loading . . . Leveraging sequences missing from the human genome to diagnose cancer communicationsmedicine Article A Nature Portfolio journal
https://doi.org/10.1038/s43856-025-01067-3
Check for updates
Ilias Georgakopoulos-Soares 1,2,3,23 , Ofer Yizhar-Barnea1,2,23, Ioannis Mouratidis3,4,23, Candace S. Y. Chan 1,2, Michail Patsakis3, Akshatha Nayak3, Rachael Bradley1,2, Mayank Mahajan5, Jasmine Sims1,2, Dianne Laboy Cintron1,2, Ryder Easterlin1,2, Julia S. Kim1,2, Emmalyn Chen6, Geovanni Pineda7, Guillermo E. Parada 8, John S. Witte 6,9, Christopher A. Maher 10,11,12, Felix Feng 13,14,15,16, Ioannis Vathiotis 17,18, Nikolaos Syrigos17,19, Emmanouil Panagiotou17, Andriani Charpidou17, Konstantinos Syrigos17, Jocelyn Chapman 14,16, Mark Kvale 2,20, Martin Hemberg 5,21 & Nadav Ahituv 1,2,22
Abstract
Background Cancer diagnosis using cell-free DNA (cfDNA) has the potential to improve treatment and survival but has several technical limitations. Methods In this study, we developed a prediction model based on neomers, DNA sequences 13–17 nucleotides in length that are predominantly absent from the genomes of healthy individuals and are created by tumor-associated mutations. Results We show that neomer-based classifiers can accurately detect cancer, including early stages, and distinguish subtypes and features. Analysis of 2577 cancer genomes from 21 cancer types shows that neomers can distinguish tumor types with higher accuracy than state-of-the-art methods. Generation and analysis of 465 cfDNAwhole-genome sequences demonstrates that neomers can precisely detect lung and ovarian cancer, including early stages, with an area under the curve ranging from 0.89 to 0.94. By testing various promoters or over 9000 candidate enhancer sequences with massively parallel reporter assays, we show that neomers can identify cancer-associated mutations that alter regulatory activity. Conclusions Combined, our results identify a sensitive, specific, and simple cancer diagnostic tool that can also identify cancer-associated mutations in gene regulatory elements.
Cancer is the second leading cause of death worldwide1,2. For most cancer types, survival is significantly higher if the tumor is detected at an early stage3,4. Currently, mass population screening is applicable only for breast and cervical cancers and utilizes physical tests like mammography and cytology screens. Detection for other cancer types, done both en masse and in a low and affordable-resource setting, still poses a major challenge for the scientific and clinical communities1. In particular, a major hurdle is to identify reliable biomarkers for the detection of cancer at a presymptomatic stage. Detection at such an early stage would allow not only for improved survival but would also decrease treatment toxicity and provide an opportunity for providing personalized treatments.
Circulating cell-free DNA (cfDNA) has a lot of potential to become an effective biomarker for cancer diagnostics5–8. It has a short life span
(16min to 2.5 h), making it an efficient temporal indicator of biological processes in the body. With improvements in sequencing technologies and their costs becoming lower, cfDNA can also be a rapid and cheap biomarker. Being able to identify circulating tumor DNA (ctDNA) is also minimally invasive and can allow not only for cancer detection but also to monitor patients already diagnosed with cancer. To identify cancerous cells, their tissue of origin, cancer type, minimal residual disease, and other cancer features, current cfDNA sequencing technologies rely on the detection of somatic mutations9 and epigenetic marks, such as DNA methylation or histone modifications that can determine the cancerous tissue of origin10,11. For example, one study showed that cell-free DNA methylation patterns can be used for the accurate detection of early-stage breast cancer12. Furthermore, cfDNA has been used as a biomarker for
A full list of affiliations appears at the end of the paper. e-mail: [email protected]; [email protected]; [email protected]
Plain language summary
DNA fragments circulating in the blood are often present in people with cancer. Current tests using DNA fragments circulating in the bloodhave limitations indetectingcancer.We developed a cancer diagnostic approach that uses neomers, short DNA sequences that are mostly absent in healthy individuals but are present in tumors as a consequence of changes in the DNA (DNA mutations) in cancer cells.We show that this approach can accurately detect lung and ovarian cancers from DNA fragments extracted from blood and identify changes in the DNA associated with cancer. Our diagnostic tool could be used to improve cancer detection and identify how changes in the DNA affect development of cancer.
Communications Medicine | (2025) 5:363 1
12 34
56 78
90 (): ,;
12 34
56 78
90 (): ,;
treatment response and recurrence13. However, ctDNA still has many hurdles and caveats that need to be overcome14. Some of the major hurdles include: (1) Circulating cfDNA is found in fragment sizes ran- ging between 120 and 220 base pairs (bp), with most of the fragments being 167 bp15 in length, thus making it challenging to extract; (2) Tumor-derived DNA makes up around 0.4% of the total cfDNA (there is some variance between tumor types), thus needing extremely sensitive biomarkers that can detect the presence of cancerous cells; (3) Prior knowledge of specific mutations or methylation marks is required for targeted screening, leading to biased screening primarily of coding mutations which make up only a small fraction of cancer-associated mutations. (4) Somatic alterations in white blood cells can confound cfDNA mutations and epigenetic changes16; (5) The techniques used to detect methylation or histone marks from cfDNA are technologically complex and have low sensitivity and specificity, in particular in early stages5,17,18. To provide the most optimal cancer treatment, it needs to be diagnosed at early stages when the tumor is small (~5mm in diameter). At these stages, the tumor produces very low levels of ctDNA that are difficult to detect using current methods5.
Nullomers are shortDNA sequences (13–17 base pairs) that are absent from the human genome19–21. While the absence of nullomers could be due to chance,we and others have shown that a significant proportion of them is under negative selection pressure20,21, suggesting that they may have a deleterious effect on the genome. We have also shown that these sequences could be used as DNA “fingerprints” to identify specific human populations21. As nullomers generally donot exist in a human genome, their appearance due to mutagenesis followed by clonal expansion could be exploited as a diagnostic method for diseases associated with a mutational burden, such as cancer. Nullomers have been used for the detection of five different cancers using cell-free RNA22.
Here, we set out to test whether nullomers could be used as a diagnostic tool to detect cancer and various additional tumor features. Throughout this manuscript, we refer to nullomers uniquely found in multiple tumor genomes as neomers to distinguish them from the more general category. We first analyzed The Cancer Genome Atlas (TCGA23) database, finding that neomers created by somatic mutations are able to detect cancer subtypes with higher accuracy than leading methods24 as well as additional cancer features. Further analyses of cfDNA whole- genome sequencing (WGS) datasets found that these neomers can also be used to detect cancer subtypes. Using WGS on cfDNA from 465 indi- viduals with ovarian or lung cancer and normal controls found an enrichment for neomers. Our neomer-based cancer detection models had area under the curve (AUC) of 0.93 for lung and 0.88 for ovarian cancers, and an AUC of 0.94 for lung and 0.93 for ovarian cancer at early stages. This is particularly noteworthy for ovarian cancer as currently there is no effective screening test25, with most women being diagnosed at later stages, IIIC or IV, where 5-year survival rates are 39% and 17%, respectively26. Finally, utilizing both promoter assays and a massively parallel reporter assay (MPRA), we show that neomers alter regulatory activity and can be used to detect cancer-associated mutations in gene regulatory elements. Combined, our results show that neomers can be used as a rapid, sensitive, specific, and simple cancer diagnostic tool and also aid in the identification of gene regulatory mutations associated with cancer.
Materials and methods Computational characterization of nullomers The GRCh38 reference assembly of the human genome was used throughout the study. Nullomer extraction was performed for kmer lengths up to17basepairs using the algorithmdescribed inourpreviouswork21. The reverse complement of a nullomer will also be a nullomer and throughout this manuscript when counting nullomers, the reverse complement of nullomer i was also considered separately, unless i is a palindrome. Sub- stitutions and indels that were not present in healthy tissues were identified fromWGS of tumor samples from 2577 individuals across 21 tissues were
obtained from open access release data at https://docs.icgc-argo.org/docs/ data-access/icgc-25k-data23. Recurrent nullomers (neomers) (ri) were annotated as those that resulted from substitutions or indels across two or more patients within a cancer type. When possible, ri was chosen to get ~10,000 neomers from each tissue; otherwise, it was set to 2 (Data S1). Drivermutation-derived neomers were defined as nullomers detected from driver mutations and were identified using the catalog of driver mutations derived from the intOGen database27, for which we grouped driver muta- tions by tissue of origin.
Classification of tumor tissue of origin using neomers We trained a classifier to distinguish classified tissue of origin for a cancer sample based on observed neomers using the libSVM package for Julia28
with default parameters to train a support vector machine (SVM) classifier with a linear kernel.We used 10-fold cross-validation, whereby the classifier was evaluated on a held-out fraction of the data. The set of neomers for each cancer type was recalculated for each round to only include the patients in the training set.
