Whole‐genome sequencing of cell‐free DNA for assessment of minimal residual disease in high‐risk smoldering multiple myeloma

Abstract

In multiple myeloma (MM), minimal residual disease (MRD) is an established endpoint for accelerated drug approval but is limited by a need for serial invasive bone marrow (BM) biopsies, which may under‐sample spatial heterogeneity and result in false negativity. Furthermore, longitudinal (i.e., sustained) MRD negativity is emerging as a powerful tool for clinical decision‐making. To these ends, reliable systemic MRD assessment is a growing need. Next‐generation sequencing approaches for plasma cell‐free (cf) DNA generally fail to achieve adequate detection limits in low tumor fraction (TF) settings (i.e., MRD), but tumor‐informed approaches leveraging whole‐genome sequencing (WGS) have thus far achieved the lowest limits of detection (LODs). We therefore performed a longitudinal analysis of MRD assessed by serial WGS of plasma cfDNA as compared to clinical standard flow‐cytometric BM MRD in high‐risk smoldering MM. 25 baseline tumor WGS served to inform detection of disease in 87 sequential plasma samples. The median LOD across patients was 1.2 × 10⁻⁴ (range 9.0 × 10⁻⁵ – 2.1 × 10⁻⁴). TF was prognostic, and tumors with high‐risk genomics had higher baseline TF (P = 0.002) and eventual disease progression (n = 7, P < 0.001). Furthermore, dynamic changes in serially tracked TF portended outcome. In comparison to BM flow, cfDNA WGS was concordant in 33/45 (73.3%) MRD samples with consistent capture of BM‐positive cases, but added resolution to apparent false‐negative BM samples. Overall, despite the deeper LOD of localized BM flow, cfDNA WGS can detect MRD when BM flow did not, adding dynamic and systemic/spatial resolution to standard hyper‐local MRD assessment.

INTRODUCTION

The assessment of minimal residual disease (MRD) in multiple myeloma (MM) has emerged as the most powerful clinical response assessment and has both prognostic and regulatory ramifications. ¹ , ² , ³ , ⁴ , ⁵ , ⁶ Currently, clinical assessment of MRD is limited to local assay via blind bone marrow biopsy and enumeration of plasma cells either by multi‐parametric flow cytometry or by PCR amplification and sequencing of the clonal CDR3 region (i.e., Adaptive Clonoseq). These localized assays, however, fail to capture known spatial heterogeneity in MM, and may be limited in settings of extramedullary disease, which increases in prevalence through advancing stages of disease. ⁷ , ⁸ , ⁹ , ¹⁰ , ¹¹ , ¹² In particular, as treatment paradigms advance toward T‐cell redirection, extramedullary disease is becoming increasingly prevalent, requiring the development of MRD approaches more sensitive to systemic disease burden. ¹³ Furthermore, the reliance on invasive bone marrow biopsy makes serial assessments logistically challenging and inconvenient for patients, limiting the potential applications of MRD for tracking disease dynamics. ¹⁴ Given these limitations, there is a need to develop reproducible peripheral assessment of MRD.

Attempts to leverage prior technologies for peripheral assessment of MRD are generally limited by the need to evaluate large blood volumes to achieve informative limits of detection (LODs). ¹⁵ , ¹⁶ Instead, the assessment of circulating tumor DNA (ctDNA) from within a patient's complement of cell‐free DNA (cfDNA) has emerged as a promising avenue for disease assessment in multiple tumor types, including MM. ¹⁷ , ¹⁸ , ¹⁹ , ²⁰ , ²¹ , ²² , ²³ Previous studies have leveraged target‐capture panels, whole‐exome sequencing, and low‐pass whole‐genome sequencing (WGS) to identify MM‐specific somatic variants in plasma cfDNA. Generally, de novo approaches fail to achieve the sensitivity and specificity needed for pushing the boundaries of detection limits in low tumor fraction (TF) settings (i.e., MRD). Specifically, with WGS, overall lower coverage limits the confidence of mutation calling and differentiation from artifact or sequencing error. Therefore, tumor‐informed approaches, where baseline bone marrow‐residing tumor is profiled as ground‐truth and somatic variants are tracked in cfDNA, have thus far been able to achieve the lowest LODs. ²³ , ²⁴ Furthermore, compared to other next‐generation sequencing technologies, WGS can reveal the complete somatic landscape of MM inclusive of prognostic single‐nucleotide variants (SNVs), large and focal copy number alterations (CNAs), structural variants (SVs), and mutational signatures, which can offer prognostic information in addition to MRD assessment. ²⁵ , ²⁶ , ²⁷ , ²⁸

