patient demographics, comprising at least one of sex, age, HIV status, or smoking status.

8. The method of claim 5, wherein the HPV genotype comprises identifying the HPV genotype as one of more than 13 HPV genotypes.

9. The method of claim 5, wherein the viral integration events comprise identifying viral genome breakpoints and human genome breakpoints.

10. The method of claim 5, wherein the select gene mutations comprise PIK3CA mutations.

11. The method of any one of the preceding claims, wherein the analysis pipeline comprises at least one trained network.

12. The method of claim 11, wherein the at least one trained network comprises a multi-label, multi-class network.

13. The method of claim 11 or claim 12, wherein an architecture of the trained network comprises at least one of a random forest, extra tress, decision trees, gradient boosting, XGBoost, LightGBM, linear, logistic regression, ridge, lasso, elastic net, kernel-based, support vector machine, instance-based, k-nearest neighbor, probabilistic, or naive bayes.

14. The method of any one of claims 11-13, wherein the at least one network comprises a first network trained on sequencing data, and a second network trained on sequencing data and HPV antibody data.

15. The method of any one of claims 11-14, wherein the trained network was trained on labeled subject data comprising control samples collected from subjects who have not beenMEEI 2024-326-02Quarles 169511.00058diagnosed with a cancer and samples collected from subjects who have been diagnosed with a cancer.

16. The method of any one of the preceding claims, wherein the cancer score is used for at least one of screening, diagnosis, detecting molecular residual disease, prognosis, or surveillance.

17. The method of any one of the preceding claims, wherein the cancer score is used longitudinally for the subject for more than one of screening, diagnosis, detecting molecular residual disease, or surveillance.

18. The method of claim 16 or claim 17, wherein the subject does not display any symptoms indicative of cancer at a time of screening.

19. The method of claim 16 or claim 17, wherein the subject displays one or more symptoms indicative of cancer at a time of screening.

20. The method of any one of claims 16-19, wherein the sensitivity of the screening is at least 25-fold more sensitive than dPCR.

21. The method of any one of claims 16-20, wherein the sensitivity of the screening is at least 80-fold more sensitive than dPCR.

22. The method of any one of the preceding claims, wherein the method is used to diagnose a subject.

23. The method of claim 22, wherein the sensitivity of the diagnosis is greater than 98% at a specificity greater than 98%.MEEI 2024-326-02Quarles 169511.0005824. The method of claim 22 or claim 23, wherein the method is able to detect a disease more than three years and a half years before disease expression.

25. The method of any one of the preceding claims, wherein the method is able to detect a disease more than twenty months before disease expression.

26. The method of any one of the preceding claims, wherein the method is used to detect molecular residual disease of the subject.

27. The method of claim 26, wherein the method enables detection of molecular residual disease more than 150 days before disease expression.

28. The method of claim 26 or claim 27, wherein the detection of the molecular residual disease has a sensitivity greater than 50%.

29. The method of any one of claims 26-28, wherein the detection of the molecular residual disease has a sensitivity greater than 75%.

30. The method of any one of claims 26-29, wherein the detection of the molecular residual disease after surgery is 50% more sensitive than dPCR.

31. The method of any one of claims 26-30, wherein the detection of the molecular residual disease after treatment completion is 25% more sensitive than dPCR.

32. The method of any one of claims 16-30, wherein a clinical decision is made based on the results of the screening, detection of MRD, or surveillance.

33. The method of any one of the preceding claims, wherein the sample is a blood sample.

34. A system for determining a cancer score, comprising:MEEI 2024-326-02Quarles 169511.00058a processor in communication with a memory, the memory having stored thereon a set of instructions, which, when executed by the processor, cause the processor to:receive a sample collected from a subject, wherein the sample comprises sequences of circulating tumor DNA (ctDNA);provide the sample to an analysis pipeline, wherein the analysis pipeline detects and synthesizes a plurality of features;determine, with the analysis pipeline, the cancer score based on the synthesized plurality of features.