Supervised selection of nullomers TheMMRstatus of each biopsy samplewas derived from ref. 29. Themodel was trained on neomers identified in MSI samples and the performance of the algorithm evaluated. Previous research has indicated that mismatch repair deficiency is associated with mutations at mononucleotide repeat tracts30,31. For the MSI versus the microsatellite stable (MSS) samples, we counted the number of neomers that contained either AAAAAAAA or TTTTTTTT repeats, since MSI cancers have been associated with muta- tions of polyA/T repeats32. The threshold for determining MSI or MSS was set as the harmonicmean of themaximumnumber of counts in theMSS set and the minimum number of counts in the MSI set. The POLE deficiency status of each biopsy sample was derived from ref. 29 and we used a similar strategy to that of MMR status, but instead counted neomers created through either a TCT > TATorTCG> TTGmutation. Since the number of patients in each category was limited, we used a 5-fold cross validation. For ovarian cancer analyses, two sets of nullomers were used. From the nullo- mers of length k = 13 identified in ref. 21, we retained only the ones that could not be created by any of the single base-pair substitutions identified in gnomAD v233. Similarly, we selected all nullomers of length k = 15 of order 1, i.e., nullomers that can only be created from the human reference genome through at least two insertions, deletions, or substitutions.
cfDNA extraction and WGS Lung and ovarian cancer samples, as well as age and gender-matched samples were purchased fromProteoGenex Inc. (Inglewood, CA,USA) and IndivuMedGmbH (Hamburg, Germany). In addition, cancer samples, and when available, healthy age and gender-matched samples, were also obtained from the labs of K. Syrigos (Sotiria Hospital, National and Kapodistrian University of Athens, School of Medicine, Athens, Greece), J. Chapman (UCSF, San Francisco, USA), J. Witte (UCSF, San Francisco, USA), and C. Maher (WashU, St. Louis, USA). The samples that included ancestry information were all Caucasian (Data S3). All samples were col- lected with informed consent. All human sample work was carried out under the approved UCSF human research protection program institu- tional review board protocol number 21-35920. cfDNAwas extracted from 1mL of plasma, following centrifugation at 4 °C at 34 × g for 3min, to remove larger debris, using the QIAamp Circulating Nucleic Acid Kit (Qiagen). cfDNA was eluted in 50 μL of elution buffer and measured using the Qubit High-Sensitivity dsDNA kit, and validated for size distribution (160–180 bp) using an Agilent BioAnalyzer 2100 Sensitivity DNA chip. Up to 10 ng of cfDNA was used for sequencing library construction using the library preparation enzymatic fragmentation kit 2.0 (Twist Bioscience) adjusted with IDT’s xGen UDI-UMI 96 barcodes system (IDT) to replace theTwist universal adapter, andusingKAPAHiFi polymerase insteadof the polymerase provided by the kit. Libraries for lung, ovarian, and their respective controls were sequenced as PE150 using a NovaSeq 6000
https://doi.org/10.1038/s43856-025-01067-3 Article
Communications Medicine | (2025) 5:363 2
S4 system (Illumina) leveragingUMIs to removePCRduplicates, aiming for 5–10X coverage per multiplexed library, via Novogene. Samples were sequenced in batches of 96 (maximalUDI by the xGenkit) andwere equally spread across multiple lanes. Each batch contains a consistent mixture of healthy controls and cancer samples.
Neomer identification in cfDNA samples To improve the detection of lung cancer mutations in cfDNA, we used neomers that are present in ≥2 patients rather than ≥3 as was used for the tissue-of-origin classifier. The FASTQ files were scanned for neomers by searching for exact matches to the 16-mers of interest. We disregarded any neomer that could arise due to a common germline variant (VAF > 1%). To reduce errors, only reads where the nullomer was found in both pairs were kept. To evaluate the ability to detect cancer from cfDNA,we first randomly split samples into training and validation sets. Prior to the split, samples were randomly removed from either the controls or patients to ensure that the two groups were of equal size. From the training set we determined a threshold, t =max (number of neomers in control samples from the training set)— 1. Samples in the validation set were thendesignated as having cancer if the number of neomers detected was ≥t. To calculate F1 scores, this procedure was repeated 100 times. For the downsampling analysis, each read pair in sample i was retained with a probability of c/Ci, where c is the target coverage and Ci is the full coverage of the sample, i.e., sampling without replacement. For each value of c the downsampling was repeated 100 times.
Classification model to detect cancer patients from cfDNA Neomers were identified for each donor and combined as a neomer den- sity × samplematrix,dij = nij/Cj, wherenij is the countofneomer i in sample j and Cj is the sample coverage. Neomer-based classification models were generated to examine the ability of neomers to detect cancer for lung, colorectal, ovarian, pancreatic, and prostate cancers. We tested PCA and feature hashing to dealwith the sparse nature of our dataset andused feature hashing with 8192 features as the best performing method. We evaluated five classification models; random forest, linear, and radial basis function SVMsand linear regression (LR)modelswith L1 andL2 regularization, each evaluated with a 10 × 10-nested cross-validation. Our benchmarks indi- cated LR with an L2 penalty as the best performing model across all cancer types. The results obtained were also examined regarding tumor staging for colorectal and lung cancers, in which we had sufficient sample sizes. The nested cross validation (CV) used a 10-fold inner CV to optimize model parameters and hyperparameters and a 10-fold outer CV to test the opti- mized classifier on data not used for parameter and hyperparameter opti- mization. This design yields an unbiased estimation of performance while allowing for full optimization of the model. Model parameters were opti- mized with respect to the balanced accuracy metric, as were hyperpara- meters across the innerCV.Agrid searchwas used to determine the optimal hyperparameters. Model performance in the outer CV was evaluated using balanced accuracy, AUC, and other binary classification metrics. Choice of model type and of number of hashed features were made outside of the nested CV framework. In practice, the classifier performance tended to plateau or achieve a mild maximum with respect to the number of hashed features, so the number of hashed features chosen was sufficiently large to place the classifier in the performance plateau. No one model type con- sistently dominated the others in terms of performance, so the model type used was determined by the needs of the analysis and the different tradeoffs with respect to complexity and interpretability.
Fragmentomics-based classification models were generated to com- pare its performance against the neomer method on the lung and ovarian cancer datasets generated from this study. Paired-end sequencing reads were aligned to the GRCh38 reference genome using Burrows-Wheeler Aligner - Maximal Exact Matches (BWA-MEM).
Genome-wide fragmentation profiles for each sample (defined as the ratio of short to long fragments) were created using https://github.com/
mhanbioinfo/fragmentomicsBEDPE, where GC-corrected counts were binned into 5Mb windows. The fragmentation profile was then used in a stochastic gradient boosting model with 10-fold cross validation as descri- bed in ref. 34. Alternatively, the read coverage on each of the 39 autosomal arms was calculated and compared to the average coverage to generate coverage z-scores for each sample as described in ref. 34 and used as features for a stochastic gradient boosting model with 10-fold cross validation. For endmotif analysis, duplicates fromaligned readswere removedusingPicard MarkDuplicates, and 167 basepair fragments were retained and converted to BEDPE format using samtools35. 256 tetranucleotide motifs at the breakpoints of 167 basepair fragments were used in a SVM and LR model using leave-one-out cross-validation as described in ref. 36.
Promoter luciferase assays Promoter sequences with and without the neomer (Data S4) were synthe- tically generated and cloned into the modified Promega promoter assay luciferase vector pGL4.11b (a gift from Dr. Rick Myers, HudsonAlpha) by BioMatik Inc. and Sanger sequence verified. LNCaP cells were plated at an initial density of 2 × 105 cells/well in 24-well tissue culture plates and maintained in RPMI medium, 10% FBS supplemented with L-Glutamine and Penicillin/Streptomycin. Plasmids together with a Renilla expressing plasmid, pGL4.74 (Promega), at a ratio of 10:1 luciferase:Renilla were transfected using the X-tremeGENE™ HP DNA Transfection Reagent (Roche) using 1:4 ratio ofDNA(μg) to reagent (μl). Seventy-twohours post- transfection, Luciferase and Renilla levels were measured using the Dual- Luciferase Reporter Assay System (Promega) following the manufacturer’s protocol using a GloMax Explorer Multimode Microplate Reader (Pro- mega). Luciferase activity was normalized to Renilla levels and presented as Relative Luciferase Units (RLU). Statistical analysis was performed using Prism version 9.0.2 (GraphPad). All values were reported as means (AVG) and standard errors (SE). p < 0.05 were considered statistically significant.
Massively parallel reporter assays The lentivirus-based MPRA was performed following a previously published protocol37. An oligonucleotide library of 230 bp long fragments bearing: (1) 4609 sequences with and without a recurring neomer that was found in prostate cancer patients; (2) 100 scrambled control sequences randomly selected from the library. A pool of 230-nt oligos containing each of these 200-nt sequences flanked by 15-base pair primer sequences on either side was synthesized (Agilent Technologies), amplified, and cloned into the lentiMPRA plasmid. Briefly, amplified inserts were cloned upstream of a minimally active promoter (mP), 5’UTR barcode (BC), and EGFP reporter transcript. We aimed for ~100 BCs per sequence when constructing and extracting the plasmid library. Each insert was associated with a set of BCs using paired-end (PE150) sequencing on an Illumina NextSeq 500(see Gordon et al. Step 8337). Lentivirus was produced and titered and used to infect LNCaP cells at an MOI of 50 virus particles per cell. Three days post-infection, to remove non-integrating virus, DNA and RNA were extracted from the three replicates and used for library construction and multiplexing for sequencing. All MPRA-related sequencing was performed using an Illumina NextSeq 500 with either PE150 for the CRS-BC association library or PE15 for the DNA/RNA BC count portion of the protocol. MPRA analysis was carried out using MPRAflow37.