To address the unmet need for peripheral MRD assessment in MM, we utilized tumor‐informed genome‐wide mutational tracking from WGS of cfDNA. By selecting a cohort of high‐risk smoldering multiple myeloma (HR‐SMM) treated on an interventional clinical trial, our approach facilitated pretreatment comprehensive genomic characterization and assessment of systemic (i.e., circulating) disease burden, as well as longitudinal assessment of response dynamics at deep LODs that predict clinical outcome.

METHODS

Sample selection

All samples and data were collected in accordance with the Declaration of Helsinki, and all patients provided written informed consent for the study and subsequent correlative analyses. All samples were collected on NCT01572480, an interventional study for HR‐SMM. ²⁹ Twenty‐five patients had baseline WGS of CD138 + BM plasma cells and longitudinal plasma samples collected in Cell‐Free DNA BCT Streck tubes. All patients underwent protocolized MRD assessment by BM flow cytometry at the end of combination therapy and yearly thereafter, from first‐pull aspirate samples to minimize the effect of hemodilution. BM flow cytometry was performed using a standardized 8‐color, two‐tube approach with established LOD to 10⁻⁵ with greater than 2,000,000 cells as input. ³⁰ cfDNA was sent for MRD analysis at the end of combination therapy, coinciding with BM MRD assessment for all cases. Serial plasma was assayed where available. In the event of BM MRD conversion or progression of disease, cfDNA MRD was also assessed at that time point, and 3 and 6 months prior, as samples allowed.

Sample processing for CD138‐selected BM plasma cells

Processing of CD138‐selected BM plasma cells for baseline WGS has been described previously ³¹ (Supplemental Methods). Briefly, however, BM mononuclear cell samples were thawed and subject to CD138+ microbead capture. Germline match was selected for each patient from peripheral blood mononuclear cells. DNA quantification was performed via PicoGreen and quality control was performed using an Agilent bioanalyzer. ∼500 ng of genomic DNA was sheared and sequencing libraries were prepared using a modified KAPA HyperPrep Kit (KK8504; Kapa Biosystems). Samples were run on a NovaSeq. 6000 in a 150 bp/150 bp paired‐end run, using the NovaSeq. 6000 SBS v1 kit and an S4 Flow Cell (Illumina). Target coverage depth was ×80 for tumor and ×40 for normal.

Sample processing for plasma cfDNA

For plasma cfDNA samples, Streck collection tubes were centrifuged to separate plasma into 1 mL aliquots. One aliquot from each sample time point was sent to the New York Genome Center (NYGC) for DNA extraction and 40× WGS as previously described. ³² Briefly, DNA extraction was performed using the Mag‐Bind cfDNA kit (Omega Bio‐Tek). ³² DNA concentration was quantified on a Qubit Fluorometer (Thermo Fisher) and fragment analysis for cfDNA fragment size and genomic DNA contamination was performed using the High Sensitivity Genomic DNA Analysis Kit (Agilent). The KAPA Hyper Library Preparation kit was used for library preparation. Samples with a mass >1 ng were then prepared for sequencing on an Illumina HiSeq X. Generally, for samples with more than 5 ng of cfDNA, five PCR cycles were performed, while 7–10 cycles were performed for cases with less than 5 ng of cfDNA.