35. The system of claim 34, wherein the ctDNA comprises viral ctDNA.

36. The system of claim 34 or 35, wherein the ctDNA comprises circulating tumor HPV DNA (ctHPVDNA).

37. The system of any one of claims 34-36, wherein the sample further comprises HPV serology of the subject.

38. The system of any one of claims 37, wherein the plurality of features comprises at least two of HPV reads as determined by unique molecular indices, HPV genome coverage, HPV genotype, HPV lineage and sublineage, normalized HPV genomes / human genomes, viral integration events, select human genes and gene mutations, ctHPVDNA fragment size features, prognostic viral single nucleotide polymorphisms, or prognostic human germline single nucleotide polymorphisms.

39. The system of any one of claims 34-38, wherein the plurality of features further comprises HPV antibodies.

40. The system of any one of claims 34-39, wherein the plurality of features further comprises patient demographics, comprising at least one of sex, age, race, HIV status, or smoking status.MEEI 2024-326-02Quarles 169511.0005841. The system of claim 38, wherein the HPV genotype comprises identifying the HPV genotype as one of more than 13 HPV genotypes.

42. The system of claim 38, wherein the viral integration events comprise identifying viral genome breakpoints and human genome breakpoints.

43. The system of claim 38, wherein the select gene mutations comprise PIK3CA mutations.

44. The system of any one of claims 34-43, wherein the analytics pipeline comprises at least one trained network.

45. The system of claim 44, wherein the at least one trained network comprises a multi-label, multi -class network.

46. The system of claim 44 or claim 45, wherein an architecture of the trained network comprises at least one of a random forest, extra tress, decision trees, gradient boosting, XGBoost, LightGBM, linear, logistic regression, ridge, lasso, elastic net, kernel-based, support vector machine, instance-based, k-nearest neighbor, probabilistic, or naive bayes.

47. The system of any one of claims 44-46, wherein the at least one network comprises a first network trained on sequencing data, and a second network trained on sequencing data and HPV antibody data.

48. The system of any one of claims 44-47, wherein the trained network was trained on labeled subject data comprising control samples collected from subjects who have not been diagnosed with a cancer and samples collected from subjects who have been diagnosed with a cancer.MEEI 2024-326-02Quarles 169511.0005849. The system of any one of claims 34-48, wherein the cancer score is used for at least one of screening, diagnosis, detecting molecular residual disease, prognosis, or surveillance.

50. The system of any one of claims 34-49, wherein the cancer score is used longitudinally for the subject for more than one of screening, diagnosis, detecting molecular residual disease, or surveillance.

51. The system of claim 49 or 50, wherein the subject does not display any symptoms indicative of cancer at a time of screening.

52. The system of claim 49 or claim 50, wherein the subject displays one or more symptoms indicative of cancer at a time of screening.

53. The system of any one of claims 49-52, wherein the sensitivity of the screening is at least 25-fold more sensitive than dPCR.

54. The system of any one of claims 49-53, wherein the sensitivity of the screening is at least 80-fold more sensitive than dPCR.

55. The system of any one of claims 34-54, wherein the system is used to diagnose a subject.

56. The system of claim 55, wherein the sensitivity of the diagnosis is greater than 98% at a specificity greater than 98%.

57. The system of claim 55 or claim 56, wherein the method is able to detect a disease more than three years and a half years before disease expression.

58. The system of any one of claims 55-57, wherein the method is able to detect a disease more than twenty months before disease expression.MEEI 2024-326-02Quarles 169511.0005859. The system of any one of claims 34-58, wherein the method is used to detect molecular residual disease of the subject.

60. The system of claim 59, wherein the method enables detection of molecular residual disease more than 150 days before disease expression.

61. The method of claim 59 or claim 60, wherein the detection of the molecular residual disease has a sensitivity greater than 30%.

62. The method of any one of claims 59-61, wherein the detection of the molecular residual disease has a sensitivity greater than 50%.

63. The system of any one of claims 59-62, wherein the detection of the molecular residual disease after surgery is 50% more sensitive than dPCR.

64. The system of any of claims 59-63, wherein the detection of the molecular residual disease after treatment completion is 25% more sensitive than dPCR.

65. The method of any one of claims 47-64, wherein a clinical decision is made based on the results of the screening, detection of MRD, or surveillance.

66. The system of any one of claims 34-65, wherein the sample is a blood sample.

Images