Comparison to validated neoantigens We downloaded a list of 1967 validated neoantigens from http://biopharm. zju.edu.cn/download.neoantigen/iedb_validated.zip. Requiring both pre- dicted strong binding and a positive validation, provided 1700 neoantigens. To evaluate the enrichment of neoantigens corresponding to neomers, we assumed a hypergeometric distribution with 1700 draws from an urn with 188,659 white balls (total number of neomers) and 186,067,892 black balls (number of nullomers foundwith lower recurrency thanwhat was required for neomers).
https://doi.org/10.1038/s43856-025-01067-3 Article
Communications Medicine | (2025) 5:363 3
Indel identification Somatic indel calls were performed using three pipelines from four somatic variant callers. These included those from the Wellcome Sanger Institute, theDKFZ/EMBL, and the Broad Institute,which reported a somatic variant false discovery rate of 2.5%23,30.
Statistics and reproducibility Statistical tests are described in the context where they are used. All code is publicly available as described above for reproducibility of computational analyses.
Results Annotation of mutations that lead to nullomers As cancer is associated with a large number of somatic DNAmutations, we investigated if a subset of them could result in the resurfacing of nullomers (Fig. 1a). Using our previously characterized human nullomers21, we ana- lyzedWGS results from 2577 patients across 21 cancer types from TCGA38
for resurfacing nullomers (Fig. S1A).We focused on 16 bpnullomers, as it is the shortest length where we detect a sufficient number of nullomers per patient, with the human reference genome having only 37.24% of all pos- sible 16mers. Themajority of the 44,599,472 single-nucleotide substitutions that were tumor-unique give rise to multiple nullomers, allowing us to identify 213,164,038 resurfacing nullomers across all cancer types. Fur- thermore, we identified 2,470,091 nullomers resulting from short insertions and deletions (1–100 base pairs). Themedian number of nullomers created by each substitution was two, and for indels, four (Fig. S1B, C). On average, 58.29% of substitutions in a patient resulted in one ormore nullomers, with 2.1% of the nullomers residing in coding regions. The median number of nullomers foundacross cancer patientswas 9107 (Fig. 1b, c) and thenumber of nullomers was directly proportional to the number of mutations (Fig. S1D). As mutations were identified by comparisons to healthy tissues, germline variants of each individual were not included38.
To further prioritize nullomers that could be used as cancer bio- markers, we focused on the subset of nullomers that are recurrent, i.e., those found inmore thanonepatient for a specific cancer type, termedhereafter as neomers. The number of neomers was proportional to the total number of mutations (Fig. 1c and Data S1). As both the number of patients per cancer type and the mutational load varied, the median number of neomers for each tissue type ranged from 0 to 98. Analysis of themost frequent neomers revealed several previously known cancer-associated driver mutations (Table 1). For example, someof themost recurrent codingneomerswere the result of either the Gly12Asp, Gly12Val, or Gly12Cys missense mutation in the KRAS proto-oncogene GTPase (KRAS), which are known to make up 80% of cancer-associated KRAS mutations and lead to KRAS being con- stitutively active39,40. Although KRAS mutations have been associated with several cancer types, 190/215 (88%) of these mutations were found in pancreatic cancers. Several frequently occurring coding neomers were also found in other known cancer-associated genes, such as tumor protein p53 (TP53), B-Raf proto-oncogene serine/threonine kinase (BRAF), and phos- phatidylinositol-4,5-bisphosphate 3-kinase catalytic subunit alpha (PIK3CA). The most frequent neomer was located in a noncoding region, within the telomerase reverse transcriptase (TERT) promoter, which is known to be associated with numerous cancer types41 and poor prognosis42. This mutation, called −124C > T or C228T, is extremely common in numerous cancer types43 and is thought to disrupt aG-quadruplex44 leading to the binding ofGABP45, an ETS transcription factor, resulting in increased TERT expression. We found this mutation in 97 patients, with the highest incidence in glioblastoma (51%), fitting with its known high prevalence rate and diagnostic use for this cancer type46.
We also identified several neomers that are frequently created by dif- ferentmutations (Table 2). Interestingly, some of these frequently recurrent neomers are created by differentmutations, yet are predominantly found in one cancer. For example, GTTTTTCTCCTAGACC is found 41 times in skin cancer at 32 different loci whileCTGGCAGTGAGCCACG is found 26 times in liver cancer across 23 loci. The majority (98%) of these frequent
neomers reside in noncoding regions, and many of them reside in intronic regions (35%). For example, CGACGTTCTGCCCACT is found in 32 loci, primarily in pancreatic and stomach cancer. Of those loci, 21/32 (65.6%) were found in noncoding regions nearby pancreatic cancer-associated genes. These include, for example, the C-C motif chemokine ligand (CCL4)47, the POM121 transmembrane nucleoporin like 12 (POM121L12), which is commonlymutated in gastrointestinal cancers48 and the potassium voltage-gated channel modifier subfamily V member 1 (KCNV1) gene, where promoter hypermethylation has been associated with both pancreatic49 and esophageal cancer50.
To further validate that nullomers can detect cancer-associated mutations, we examined whether they are enriched in known cancer dri- vermutations.We annotated a cohort of drivermutations from 28 different cancer types27 for their overlap with 16 bp nullomers. In the pan-cancer analysis, we identified 19,594,212 nullomers resulting from 50,167 putative driver mutations (Fig. 1f and Data S2). For specific cancer types, we found that on average, 81% of driver mutations resulted in one or more nullomer, ranging between 63.88% and 86.50% in pancreatic and lung cancer, respectively (Data S2). The number of nullomers also varied by tissue type, ranging between 92 in thyroid and 9434 in skin (Fig. 1d). In general, driver mutations were 1.4-fold more likely to result in the creation of a nullomer (permutation test, p-value < 0.001). Taken together, our results suggest that a large subset of clinically relevant cancer mutations are associated with neomers, i.e., recurrent nullomers.
Cancer subtype neomer classifier We next set out to assess whether neomers can be used to distinguish between cancer types. We filtered neomers by keeping only those that appeared >=ri times in specific cancer type i (Data S1). Comparison of the set of neomers associated with each cancer type revealed a small overlap, as indicated by the Jaccard index which is <0.04, suggesting that each cancer type has a distinct neomer signature (Fig. 1e). We also counted the number of times neomers are found in eachpatient,finding that patients are strongly enriched for only one set of cancer-specific neomers (Fig. 1f).
We then tested if our annotated cancer-specific neomers can be used to classify tumor samples. We trained a SVM classifier to identify tumor type. The classifier takes as input a 21-dimensional vector indicating the number of neomers found for each cancer-specific set. For the classification model, we divide the samples into train and validation and test groups, and only filter the neomers and construct a classifier, based on the train samples. Evaluation using 10-fold cross-validation revealed that our classifier achieves both high sensitivity and specificity, with an F1 score of 0.92 and an accuracy of 0.99 (Fig. 2a, b). Performance was better than a recent deep learning model24 and also requires less computational resources to train as once neomers have been extracted, training and testing take only a few minutes.
Neomers distinguish additional cancer features We next tested whether a supervised feature selection approach, i.e., using neomers that are thought to be informative based on prior biological knowledge, would improve performance. For this approach, we initially utilized microsatellite unstable (MSI) and MSS cancers. MSI is associated with better cancer prognosis, increased benefits from surgery and higher sensitivity to immunotherapy, but with a lack of efficacy from adjuvant treatment51. Since MSI cancers are associated with mutations at polyA and polyT stretches52, we hypothesized that neomers containing these motifs would be able to effectively distinguish these two cancer types.We identified ten MSI samples from a cohort of 560 breast cancers32,53,54 and compared them to ten randomly selected MSS samples from the same cohort. We found that the polyA/T neomers were able to separate the two categories with an accuracy of 100% (Fig. 2c).
We next applied a similar strategy to distinguish patients with DNA polymerase epsilon catalytic subunit (POLE) deficiency, as these tumors are known to respond more favorably to immune checkpoint inhibitors54. We identified 25patients from theTCGAdataset labeled as POLEdeficient, and
https://doi.org/10.1038/s43856-025-01067-3 Article
Communications Medicine | (2025) 5:363 4
searched for neomers created through a TCT > TAT or TCG >TTG mutation, which are the most common types of mutations in this context55,56. Comparing against POLE proficient tumors, we found that the number of neomers identified for each group have very little overlap (Fig. 2d), and the classifier achieved an accuracy of 96%.