Whole‐genome analytical pipeline

The analytical pipeline for the baseline BM tumor/normal pairs has been previously described using a combination of the NYGC and MGP1000 somatic pipelines. ³¹ , ³³ Reads were aligned to the reference genome (GRCh38) using the Burrows–Wheeler Aligner (v0.7.15). Somatic mutations were identified using high‐confidence consensus calls from at least two of Mutect2, Lancet, and/or Strelka2. ³⁴ , ³⁵ , ³⁶ Indels were identified using Lancet, Mutect2, SvABA, and Strelka2. ³⁴ , ³⁵ , ³⁶ Clonality of SNV was determined by calculating a cancer cell fraction for each variant defined as the variant allele frequency corrected for ploidy and purity. ³⁷ Copy number and tumor purity were evaluated using ASCAT ³⁸ ; IgCaller was used for translocations at the immunoglobulin loci ³⁹ ; and SVs were defined using Manta, DELLY, and SvABA. ³⁵ , ⁴⁰ , ⁴¹ To consider the preferential representation in cfDNA for SNV duplicated across gains, specifically in hyperdiploid disease states, we corrected allele fraction for purity and considered SNV with allele fraction 0.66 as duplicated and 0.33 as unduplicated in a 3:1 total:minor copy number state.

Mutational signatures were analyzed across all WGS samples as has been previously described. ³¹ Signature fitting is error‐prone when <50 SNVs are used as input. ⁴² Therefore, when fitting mutational signatures to ctDNA SNVs, subgroups were collapsed and pooled for accuracy. ⁴³ For genomic feature selection, we adapted a locus‐based classification scheme that has been previously used in newly diagnosed MM ²⁶ (https://github.com/UM-Myeloma-Genomics/GCP_MM). To consider feature association with cfDNA TF, we performed Wilcoxon's Signed Rank test for tumor fraction in reference to combinations of monoallelic events, biallelic events, and absence of the genomic feature in question. Sensitivity analysis, where appropriate, was performed by removing two HR‐SMM cases that would be reclassified as MM per IMWG 2014 diagnostic criteria. ⁴⁴

For cfDNA analysis, WGS reads for all primary BM tumor, normal germline match, and plasma cfDNA samples were demultiplexed into FASTQ files with Illumina bcl2fastq (v2.17.1.14). Adapter trimming, duplicate marking and sorting, base quality score realignment, and contamination quality control were performed as previously described and a read‐centric support vector machine (SVM) trained on Novaseq data was used for further denoising and SNV filtering. ³² Total aligned bases were converted into read counts using mean aligned read length. IchorCNA was additionally used for all samples to attempt measurement of copy number alterations and TF in plasma, though TF was sufficiently low that copy number alterations were undetectable in all plasma samples. Detection of patient‐specific SNVs was carried out by intersecting patient‐specific compendia (from primary tumor high‐confidence calls) with corresponding sites in plasma samples (MRDetect pipeline). ³² Fragmentation profiles were examined by calling the density function in base R, using the “Gaussian” kernel, on fragments supporting tumor‐derived SNV as compared to those supporting reference variants. Specifically, fragments supporting the tumor SNV compendium were compared to reference fragments at these loci.

Calculation of assay limit of detection (LOD) and tumor fraction

For each plasma sample, a site‐level detection rate was calculated as the number of tumor‐supporting SNV divided by the total number of reads interrogated at patient‐specific SNV loci.

To define a patient‐specific LOD, each patient's somatic SNV compendium was queried across unrelated plasma samples to estimate a background error distribution, excluding shared variants. Detection rates were converted into Z‐scores (detection rate–mean control detection rate/standard deviation across control detection rates) and a cutoff was selected that allowed rejection of 95% of controls (95% specificity) based on the precedent set by prior analyses and best fit to the data set. ³² Plasma sample detection rates exceeding the cutoff (LOD) were considered positive.

Tumor fraction was estimated using

$$\mathrm{TF}=1-{\left(1-\frac{M-\mu R}{N}\right)}^{\frac{1}{\text{cov}}},$$

where M denotes the number of tumor‐specific SNVs detected in plasma, μ is the mean background sequencing error rate, R is the total number of reads covering tumor‐specific loci, N is the tumor mutation burden, and cov is the mean local sequencing coverage across those loci. ³²

RESULTS

Patient cohort

To assess the utility of genome‐wide WGS of cfDNA, we performed a longitudinal analysis of patients with high‐risk SMM who had participated in an interventional clinical trial of carfilzomib, lenalidomide, and dexamethasone (8 cycles; 32 weeks), followed by lenalidomide maintenance for 2 years (NCT01572480). ²⁹ Clinical records were updated from prior publications to a new data cutoff of January 10, 2025 (median follow‐up; 65.7 months) for correlation of outcomes with cfDNA features (Supporting Information S1: Table S1). This trial enrolled 54 patients and baseline 80x WGS was available for 27 (50%). ³¹ It is noteworthy that, as previously reported, accrual on the clinical study began in 2012 and two patients who would today have been classified as having MM by SLiM criteria were included (BM plasma cells ≥60%; 2014 International Myeloma Working Group Criteria). Twenty‐five of these patients had serially collected longitudinal plasma samples available for assay by cfDNA WGS and were the focus of this study. Seven of 25 patients experienced clinical or biochemical disease progression and 9 (36%) experienced a loss of MRD negativity by flow cytometry (i.e., MRD conversion; Supporting Information S1: Table S1). Overall, we analyzed 112 WGS samples from 25 patients, including 25 baseline BM tumor WGS and 25 baseline plasma samples, with 62 follow‐up plasma samples. Of the follow‐up samples, 50 (81%) corresponded to bone marrow sampling time points assayed by standard‐of‐care bone marrow flow cytometry for MRD (Figure 1A). ³⁰ Each of the 87 cfDNA WGS was analyzed with the MRDetect pipeline, ³² in which cfDNA was interrogated for reads supporting a catalogue of high‐confidence SNV calls from each patient's primary BM WGS (tumor‐informed detection; Methods).

Figure 1.

Clinical and genomic features associated with cfDNA tumor fraction. (A) Study schema. Two patients colored in red would be classified as having MM by IMWG 2014 criteria as previously described. Twelve intercurrent/independent plasma samples do not have a corresponding BM sample. (B) Range of Estimated TFs of baseline plasma samples. High_TF is the top quartile of baseline TF. (C) Association of estimated TF with clinical features, including SMM risk assessment by 3 models. Progression of disease (PD) is defined as either of biochemical or clinical progression. Colored bars represent the interquartile range and gray bars are the range. HR, high‐risk, Sus, sustained MRD negativity. (D) Association of estimated TF with MM high‐risk genomic features. Mono; monoallelic event. (E) Patients in the top quartile of baseline TF have suboptimal outcome when treated with early KRd and R maintenance. BM, bone marrow; cfDNA, cell‐free DNA; MM, multiple myeloma; MRD, minimal residual disease; TF, tumor fraction.

Clinical and genomic features associated with circulating tumor DNA

We first examined whether the burden of circulating tumor DNA (ctDNA; i.e., tumor fraction; TF) was prognostic. The median TF of baseline plasma samples was 6 × 10⁻⁴ (range: 1 × 10⁻⁴–9 × 10⁻³; Figure 1B, Supporting Information S1: Table S2). A higher TF at baseline was associated with biochemical/clinical PD (Wilcoxon's Ranked‐Sum; P < 0.001). Similarly, higher TF was associated with failure to sustain MRD negativity (by BM flow) at last follow‐up (Wilcoxon; P = 0.011). Regarding TF and SMM risk, high‐risk status as assigned by 3 independent SMM risk scores was not significantly associated with plasma TF (Figure 1C). These observations are in line with the prior finding that SMM risk scores were unable to predict clinical outcome with intervention in this study. ³¹ Importantly, there was only a weak association between baseline BM plasma cell infiltration and plasma TF as has been described previously (Supporting Information S1: Figure S1A). ⁴⁵ We additionally repeated these comparisons, as sensitivity analyses, with removal of biochemical MM from the pool, without any change in the results. Altogether, these data suggest that baseline TF might be associated with biologically higher risk disease and is not solely a reflection of tumor burden.