Neomers detect early-stage lung cancer We next tested whether neomers could be used to diagnose cancer in cfDNA.We first used lung cancer as our test case, due to having an ample amount of cancer-associated neomers (Data S1), difficulty to diagnose using current techniques, with early stages being mostly
https://doi.org/10.1038/s43856-025-01067-3 Article
Communications Medicine | (2025) 5:363 5
asymptomatic55,57, and being the leading cause of cancer-associated mortality worldwide55. We assembled a cohort of cfDNA from the plasma of 381 individuals. These included 200 lung cancer patients from all stages with a strong emphasis on early stages: 77 stage I (38% of cancer cases), 36 stage II (18%), 34 stage III (17%), and 53 stage IV (27%). The majority of our cases were adenocarcinoma (50.2% of cancer samples), followed by squamous cell carcinoma (37.8%) (Data S3). The median age was 64 ± 7.8 years old, with patients as young as 42 years old.We also had 180 controls, considered cancer free (e.g., “healthy”, no cancer of any kind detected >2 years after sampling) with a median age of 58 ± 8.1 years and 90% fitting the criteria for the United States Preventive Services Task Force (USPTF) age interval for early screening for lung cancer. Of the healthy individuals, 21% were active smokers (37.2% were noted as unknown), while 64.2% of the cancer patients were noted as active or past smokers (complying with USPTF guidelines for lung cancer screening) (additional information on the cohort is available in Data S3). Samples were paired-end sequenced at a genomic depth of 5-10× coverage (C) and analyzed for neomers. For each sample and each neomer, all occurrences of the neomer were extracted from the FASTQ file, and the resulting neomer count was normalized by dividing by the sample read depth, forming a neomer density measure. Putting all the neomer den- sities into a table with sample rows and neomer columns yields a neomer density matrix (NDM). The NDM with an added column for the sample status forms the input for our classification models. To reduce the impact of sequencing errors, we only considered neomers that were found on both read-pairs from a single pair-end read, resulting in 18,090–55,409 read-pairs per sample. To further avoid spurious neomers, only neomers that overlapped at least two read-pairs and overlapped the human
genome were used for subsequent analyses. Of note, an important computational advantage compared to conventional mutation calling pipelines is that only a single pass is made across the reads to identify those containing neomers, and only those reads are aligned to the reference genome. Using these data, we trained an LR classifier, achieving an AUC score of 0.93 using 10-fold cross validation and a sensitivity of 92.5% (specificity of 85%) (Fig. 3a, b). Importantly, our classifier also performed well for early stages, with an AUC of 0.94 for stage I and sensitivity of 96.1% (specificity of 85%) (Fig. 3c, d). To showcase the ability of the method to recognize different cancer types, we tested the performance of the classifier on lung cancer samples when utilizing the ovarian neomer list. The performance for all models is clearly lower than the performance of the optimal model trained on lung cancer neomers, indicating the ability of the classifier to discriminate between cancer types —a sample will be given a much higher chance of being from a cancer patient when using the corresponding cancer neomer list (Fig. S3). Finally, by carrying out a downsampling analysis, we find that already at 3x coverage we can obtain similar accuracy (Fig. S2). In summary, these data suggest that our neomer lung cancer profile can effectively detect lung cancer from cfDNA with high accuracy, including early stages.
Neomers detect early-stage ovarian cancer We next set out to test how our neomer classifier performs with a low amount of tumor-associatedneomers.We focused on ovarian cancerwhich only has 12,788 associated neomers (amongst our analyzed TCGA WGS datasets) (Data S1). In addition, as ovarian cancer does not have an effective screening test25, most women are diagnosed at stages IIIC and IV, when 5-year survival rates are 39% and 17%, respectively26, reflecting a pressing
Fig. 1 | Neomers can detect cancer tissue of origin. a Schematic overview of neomer cancer diagnostic pipeline. Created in Biorender. Ahituv, N. (2025) https://BioRender.com/9sgstbs. b–f 2577 patients across 21 cancer types were used. b Number of neomers per patient sample across tissues. Each dot represents a patient sample. The boxplot displays the median as the central line. The bottom and top edges of the box represent the first quartile (25th percentile) and third quartile (75th percentile), respectively. The lower and upper whiskers extend to the smallest and largest data values that fall within 1.5 times the interquartile range above the 25th percentile and below the 75th percentile. c Number of
neomers and the number of substitutions for 2577 patients (Spearman’s ρ = 0.75). d Identification of nullomers at driver mutation sites across cancer tissues. The majority of cancer driver mutations result in the resurfacing of nullomers. The proportion of mutations that cause the resurfacing of nullomers is indicated under each cancer type. e Heatmap showing the Jaccard index for the overlap of neomer sets associated with different cancer types. f Heatmap showing the occurrence of neomers across patients for each cancer type. Each row represents a cancer type and each column a patient. The intensity of the heatmap (log2-scale) shows the number of neomers for each tissue set.
Table 1 | Common cancer-associated neomers
Neomer Locus Coordinate (hg38) Mutation Name Patient # Cancer #
AGGGCCCGGAAGGGGC TERT chr5:1295113 G > A −124C > T C228T 97 7
TCTTGCCTACGCCATC KRAS chr12:25245350 C > T c.35G > AGly12Asp 89 6
CTGTTGGCGTAGGCAA KRAS chr12:25245350 C > A c.35G > TGly12Val 79 6
ACGCCACGAGCTCCAA KRAS chr12:25245351 C >G c.34G > T Gly12Cys 47 1
GGTGCATGTTTGTGCC TP53 chr17:7673802 C > T c.818G > A Arg273Pro 35 13
GTGGGGGCAGTGCCTC TP53 chr17:7675088 C > T c.524G > A Arg175His 28 11
Six of the most common neomers created by a single mutation. The nucleotide in bold font and underlined is the neomer causing mutation.
Table 2 | Frequent cancer-associated neomers
Neomer Loci # Patient # Cancer # Cancer type
CGACGTTCTGCCCACT 32 74 7 Mainly pancreatic and stomach
GTTTTTCTCCTAGACC 32 41 2 All but one in skin
CTGCAGTGGCGCAATA 30 30 11 33%of the neomers in colon cancer
TTATAGGGGTCCAGTG 25 26 1 (Five of the top frequently recurring neomers created by several different mutations)
Colorectal only
CTGGCAGTGAGCCACG 23 26 2 21 in liver, 5 esophagus
Five of the top frequently recurring neomers created by several different mutations.
https://doi.org/10.1038/s43856-025-01067-3 Article
Communications Medicine | (2025) 5:363 6
need for early detection methods. We obtained plasma samples and gen- erated cfDNA from 70 patients (Data S3) with a median age of 57 ± 16.1 years. Our analyzed cohort comprises 45.7% early stage samples (24 stage I and 8 stage II), and the rest from late stage (34 stage III and 4 stage IV). Samples were similarly paired-end sequenced at a genomic depth of 5–10× coverage (C) and analyzed for neomers. As controls, we used female donors from our control donor group, totaling 123 control cfDNA samples with a median age of 58 ± 7.9 years (additional information on this cohort is provided in Data S3). Using the same classifier as before, we obtained an AUC of 0.89 and a sensitivity of 62.86% with a specificity of 95.12% for all stages (Fig. 3e, f). Importantly, for early stages (stage I and II), we obtained an AUC of 0.93 and sensitivity of 71.9% with specificity of 95.12% (Fig. 3g, h). In addition, we carriedout a downsampling analysis andwe find that already at 3× coverage we can obtain high accuracy (Fig. S2). Com- bined, these results suggest that neomers can detect cancer even with a low tumormutation burden and,more importantly, could potentially be used to diagnose early-stage ovarian cancer for which there is currently no available detection method. We also compared our approach to fragmentomics- based algorithms. We found that after implementing a fragmentomics- based stochastic gradient boosting model on our dataset, using 10-fold cross-validation, we obtain AUCs of 0.83 for lung cancer and 0.87 for ovarian cancer (Fig. S6A, B)34. Additionally, we extracted 256 end motifs from all 167 basepair fragments across the genome and used a SVM and LR model to differentiate healthy from cancer samples36. This model had an AUCof 0.75 and0.81, respectively for lung cancer andAUCof 0.93 and0.94
for ovarian cancer (Fig. S6C, D). Finally, we also implemented a stochastic gradient boosting model for lung cancer detection using coverage Z-scores derived from 39 autosomal arms (AUC 0.87 for lung cancer, AUC 0.93 for ovarian cancer) (Fig. S6E, F)34. Our analyses show that the neomer-based method performs comparably or better than existing fragmentomic approaches for cancer detection using the same cohort and preprocessing pipeline.Notably, it achieved higher or similarAUCs acrossmodels for both lung and ovarian cancer, highlighting its robustness and potential clinical utility.
Neomers alter promoter activity Only a small number of mutations in gene regulatory elements have been found to be associated with cancer38,58,59. As our results show that driver mutations are enriched for neomers (Fig. 1d) and the majority of our neomers (98%) lie in noncoding regions, we hypothesized that neomer- creating mutations in noncoding regions could be used to detect cancer- associated mutations in gene regulatory elements that have a functional consequence. Of note, the top patient recurrent neomer was in the TERT promoter (Table 1), which is associatedwith numerous cancers59.We chose to focus on prostate cancer since it is associated with a relatively small number of neomers (N = 4621; median per patient = 29.5), allowing us to characterize all the detected neomers for this cancer via MPRA in the following section. Moreover, 80–90% of prostate cancer is dependent on androgen receptor (AR) signaling60, allowing us to use LNCaP-FGC cells, which are androgen-dependent, for our reporter assays. We initially set out
Fig. 2 | Neomers can distinguish cancer features. Boxplots of classifier accuracy (a) using an unsupervised classifier and F1 (b) score for the same classifier for each of the twenty-one cancer types. Box plot lines indicate lower, middle, and upper
interquartile values. c Separation of 10 MSI and 10 MSS samples using a supervised selection of neomers. d Separation of 25 POLE proficient and 25 deficient samples using nullomers. In (c, d) the vertical line displays the harmonic mean.
https://doi.org/10.1038/s43856-025-01067-3 Article
Communications Medicine | (2025) 5:363 7
to test neomers found in promoters for luciferase reporter assays (Fig. 4a) using the following criteria: (i) neomers that reside in a promoter based on ENCODE annotations61; (ii) the gene regulated by the promoter is asso- ciated with prostate cancer. Our list (Data S4) included neomers in: (1) a promoter between two divergent genes,RPS2 and the lncRNAgene SNHG9 (Fig. 4b), both of which are overexpressed in prostate cancer62; (2) a pro- moter between two divergent genes, TMEM127 and CIAO1 (Fig. 4c), with
the former being downregulated in prostate cancer63; (3) a promoter between twodivergent genes,TTC23 andLRRC28, with the former showing aberrant splicing that relates to therapy resistance in prostate cancer cells64; (4) The promoter of GNAI2, a protein that interacts with CXCR5, which positively correlates with prostate cancer progression65; (5) A promoter between two divergent genes, PRICKLE4 and FRS3, with the latter thought to affect malignant but not benign prostate cells66.