To assess whether genomic features might predict the propensity of HR‐SMM to shed ctDNA, we examined the genomic profiles obtained from the primary BM WGS. Genomic features (including driver mutations, copy number abnormalities, canonical translocations, and mutational signatures) were selected from a de novo driver discovery performed across 1933 newly diagnosed MM used in a model to predict outcomes. ²⁶ , ⁴⁶ We then examined the association between each genomic feature and plasma TF (Supporting Information S1: Table S3). APOBEC mutational signature presence (SBS2/SBS13), deletion of 14q24.3, loss of the tumor suppressor MAX, and t(4;14) were each associated with higher plasma TF (Wilcoxon; P < 0.05, FDR; q < 0.1), although these features were generally not mutually exclusive. Importantly, these associations held true upon correction of TF for baseline bone marrow plasma cell infiltration. Furthermore, in line with this, previously defined genomic high risk (i.e., combinations of high‐risk features, including lesions contributing to MYC and NF‐KB deregulation, genomic instability in the form of APOBEC and chromothripsis, and t(4;14)) was associated with ctDNA burden, providing further evidence for the higher likelihood of biologically aggressive SMM to shed DNA (P = 0.002, Methods, Figure 1D, Supporting Information S1: Figure S1B, Supporting Information S1: Table S4). ³¹ Finally, we observed that plasma TF burden in the fourth quartile was associated with poor outcomes, in that these patients were more likely to experience biochemical or clinical progression despite early intervention (Figure 1E). Notably, TF adds further resolution to prognostication and interventional outcome prediction, given that all cases of HR‐SMM included in this analysis have been classified as genomic MM (i.e., progressive precursors) per a model built on 374 patients with precursor disease ⁴⁶ (Supporting Information S1: Table S4).

Calculating the LOD and comparison to bone marrow flow MRD

The calculation of the assay LOD is described in detail in the Methods. Briefly, however, each patient (i.e., tumor) is determined to have a unique LOD based on the mutational burden of the primary tumor, among other factors. In cfDNA fragments, reads that support high‐confidence SNVs identified in the primary tumor (baseline BM biopsy) are annotated. The number of SNV sites detected in cfDNA as a fraction of the total number of fragments (i.e., reads) provides the detection rate, which can facilitate calculation of a patient‐specific LOD. By comparing the detection rates to control detection rates obtained by querying tumor‐supporting SNVs in all unrelated plasma samples, the LOD for each patient was obtained. The median LOD across patients was 1.2 × 10⁻⁴ (range 9.0 × 10⁻⁵ – 2.1 × 10⁻⁴), and 59/87 (67.8%) plasma samples had detectable disease (i.e., detection rate above the LOD; Figure 2A, Supporting Information S1: Table S2). We then used the detection rate to estimate a TF for each patient with detectable disease. In the MRD setting (i.e., not baseline or at PD), these LODs translated into a median estimated TF of 1.4 × 10⁻⁴ (range 8.5 × 10⁻⁵–5.8 × 10⁻⁴, Supporting Information S1: Figure S2A).

Figure 2.

Individualized limits of detection and comparison with focal bone marrow flow cytometry MRD. (A) Site‐level SNV detection rates for each patient in the study. All plasma sample detection rates are plotted per patient and control detection rates are calculated by querying tumor‐supporting SNV sites in all unrelated plasma samples. The LOD (horizontal dashed line) is the sensitivity threshold at which 95% of controls are negative. (B) Sankey plot showing concordance of cfDNA WGS with BM flow cytometry. Only samples with corresponding plasma and BM collections are shown (n = 75). 12 intercurrent (plasma‐only) samples are excluded. (C) Swimmer plot demonstrating concordance of MRD assessments in clinical context. Patients with dark bars have either clinical or biochemical progression (PD). BM, bone marrow; cfDNA, cell‐free DNA; LOD, limit of detection; MRD, minimal residual disease; SNV, single‐nucleotide variants; WGS, whole‐genome sequencing.

Overall, these assay characteristics resulted in 59/87 (67.8%) plasma samples with detectable disease, including in baseline samples from all patients and in all samples collected at the time of disease progression. Many of the MRD assessments, however, were at odds with BM‐based MRD assessment with flow cytometry (Figure 2B). In longitudinal analysis of MRD time points, plasma cfDNA WGS recapitulated the results of BM flow in 33/45 (73.3%) cases. Specifically, plasma MRD positivity was concordant with BM flow for 11/12 (91.7%) BM MRD‐positive assays. However, one t(11;14) case seen to be MRD‐positive in BM at the end of combination therapy with simple genomic features (no complex SVs and few copy number abnormalities), and the lowest SNV count (~1800) was not detected in plasma (NIH005‐MRD‐pos; Supporting Information S1: Figure S2B). Sample QC revealed adequate coverage (×49) and expected reference cfDNA fragment characteristics (modal peak 167 base pairs [bp]; Supporting Information S1: Figure S2C). This suggests a limitation of the assay with very low tumor mutational burden compared to conventional approaches, but overall displays high concordance for MRD positivity in plasma when MRD is present in the bone marrow.

Conversely, 11/33 (33.3%) BM MRD‐negative time points were seen to instead be MRD‐positive in plasma (BM−/cfDNA+). Most discrepancies occurred in cases where BM MRD conversion (e.g., negative to positive) or PD later occurred, or in cases where flanking MRD BM assessments were MRD+ (8/11, 72.3%), emphasizing the value added of systemic cfDNA assessment in informing of potential false‐negative BM assessments (Figure 2C).

Addressing the discrepancy between cfDNA WGS and BM flow cytometry MRD Assessments

Given the discordance between cfDNA and BM assessments at low‐TF time points (e.g., MRD), specifically in cases where residual disease was detected in plasma, but not in BM (BM‐cfDNA+), we further interrogated the tumor‐supporting calls in plasma for possible false positivity in these 11 cases. First, we examined the ability of plasma cfDNA WGS to detect immunoglobulin genes with evidence of the somatic hypermutation variants found in the primary tumor sample, as these would be supportive of true residual disease. Predictably, these variants were most frequently detected in samples with higher TF, as tumor‐derived genomic elements spanning the immunoglobulin loci are more abundant (Figure 3A). As TF decreased, the likelihood of detecting these variants declined, even in cases with concordant positive MRD calls as compared to those showing discordance. The inability to reliably discern tumor‐specific immunoglobulin variants at low TF highlights an inherent limitation of single‐locus approaches in plasma, where informative tumor‐derived elements are sparse. These observations further underscore the advantage of genome‐wide interrogation to leverage a broader representation of tumor‐derived genomic elements.

Figure 3.

Support for discordant plasma MRD‐positive samples and fragmentomics. (A) Heatmap of all detected immunoglobulin gene variants emphasizing that detection likely hinges on ctDNA tumor fraction. (B) Proportion of clonal (to subclonal) variants in discordant BM‐negative, plasma‐positive cases (n = 11), with no significant change in clonal proportion across baseline marrow (left) and plasma (right). (C) Pooled sample mutational signature contributions to all 25 baseline plasma cases, all 22 plasma‐positive cases, and 5 of 11 discordant, plasma‐positive cases that had measurable baseline APOBEC contribution. SBS84 fitting was allowed, but contribution was 0. (D) Within‐patient changes in the proportion of cfDNA fragments <150 bp at minimal residual disease (MRD) and progression (PD) relative to baseline (BL). For patients with multiple MRD samples, fragmentation metrics were averaged to obtain a single MRD value per patient. Each line represents a patient. MRD values represent one collapsed MRD sample per patient. The dashed line indicates no change from baseline. Black horizontal ticks indicate the mean change; error bars show 95% confidence intervals. (E) Fragment density plots for 2 discordant plasma‐positive, BM‐negative MRD samples with visible tumor‐supporting fragment density peaks <150 bp. (F) Longitudinal fragmentomics reveal downshift in tumor‐supporting fragment size peaks with increasing tumor fraction from MRD time points until progression of disease.