Fig. 3 | Neomers can detect lung and ovarian cancer from cfDNA. Receiver operator character- istics (ROC) curve (a) and confusion matrix (b) for the neomer based classifier for cfDNA from200 lung cancer patients and 180 controls. ROC curve (c) and confusion matrix (d) from 77 stage I lung cancer cases and 180 controls. ROC curve (e) and confusion matrix (f) for ovarian cancer from 70 cases and 123 controls. ROC curve (g) and confusion matrix (h) for 24 stage I and 8 stage II ovarian cancer and 123 controls.
https://doi.org/10.1038/s43856-025-01067-3 Article
Communications Medicine | (2025) 5:363 8
We cloned the promoter sequences with and without the neomer into a luciferase promoter assay vector and compared their activity in LNCaP cells (Fig. 4a). Two out of the five assayed promoters showed a significant effect on reporter activity between the neomer and wild type allele (Fig. 4d). For the RPS2-SNHG9 promoter, the neomer led to significantly increased activity, in line with this gene being
overexpressed in cancer62. For the TMEM127-CIAO1 promoter, the neomer completely abolished activity, fitting with its observed down- regulation in prostate cancer63. Combined, our experimental results show that neomers could have a significant effect on promoter activity and could potentially be used to identify cancer-associated gene reg- ulatory mutations.
https://doi.org/10.1038/s43856-025-01067-3 Article
Communications Medicine | (2025) 5:363 9
Neomers alter enhancer activity Wenext setout to test in ahigh-throughputmannerwhether neomers could affect the functionof active gene regulatory elements, suchas promoters and enhancers. We annotated all the prostate-associated neomers that reside in noncoding regions, selected for those that were recurrent in at least three patients (to increase likelihood to be cancer-associated and reduce sequences to be assayed due toMPRA limitations), obtaining 4609 loci and generated a lentivirus-based MPRA (lentiMPRA) to test the regulatory activity with and without the neomer of all these sequences (Fig. 4e). Two hundred base pair sequences, where the position of the neomer is used as a center, were synthesized and cloned upstream of a minimal promoter fol- lowed by a GFP reporter gene. Lentivirus was generated and the library was infected using three technical replicates into LNCaP cells and RNA and DNA barcodes were sequenced to determine regulatory activity, as pre- viously described37. We observed a good correlation between RNA/DNA ratios (Pearson 0.72-0.77; Fig. S4A) as well as for DNA and RNA barcodes separately (Fig. S4B, C) across all three technical replicates. Out of the 4609 loci, 567 showed significant regulatory activity due to the neomer (RNA/ DNA ratio ≥ 1.5). Amongst them, 94 neomers andwild type sequence pairs showed significant (p-value ≤ 0.05, U-Mann–Whitney test) differential enhancer activity, with 48 having increased activity due to the neomer (NEO/WT ≥ 2) and 46 decreased activity (NEO/WT ≤ 0.5) (Fig. 4F and Data S5).
The neomer that showed the highest level of increased activity com- pared to reference (4.09-fold) is located in an intron of the catenin alpha 1 (CTNNA1) gene. This gene is a core member of the cadherin/catenin complex and is involved in the regulation of theWnt/beta-catenin pathway, whichhas beenwidely studied in cancer67, includingprostate cancer68, and is the target of several therapies69,70. TFBS analysis of this neomer using FIMO71 and JASPAR72 found that it leads to a gain of a TCF7/TCFL1motif (MA0769.2) (Fig. 4G), which is known to play a role in prostate cancer malignancy73,74.Aneomer residing in the third intronof theHECTandRLD domain containing E3 ubiquitin protein ligase (HERC3) gene led to the lowest downregulation of reporter activity (2.7-fold; Fig. 4F). HERC3 is an AR target and its dysregulation has been associated with epigenetic mod- ulations and poor outcome in prostate cancer75. In addition, for colorectal cancer, this gene was found to be downregulated, is known to inhibit metastasis and is associated with overall lower survival74. Taken together, our MPRA results suggest that neomers could be utilized to identify gene regulatory mutations in cancer.
Discussion Cancer is a DNA mutation-associated disease. Here, we show that by analyzing cancerWGS, both from tumors and cfDNA, we can find cancer- associated DNA mutations that lead to the generation of neomers, short sequences that are predominantly absent from genomes of healthy indivi- duals. Further analyses of these sets of neomers show that they can be used not only to classify cancer tissue of origin, but also to identify additional cancer features, such as MSI or POLE deficiency, with high accuracy. Analysis of cfDNAWGS generated from lung and ovarian cancer patients finds that neomers could be used to detect cancer, including in early stages, with high specificity and sensitivity. This is particularly important for ovarian cancer, for which there is currently no effective screening test25 with most women being diagnosed at later stages when survival rates are low26. We were able to detect early-stage (stage I and II) cancer with an AUC of
0.93. Finally, using promoter assays andMPRA, we show that neomers can be used to detect cancer-associated mutations that have a functional effect on regulatory element activity.
We utilized 2577 patients from 21 different cancerous tissues to develop a cancer tissue of origin classifier. Overall, we observed almost no overlap between neomers found in different cancer types. This allowed us to detect cancer tissue of origin with extremely high specificity and accuracy (F1 score of 0.92 and an accuracy of 0.99), performing better than recent deep learning models24. In general, the classifier has better performance for cancer types withmore patients and highmutation burden (Fig. 2a, b). Our analyses also showed that other than tissue origin, nullomers can also be used to detect additional cancer features. Future work that will utilize additional cases andmetadatawould allowus to test whether nullomers and neomers can diagnose additional tumor features and also detect other cancer characteristics such as chance of recurrence, predict and/or monitor response to treatment, disease-free survival, and overall survival.
For cfDNA, despite various known challenges, including fragmenta- tion and low levels of ctDNA, we obtained highAUC scores, and our results for lung cancer compare favorably with a recent study based on metabolite profiling76. Lung cancer, being one of five cancer types with a widespread approved screening option, utilizes a low-dosage CT scan (LDCT) for populations at risk (age, smoking, etc.). LDCT’s sensitivity was 93%, spe- cificity was 84% when taking all stages together77–80. Recently, the employ- ment of a sequencing-based fragmentomics assay followed by LDCT provided a stage-weightedoverall sensitivityof 84%(95%CI: 78–90%)81. For ovarian cancer, which is associated with a low tumor mutation burden and has hitherto been extremely challenging to detect at early stages, we man- aged to detect, even with low sample numbers (N = 70), both late andmore importantly early stages (stages I and II), obtaining an AUC of 0.93 and sensitivity of 71.9% with specificity of 94.44%. Our higher AUC scores in early stages could be due to larger sample sizes at this stage and are in line with other publications that find higher scores in these early stages com- pared to late stages34,82,83.
It is worth noting that our classifier did not benefit from removing neomers that can arise due to common germline variants. Nevertheless, we expect that knowing the donor’s ancestry could allow us to further improve the classifier. However, this would requiremore detailedmetadata thanwas available in the present study. Since the number of neomers increases with the number of tumor samples profiled, obtaining additional WGS datasets from tumors and matched controls would also be beneficial in improving the ability of our classifier to diagnose both these and other cancers. Additional testing across multiple cancer types and in a larger number of samples is required to determine if the specificity that we obtained when using neomers to classify ICGC primary tumor samples can bemaintained. Weassume that, as cfDNAonly contains a small fractionof tumorDNA, the signal will be much weaker.
Our cfDNA detection approach has several advantages over current methods: (1) Detecting short DNA sequences that are enriched in cancer samples provides an easy-to-use diagnostic that could allow detection from low amounts of ctDNA. In addition to sequencing-based assays, alternate techniques could potentially be used, such as CRISPR-based diagnostic tools84, that can also allow the testing of thousands of sequences in parallel85. In addition, with neomer-based diagnostics potentially not needing large amounts of startingmaterial, cfDNAcould be collected fromurine, sputum, saliva, or other bodily fluids, which were shown to be a viable but reduced
Fig. 4 | Neomers alter the activity of gene regulatory elements. a Schematic describing the luciferase reporter assay. Created in Biorender. Ahituv, N. (2025) https://biorender.com/08yc3gj. Integrative Genomics Viewer track snapshot showing the location of a prostate cancer-associated neomer in the promoters of RPS2-SNHG9 (b) and TMEM127-CIAO1 (c). The top track shows the gene location (GENCODE V36), beneath it the ENCODE layered H3K27ac marks in this region, below that the promoter location (red line) according to ENCODE cCRE annota- tions and the most lower track shows the neomer sequence. d Relative luciferase units from a luciferase promoter assay (n = 4 per condition) containing either the
wild-type (WT) or nullomer variant. Transfection efficiency was normalized using Renilla luciferase and significance is calculated using a two-way ANOVA with multiple testing and Šidák correction. e Schematic showing the lentiMPRA work- flow testing 4906 wild type (WT) and neomer (NEO) sequences. f The regulatory impact of 4609 neomer creatingmutations, presented as log2 fold change of neomers containing sequence over wild type sequence (x-axis) and ordered from highest to lowest (y-axis). g A neomer in the CTNNA1 locus had the highest activating score and is postulated to lead to a gain of a TCF7 motif, as predicted by FIMO71.
https://doi.org/10.1038/s43856-025-01067-3 Article
Communications Medicine | (2025) 5:363 10
source of cfDNA86,87. (2) OurWGS approach is applicable to more sparsely sequenced samples. As also shown in our downsampling analysis, we obtain similar accuracy for both lung and ovarian cancer already at 3× coverage (Fig. S2). It is worth noting that this study was performed using standard Illumina sequencing. The use of sequencing technologies that provide longer reads88 and/or improved accuracy89 could further improve the ability of neomers to detect cancer. (3) As our diagnostic is based on neomer detection followed by genome alignments of nullomer encompassing reads, it is extremely effective from a computational standpoint. 4)Neomers could easily be combinedwithother sequence or analyte-based cancer biomarkers and risk factors to improve the diagnostic positive predictive value. For example, it was shown that combining a blood test that detects both protein biomarkers and DNA mutations, along with positron emission tomography-computed tomography (PET-CT), could detect multiple cancers90. In another example, the use of cfDNA fragmentation patterns combined with CT imaging, clinical risk factors, and serum levels of car- cinoembryonic antigen significantly increased the ability to diagnose lung cancer34. Adding neomers to other known cancer-associated mutations (coding, regulatory, CNVs, etc.) in the screening of cfDNA could also increase sensitivity and specificity. This should be relatively straightforward, as our neomer detection method is based on standard sequencing tech- nologies. In summary, coupling neomer-based diagnostics to existing can- cer biomarkers and risk factors could improve the power to detect various cancer subtypes.