We next assessed the clonality of ctDNA‐derived variants with regard to the primary tumor. We reasoned that variants of clonal origin in primary tumor (as opposed to subclonal variants) would be less likely to represent artifact or contamination by other somatic clonal processes (i.e., clonal hematopoiesis or monoclonal B‐cell lymphocytosis) when seen in plasma cfDNA and that maintenance of a stable ratio of detected clonal to subclonal variants would further support true positivity. In line with this, we observed no significant decrease in the proportion of clonal to subclonal variants between the original BM CD138+ selected WGS and the cfDNA tumor‐supporting variants (Figure 3B; Methods).

Next, we performed a mutational signatures analysis of tumor‐supporting variants in cfDNA. Here, we reasoned that the presence of APOBEC mutational activity would reflect true MM variants and that an absence of localized somatic hypermutation SBS84 would further exclude contaminant monoclonal B‐cell lymphocytosis. A challenge in this SMM setting is the relative low genomic complexity and low APOBEC contribution as compared to what is expected in more advanced disease stages. ²⁷ , ³¹ , ⁴⁷ , ⁴⁸ Despite this, for the 12 of 25 BM WGS (48%) with measurable APOBEC contribution, 7 (58%) cases had concordant detectable APOBEC contribution in baseline plasma ctDNA variants (including cases with too few plasma variants for signature fitting; Supporting Inforation S1: Figure S3, Supporting Information S1: Table S5, Methods). Next, in the MRD setting, due to the overall low numbers of ctDNA variants in each sample, we pooled SNVs to facilitate signature analysis. ⁴² In the longitudinal MRD samples and, importantly, in samples with discordant positivity compared to BM flow, APOBEC contribution was detectable, supporting the MM origin of these ctDNA reads (Figure 3C, Supporting Information S1: Table S4). However, for some discordant cases where BM MRD conversion or progression never occurred (e.g., NIH001), we cannot rule out a false plasma MRD call with certainty.

We also queried whether hyperdiploidy (HRD), a common phenomenon in MM, might skew detection rates based on the presence of SNVs that had been duplicated within chromosomal gains (Methods). At baseline plasma sampling, HRD cases (18/25, 72%) demonstrated a greater enrichment for duplicated variants in cfDNA relative to their matched BM tumor compared to non‐HRD cases. The mean change in proportion of duplicated variants was +6.5% in HRD cases versus +0.5% in non‐HRD tumors (linear model, P = 0.44). This effect is biologically consistent with the increased burden of chromosome gains in HRD disease, suggesting preferential representation of high‐copy regions in cfDNA (Supporting Information S1: Figure S4A,B). However, among cases with detectable disease at MRD and PD time points, where TF was lower (n = 33), there was no such enrichment in duplicated SNVs in HRD cases, such that the overall effect of this unique aspect of MM disease biology on cfDNA MRD detection may be small and unlikely to be contributing to biased detection in this set (Supporting Information S1: Figure S4C).

Finally, we questioned whether there might be distinct high‐confidence SNVs that might be overrepresented throughout longitudinal samples at low TF, which is biologically unlikely. Pairwise comparison of SNVs across longitudinal cfDNA samples within patients showed a modest number of shared variants between baseline and follow‐up samples, which is appropriate, given higher TF at baseline and the stochastic nature of SNV sampling. Importantly, the overlap between MRD samples was minimal, arguing against widespread technical duplication or propagation of artifact throughout serial samples, and supported by a low Jaccard similarity across longitudinal samples (Supporting Information S1: Figure S5).