As nullomers/neomers do not exist in the human genome they could alsobe exceptional candidates as suicide genes, as shownpreviously for three nullomers91, or for neoantigens, to be targeted via immunotherapy. Previous work has shown thatminimal absentwords, short sequences that are absent from a genome or proteome, could be used to identify phosphorylation sites of high confidence, some of which could be associated with cancer92. Ana- lysis of the Immune Epitope Database of validated antigens93 found that 13 of the recurrent coding neomers can create neoantigens with predicted strong binding levels that were subsequently validated (Data S6). From the 1700 neoantigens with strong binding levels, only 1.72 (p-value < 1e-8, hypergeometric test) is expected to correspond to a neomer, suggesting that missense mutations also resulting in neomers are 7-fold more likely to also generate strongly binding neoantigens.
Neomers can be used as a tool to identify cancer-associated gene reg- ulatorymutations. Amongst the 210 prostate cancer promoter neomers, we selected five promoters and found that two of them significantly affected promoter activity due to the neomer. Their difference in activity was in line with the gene’s expression change in prostate cancer, with RPS2-SNHG9 having increased activity, fitting with RPS2 overexpression in prostate cancer94 and TMEM127-CIAO1 abolishing activity, in line with TMEM127 observed downregulation in cancer63. Our MPRA library of 4609 neomer- causing mutations in enhancers revealed that 2.6% can change gene expression by >1.5-fold, suggesting that a subset of these mutations could have important functional consequences in cancer.While we obtained good correlations between our three MPRA replicates, MPRA has some caveats that should be noted: (1) testing genes outside of their genomic context; (2) not being able to assign the target gene; (3) utilizing a minimal promoter instead of the target promoter; (4) using cell lines and one prostate cancer cell type. Nonetheless, understanding tumor onset and its transition to a metastatic cancer has been limited mostly to coding genes or their pre- determined regulatory sequences95,96. With the vast majority of mutations falling into non-coding regions, making them much more challenging to interpret, novel approaches are required for identifyingmutations that have functional consequences. Here, we show that by prioritizingmutations that result in neomers and utilizing MPRA, it is possible to shortlist candidates that can have a phenotypic impact. This approach can be applied to any cancer type and thus holds the potential to expand our catalog of func- tionally relevant cancer driver non-coding mutations.
In summary, we show that neomers can provide a powerful tool for cancer diagnosis. As they can easily be detected via sequence or CRISPR- based tools, it should be straightforward to integrate them in current routine
cancer diagnostic tests and their use could increase the sensitivity and specificity of these tests. Combining neomer-based screening with clinical characteristics and additional diagnostic tools and features could increase the positive predictive value. In addition, as cfDNA could also be isolated from urine and saliva, and detection of these sequences only requires a relatively small amount of DNA, neomer-based diagnosis could be carried out in a non-invasive manner. The utilization of neomers in the clinic will require additional work. This includes standardization of the collection and processing of patient samples, clinical trials that use neomers for cancer detection in a Clinical Laboratory Improvement Amendments certified lab and further refinement and standardization of our algorithms. Our work also suggests that neomers could be used to highlight cancer-associatedgene regulatory mutations, which have been difficult to identify. Further high- throughput characterization of thesemutations could allow the detection of bona fide cancer-associated functional regulatory mutations that could be used for diagnosis and treatment.
Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.
Data availability The source data for Fig. 1 is found in Data S1, for Fig. 3 in Data S3, and for Fig. 4 in Data S5. MPRA sequencing data was deposited in the NCBI short read archive (SRA) with accession code PRJNA917083. ICGC samples are publicly available and can be accessed fromhttps://docs.icgc-argo.org/docs/ data-access/icgc-25k-data. Due to patient data confidentiality, WGS sequencing data is only available upon request. For WGS sequencing, contact Nadav Ahituv.
Code availability We generated an easy-to-use software package that enables performing nullomer cancer analyses from sequence-based datasets97. The package is composed of six functions: (1) EnumerateNullomers, which extracts all nullomers of specified kmer lengths in a FASTA sample; (2) Extra- ctMutationNullomers, which finds all mutations that cause the resurfacing of a list of nullomers; (3) IdentifyRecurrentNullomers, which identifies nullomers that recur in a dataset through mutagenesis; (4) FindAlmost- Nullomers, which identifies the positions that can create a list of nullomers genome-wide for every possible substitution and single base-pair insertion and deletion; (5) FindNullomerVariants, which removes nullomers that are likely to result fromcommonvariants in a user specified variantVCFfile; (6) FindDNANullomersFromReads, which performs the identification of nullomers in raw read samples. The package can be found at https://doi.org/ 10.5281/zenodo.1091017397 and the documentation on Read the Docs (https://nullomerator.readthedocs.io/en/latest/#) provides in-depth details on how to use the software.
References 1. Cancer. https://www.who.int/news-room/fact-sheets/detail/cancer. 2. Siegel, R. L., Miller, K. D. & Jemal, A. Cancer statistics, 2020. CA
Cancer J. Clin. 70, 7–30 (2020). 3. Hawkes, N. Cancer survival data emphasise importance of early
diagnosis. BMJ 364, l408 (2019). 4. Etzioni, R. et al. The case for early detection. Nat. Rev. Cancer 3,
243–252 (2003). 5. Heitzer, E., Auinger, L. & Speicher,M. R. Cell-freeDNAand apoptosis:
how dead cells inform about the living. TrendsMol. Med. 26, 519–528 (2020).
6. Bronkhorst, A. J., Ungerer, V. & Holdenrieder, S. The emerging role of cell-free DNA as amolecular marker for cancermanagement.Biomol. Detect. Quantif. 17, 100087 (2019).
https://doi.org/10.1038/s43856-025-01067-3 Article
Communications Medicine | (2025) 5:363 11
7. Brill, J. V. Screening for cancer: the economic, medical, and psychosocial issues. Am. J. Manag. Care 26, S300–S306 (2020).
8. Zill, O. A., Banks, K. C., Fairclough, S. R. & Mortimer, S. A. The landscape of actionable genomic alterations in cell-free circulating tumor DNA from 21,807 advanced cancer patients.Clin. Cancer Res. 3528–3538 (2018).
9. Saghafinia, S., Mina, M., Riggi, N., Hanahan, D. & Ciriello, G. Pan- cancer landscapeof aberrantDNAmethylationacrosshuman tumors. Cell Rep. 25, 1066–1080.e8 (2018).
10. Sadeh, R. et al. ChIP-seq of plasma cell-free nucleosomes identifies gene expression programs of the cells of origin. Nat. Biotechnol. https://doi.org/10.1038/s41587-020-00775-6 (2021).
11. Barbany, G. et al. Cell-free tumour DNA testing for early detection of cancer-a potential future tool. J. Intern. Med. 286, 118–136 (2019).
12. Liu, J. et al. Genome-wide cell-free DNA methylation analyses improve accuracy of non-invasive diagnostic imaging for early-stage breast cancer.Mol. Cancer 20, 36 (2021).
13. Magbanua, M. J. M. et al. Cell-free DNA concentration as a biomarker of response and recurrence inHER2-negative breast cancer receiving neoadjuvant chemotherapy. Clin. Cancer Res. 30, 2444–2451 (2024).
14. Razavi, P. et al. High-intensity sequencing reveals the sources of plasma circulating cell-free DNA variants. Nat. Med. 25, 1928–1937 (2019).
15. Shi, J., Zhang, R., Li, J. & Zhang, R. Size profile of cell-free DNA: a beacon guiding the practice and innovation of clinical testing. Theranostics 10, 4737–4748 (2020).
16. Ji, L. et al. Methylated DNA is over-represented in whole-genome bisulfite sequencing data. Front. Genet. 5, 341 (2014).
17. WormØrntoft,M.-B.Reviewofblood-basedcolorectal cancer screening: how far are circulating cell-free DNAmethylation markers from clinical implementation? Clin. Colorectal Cancer 17, e415–e433 (2018).
18. Warton, K. & Samimi, G.Methylation of cell-free circulatingDNA in the diagnosis of cancer. Front. Mol. Biosci. 2, 13 (2015).
19. Hampikian, G. & Andersen, T. Absent sequences: nullomers and primes. in Biocomputing 2007 355–366 https://doi.org/10.1142/ 9789812772435_0034 (World Scientific, 2006).
20. Vergni, D. & Santoni, D. Nullomers and high order nullomers in genomic sequences. PLoS ONE 11, e0164540 (2016).
21. Georgakopoulos-Soares, I., Yizhar-Barnea, O., Mouratidis, I., Hemberg, M. & Ahituv, N. Absent from DNA and protein: genomic characterization of nullomers and nullpeptides across functional categories and evolution. Genome Biol. 22, 245 (2021).
22. Montgomery, A. et al. Utilizing nullomers in cell-free RNA for early cancer detection. Cancer Gene Ther. 31, 861–870 (2024).
23. The Cancer Genome Atlas Program. https://www.cancer.gov/tcga (2018).
24. Jiao,W.et al. Adeep learningsystemaccurately classifiesprimaryand metastatic cancers using passenger mutation patterns. Nat. Commun. 11, 728 (2020).
25. Henderson, J. T.,Webber, E.M. &Sawaya,G. F.Screening forOvarian Cancer:AnUpdatedEvidenceReview for theU.S. PreventiveServices Task Force (Agency for Healthcare Research and Quality (US), 2018).
26. Heintz, A. P. M. et al. Carcinoma of the ovary. Int. J. Gynecol. Obstet. 95, S161–S192 (2006).
27. Gonzalez-Perez, A. et al. IntOGen-mutations identifies cancer drivers across tumor types. Nat. Methods 10, 1081–1082 (2013).
28. Chang, C.-C. & Lin, C.-J. LIBSVM: a library for support vector machines. ACM Trans. Intell. Syst. Technol. 2, 1–27 (2011).
29. Zou, X. et al. A systematic CRISPR screen defines mutational mechanisms underpinning signatures caused by replication errors and endogenous DNA damage. Nat. Cancer 2, 643–657 (2021).
30. Georgakopoulos-Soares, I. et al. Transcription-coupled repair and mismatch repair contribute towards preserving genome integrity at mononucleotide repeat tracts. Nat. Commun. 11, 1980 (2020).
31. Pećina-Šlaus, N., Kafka, A., Salamon, I. & Bukovac, A. Mismatch repair pathway,genomestability andcancer.Front.Mol.Biosci.7, 122 (2020).
32. Garmezy, B. et al. Correlation of pathogenic POLE mutations with clinical benefit to immune checkpoint inhibitor therapy. J. Clin. Orthod. 38, 3008–3008 (2020).
33. Karczewski, K. J. et al. Themutational constraint spectrum quantified from variation in 141,456 humans. Nature 581, 434–443 (2020).
34. Cristiano, S. et al. Genome-wide cell-free DNA fragmentation in patients with cancer. Nature 570, 385–389 (2019).
35. Danecek, P. et al. Twelve years of SAMtools and BCFtools. Gigascience 10, giab008 (2021).
36. Wong, D. et al. Cell-free DNA from germline TP53 mutation carriers reflect cancer-like fragmentation patterns. Nat. Commun. 15, 7386 (2024).
37. Gordon, M. G. et al. lentiMPRA and MPRAflow for high-throughput functional characterization of gene regulatory elements. Nat. Protoc. 15, 2387–2412 (2020).
38. ICGC/TCGA Pan-Cancer Analysis of Whole Genomes Consortium. Pan-cancer analysis of whole genomes. Nature 578, 82–93 (2020).
39. Prior, I. A., Lewis, P. D. & Mattos, C. A comprehensive survey of Ras mutations in cancer. Cancer Res. 72, 2457–2467 (2012).
40. Muñoz-Maldonado, C., Zimmer, Y. & Medová, M. A comparative analysis of individual RAS mutations in cancer biology. Front. Oncol. 9, 1088 (2019).
41. Vinagre, J. et al. Frequency of TERT promoter mutations in human cancers. Nat. Commun. 4, 2185 (2013).
42. Dratwa, M., Wysoczańska, B., Łacina, P., Kubik, T. & Bogunia-Kubik, K. TERT-regulation and roles in cancer formation. Front. Immunol. 11, 589929 (2020).
43. Heidenreich, B., Rachakonda, P. S., Hemminki, K. & Kumar, R. TERT promoter mutations in cancer development. Curr. Opin. Genet. Dev. 24, 30–37 (2014).
44. Song, J. H. et al. Small-molecule-targeting hairpin loop of hTERT promoter G-quadruplex induces cancer cell death. Cell Chem. Biol. 26, 1110–1121.e4 (2019).
45. Bell, R. J. A. et al. Cancer. The transcription factor GABP selectively binds and activates the mutant TERT promoter in cancer. Science 348, 1036–1039 (2015).
46. Powter, B. et al. Human TERT promoter mutations as a prognostic biomarker in glioma. J. Cancer Res. Clin. Oncol. 147, 1007–1017 (2021).
47. Romero, J.M. et al. A four-chemokine signature is associatedwith aT- cell-inflamedphenotype in primary andmetastatic pancreatic cancer. Clin. Cancer Res. 26, 1997–2010 (2020).
48. Antal,C. E. et al. Cancer-associatedprotein kinaseCmutations reveal kinase’s role as tumor suppressor. Cell 160, 489–502 (2015).
49. Vincent, A., Omura, N., Hong, S.M., Jaffe, A. & Eshleman, J. Genome- wide analysis of promoter methylation associated with gene expression profile in pancreatic adenocarcinoma. Clin. Cancer Res. 17, 4341–4354(2011).
50. Xu, E. et al. Genome-widemethylation analysis showssimilar patterns in Barrett’s esophagus and esophageal adenocarcinoma. Carcinogenesis 34, 2750–2756 (2013).
51. Strand, M., Earley, M. C., Crouse, G. F. & Petes, T. D. Mutations in the MSH3 gene preferentially lead to deletions within tracts of simple repetitive DNA in Saccharomyces cerevisiae. Proc. Natl. Acad. Sci. USA 92, 10418–10421 (1995).
52. Nik-Zainal, S. et al. Landscape of somatic mutations in 560 breast cancer whole-genome sequences. Nature 534, 47–54 (2016).
53. Wang, F. et al. Evaluation of POLE and POLD1 mutations as biomarkers for immunotherapy outcomes across multiple cancer types. JAMA Oncol. 5, 1504–1506 (2019).
https://doi.org/10.1038/s43856-025-01067-3 Article
Communications Medicine | (2025) 5:363 12
54. Alexandrov, L. B. et al. Signatures of mutational processes in human cancer. Nature 500, 415–421 (2013).
55. Thai, A. A., Solomon, B. J., Sequist, L. V., Gainor, J. F. & Heist, R. S. Lung cancer. Lancet 398, 535–554 (2021).
56. Blandin Knight, S. et al. Progress and prospects of early detection in lung cancer. Open Biol. 7, 170070 (2017).
57. de Groot, P. M., Wu, C. C., Carter, B. W. & Munden, R. F. The epidemiologyof lungcancer.Transl. LungCancerRes.7, 220–233 (2018).
58. Poulos,R.C., Sloane,M.A.,Hesson, L.B.&Wong, J.W.H. Thesearch for cis-regulatory driver mutations in cancer genomes. Oncotarget 6, 32509–32525 (2015).
59. Elliott, K. & Larsson, E. Non-coding drivermutations in human cancer. Nat. Rev. Cancer 21, 500–509 (2021).
60. Aurilio, G. et al. Androgen receptor signaling pathway in prostate cancer: from genetics to clinical applications. Cells 9, 2653 (2020).
61. ENCODE Project Consortium. An integrated encyclopedia of DNA elements in the human genome. Nature 489, 57–74 (2012).
62. Ohkia, A., Hu, Y., Wang, M., Garcia, F. U. & Stearns, M. E. Evidence for prostate cancer-associated diagnostic marker-1: immunohistochemistry and in situ hybridization studies. Clin. Cancer Res. 10, 2452–2458 (2004).
63. Qin, Y. et al. The tumor susceptibility gene TMEM127 is mutated in renal cell carcinomas and modulates endolysosomal function. Hum. Mol. Genet. 23, 2428–2439 (2014).
64. Bowler, E. et al. Hypoxia leads to significant changes in alternative splicing and elevated expression of CLK splice factor kinases in PC3 prostate cancer cells. BMC Cancer 18, 355 (2018).
65. El-Haibi, C. P. et al. Differential G protein subunit expression by prostate cancer cells and their interactionwith CXCR5.Mol. Cancer 12, 64 (2013).
66. Valencia, T. et al. Role and expression of FRS2 and FRS3 in prostate cancer. BMC Cancer 11, 484 (2011).
67. Yu, F. et al.Wnt/β-catenin signaling in cancers and targeted therapies. Signal Transduct. Target. Ther. 6, 307 (2021).
68. Kypta, R. M. & Waxman, J. Wnt/β-catenin signalling in prostate cancer. Nat. Rev. Urol. 9, 418–428 (2012).
69. Schneider, J. A. & Logan, S. K. Revisiting the role of Wnt/β-catenin signaling in prostate cancer.Mol. Cell. Endocrinol. 462, 3–8 (2018).
70. Wang, C., Chen, Q. & Xu, H. Wnt/β-catenin signal transduction pathway in prostate cancer and associated drug resistance. Discov. Oncol. 12, 40 (2021).
71. Grant, C. E., Bailey, T. L. & Noble, W. S. FIMO: scanning for occurrences of a given motif. Bioinformatics 27, 1017–1018 (2011).
72. Castro-Mondragon, J. A. et al. JASPAR 2022: the 9th release of the open-accessdatabaseof transcription factorbindingprofiles.Nucleic Acids Res. 50, D165–D173 (2022).
73. Vatapalli, R. et al. Histone methyltransferase DOT1L coordinates AR and MYC stability in prostate cancer. Nat. Commun. 11, 4153 (2020).
74. Dhingra, P. et al. Identification of novel prostate cancer drivers using RegNetDriver: a framework for integration of genetic and epigenetic alterations with tissue-specific regulatory network. Genome Biol. 18, (2017).
75. Zhang, Z. et al. HERC3 regulates epithelial-mesenchymal transition by directly ubiquitination degradation EIF5A2 and inhibits metastasis of colorectal cancer. Cell Death Dis. 13, 1–12 (2022).
76. Schult, T. A. et al. Screening human lung cancer with predictive models of serum magnetic resonance spectroscopy metabolomics. Proc. Natl. Acad. Sci. USA. 118, e2110633118 (2021).
77. Moyer, V. A. & Preventive Services Task Force, U. S. Screening for lung cancer: U.S. Preventive Services Task Force recommendation statement. Ann. Intern. Med. 160, 330–338 (2014).
78. Smith, R. A. et al. Cancer screening in the United States, 2019: a review of current American Cancer Society guidelines and current issues in cancer screening. CA Cancer J. Clin. 69, 184–210 (2019).
79. Wood, D. E. et al. Lung Cancer Screening, version 3.2018, NCCN Clinical Practice Guidelines in Oncology. J. Natl. Compr. Cancer Netw. 16, 412–441 (2018).
80. Rendle, K. A. et al. Rates of downstream procedures and complications associated with lung cancer screening in routine clinical practice: A retrospective cohort study. Ann. Intern. Med. 177, 18–28 (2024).
81. Mazzone, P. J. et al. Clinical validationof a cell-freeDNA fragmentome assay for augmentationof lungcancer early detection.CancerDiscov. https://doi.org/10.1158/2159-8290.CD-24-0519 (2024).
82. Lin, W.-H. et al. Circulating tumor DNA methylation marker MYO1-G for diagnosis and monitoring of colorectal cancer. Clin. Epigenetics 13, 232 (2021).
83. Bae,M. et al. Integrativemodelingof tumorgenomesandepigenomes for enhanced cancer diagnosis by cell-free DNA. Nat. Commun. 14, 2017 (2023).
84. Kaminski, M. M., Abudayyeh, O. O., Gootenberg, J. S., Zhang, F. & Collins, J. J. CRISPR-based diagnostics. Nat. Biomed. Eng. 5, 643–656 (2021).
85. Ackerman, C. M. et al. Massively multiplexed nucleic acid detection with Cas13. Nature 582, 277–282 (2020).
86. Augustus, E. et al. The art of obtaining a high yield of cell-free DNA from urine. PLoS ONE 15, e0231058 (2020).
87. Ding, S. et al. Saliva-derived cfDNA is applicable for EGFR mutation detection but not for quantitation analysis in non-small cell lung cancer. Thorac. Cancer 10, 1973–1983 (2019).
88. Logsdon, G. A., Vollger, M. R. & Eichler, E. E. Long-read human genome sequencing and its applications. Nat. Rev. Genet. 21, 597–614 (2020).
89. Cheng,C., Fei, Z. & Xiao, P.Methods to improve the accuracy of next- generation sequencing.Front. Bioeng. Biotechnol. 11, 982111 (2023).
90. Lennon, A.M. et al. Feasibility of blood testing combinedwith PET-CT to screen for cancer and guide intervention. Science 369, eabb9601 (2020).
91. Alileche, A. & Hampikian, G. The effect of Nullomer-derived peptides 9R, 9S1R and 124R on the NCI-60 panel and normal cell lines. BMC Cancer 17, 533 (2017).
92. Koulouras, G. & Frith,M. C. Significant non-existence of sequences in genomes and proteomes. Nucleic Acids Res. https://doi.org/10. 1093/nar/gkab139 (2021).
93. Vita, R. et al. The Immune Epitope Database (IEDB): 2018 update. Nucleic Acids Res. 47, D339–D343 (2019).
94. Edlind, M. P. & Hsieh, A. C. PI3K-AKT-mTOR signaling in prostate cancer progression and androgen deprivation therapy resistance. Asian J. Androl. 16, 378–386 (2014).
95. Chiang, A. C. & Massagué, J. Molecular basis of metastasis. N. Engl. J. Med. 359, 2814–2823 (2008).
96. Nguyen, D. X. & Massagué, J. Genetic determinants of cancer metastasis. Nat. Rev. Genet. 8, 341–352 (2007).
97. Easterlin, R. et al. Nullomerator. Zenodo. https://doi.org/10.5281/ zenodo.10910173 (2024).
Acknowledgements We wish to thank the individuals who participated in this study. We dedicate this article to Felix Feng (UCSF) and thank him for his advice, support and inspiration for this study. This work was supported in part by: Benioff Initiative for Prostate Cancer Research (NA). UCSF Catalyst award (NA). UCSF Innovations Ventures Philanthropy Fund (NA). National Human Genome Research Institute grant number UM1HG011966 (NA). Food and Drug Administration (FDA) of the U.S. Department of Health and Human Services (HHS) as part of a financial assistance award [Center of Excellence in Regulatory Science and Innovation grant to University of California, San Francisco, U01FD005978] award number 3U01FD005978-09S3 (NA, JC, MH). Core funding from the Wellcome Trust (MH). Core funding from the Evergrande Center (MH). Core funding from the Penn State College of Medicine (IGS). ElsaU.PardeeFoundation (IGS). FourDiamonds (IGS).Penn State Cancer Institute (IGS). The contents are those of the author(s) and do not necessarily represent the official views of, nor an endorsement, by FDA/ HHS, or the U.S. Government.
https://doi.org/10.1038/s43856-025-01067-3 Article
Communications Medicine | (2025) 5:363 13
Author contributions Conceptualization: I.G.S.,O.Y.B., I.M.,M.H.,N.A.Computationalmethodology: I.G.S., I.M.,C.C.,R.E.,E.C.,G.E.P.,M.K.,M.H.Analysisandvisualization: I.G.S., O.Y.B., I.M., C.C., M.P., A.N., M.M., M.K., M.H., N.A. Provided samples: G.P., J.S.W., C.A.M., F.F., I.V., N.S., E.P., A.C., K.S., J.C. Experimental assays: O.Y.B.,C.C., R.B., J.S., D.L.C., R.E., J.K. Fundingacquisition: I.G.S.,M.H.,N.A. Project administration: I.G.S.,M.H., N.A. Supervision: I.G.S.,M.H., N.A.Writing —original draft: I.G.S., O.Y.B., I.M., M.H., and N.A.
Competing interests The authors declare the following competing interests: I.G.S., O.Y.B., I.M., M.H., and N.A. have filed patent applications covering embodiments and concepts disclosed in the manuscript.
Additional information Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s43856-025-01067-3.
Correspondence and requests for materials should be addressed to Ilias Georgakopoulos-Soares, Martin Hemberg or Nadav Ahituv.
Peer review information Communications Medicine thanks the anonymous reviewers for their contribution to the peer review of this work.
Reprints and permissions information is available at http://www.nature.com/reprints
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in anymedium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
© The Author(s) 2025
1Department of Bioengineering and Therapeutic Sciences, University of California San Francisco, San Francisco, CA, USA. 2Institute for HumanGenetics, University of California San Francisco, San Francisco, CA, USA. 3Institute for Personalized Medicine, Department of Molecular and Precision Medicine, The Pennsylvania State University College of Medicine, Hershey, PA, USA. 4Huck Institutes of the Life Sciences, Penn State University, University Park, PA, USA. 5The Gene Lay Institute of Immunology and Inflammation, Brigham andWomen’s Hospital, Massachusetts General Hospital and Harvard Medical School, Boston, MA, USA. 6Department of Epidemiology and Biostatistics, University of California San Francisco, San Francisco, CA, USA. 7Division of Gynecologic Oncology, University of California San Francisco, San Francisco, CA, USA. 8Donnelly Centre for Cellular and Biomolecular Research, University of Toronto, Toronto, ON, Canada. 9Department of Epide- miology and Population Health, Stanford University, Stanford, CA, USA. 10Division of Oncology, Department of Internal Medicine, Washington University School of Medicine, St. Louis, MO, USA. 11Siteman Cancer Center, Washington University School of Medicine, St. Louis, MO, USA. 12Department of Biomedical Engineering, Washington University School of Medicine, St. Louis, MO, USA. 13Division of Hematology/Oncology, Department of Medicine, University of California San Francisco, San Francisco, CA, USA. 14Helen Diller Comprehensive Cancer Center, University of California San Francisco, San Francisco, CA, USA. 15Department of Radiation Oncology, University of California San Francisco, San Francisco, CA, USA. 16Department of Urology, University of California San Francisco, San Francisco, CA, USA. 17Third Department of Internal Medicine, Sotiria Hospital, National and Kapodistrian University of Athens, School of Medicine, Athens, Greece. 18Department of Pathology,YaleSchool ofMedicine,NewHaven,CT,USA. 19BreastOncology,Dana-FarberBrighamCancerCenter,Boston,MA,USA. 20KaiserPermanente,Oakland, CA, USA. 21Wellcome Sanger Institute, Hinxton, Cambridge, UK. 22Helen Diller Family Comprehensive Cancer Center, University of California San Francisco, San Francisco, CA, USA. 23These authors contributed equally: Ilias Georgakopoulos-Soares, Ofer Yizhar-Barnea, Ioannis Mouratidis. e-mail: [email protected]; [email protected]; [email protected]
https://doi.org/10.1038/s43856-025-01067-3 Article
Communications Medicine | (2025) 5:363 14 https://bit.ly/3W3t38u