Longitudinal fragmentomic profiles characterize disease states

Multiple prior analyses have suggested that tumor‐derived cfDNA fragments have shorter fragment length (145–150 bp) as compared to the canonical nucleosome‐protected distance of normal apoptotic cell‐derived cfDNA (167 bp). ⁴⁹ , ⁵⁰ , ⁵¹ We therefore queried whether fragmentomic profiles varied across disease states, in part, to determine whether fragmentation patterns might aid in the identification of true‐positive MRD calls in low‐TF states. To control for inter‐patient differences in fragmentation patterns, we analyzed longitudinal data from within‐patient samples, relative to baseline, where TF was highest. Across patients, baseline samples showed a higher proportion of fragments less than 150 bp, ⁵² compared to MRD and PD time points (Supporting Information S1: Table S6). A linear mixed‐effects model accounting for patient‐specific baseline fragmentation confirmed that the proportion of fragments under 150 bp was consistently reduced at MRD time points compared to baseline, with no strong evidence for reversion at disease progression, though the latter could be limited by the small sample size and biochemical vs clinical PD (Figure 3D). Generally, these findings indicate that the cfDNA fragmentation profiles reflect disease state, rather than the absolute tumor burden, but suggest that sample‐level fragmentation patterns alone may struggle to discern true ctDNA presence in low‐TF states.

We next addressed whether the fragmentomic profiles of each of the 11 cases with discordant MRD calls (BM−/cfDNA+) might provide further resolution for disease presence (Methods). Specifically, we compared the size of fragments supporting the primary tumor compendium to those with reference sequence at these loci. As expected, the overall low number of tumor‐supporting variants precluded statistical confirmation of enrichment for these fragment sizes. However, many BM−/cfDNA+ cases did display a non‐significant leftward shift in tumor‐supporting fragment size with visually prominent peaks at ~150bp in additional limited support of the MM origin of these cfDNA fragments (Figure 3E and Supporting Information S1: Figure S6). Altogether, support for the true tumor origin of these fragments is strengthened when added to the preceding mutational signatures and clonality analyses, and in the context of agreement with the clinicotemporal pattern of disease behavior (Figure 2C). Additionally, in cases of sequential samples from the same patient, the number of tumor‐supporting fragments with size <150 bp increased as TF increased and disease progressed (Figure 3F). Overall, within patients, fragmentomic profiles may be used to glean disease state and dynamics.

Longitudinal tracking of cfDNA tumor fraction and sustained MRD negativity

CfDNA WGS presents a logistically viable and minimally invasive approach to track TF over time. With sequential samples taken over the course of each patient's treatment, we sought to track dynamic changes in TF as they pertain to outcome. After grouping patients according to outcome, we identified that all patients with clinical/biochemical disease progression and/or MRD conversion (i.e., failure to sustain MRD negativity per BM assessment) had concomitant or, in some cases, preceding increases in TF. For each patient with more than two longitudinal cfDNA assays spanning disease progression or MRD conversion, there was a dynamic increase in TF over time, including at earlier interval time points 3 and 6 months prior to the event (Figure 4). In some cases, increasing plasma TF can be seen to precede disease resurgence in marrow owing to the ability for more frequent plasma sampling over BM biopsy. These observations suggest added value in serial peripheral assessment of MRD status to more readily inform or predict changes in disease status.

Figure 4.

Longitudinal tumor fraction dynamics. Longitudinal TF is grouped by clinical outcome. Patients with sustained MRD negativity (or stable/sustained MRD positivity) per BM assessment are shown on the left, while those with failure to sustain BM MRD negativity (MRD conversion) or with progression of disease under follow‐up are shown on the right panel. For graphical purposes, undetectable TF (i.e., 0) is set to 5 × 10⁻⁵. BM, bone marrow; MRD, minimal residual disease; TF, tumor fraction.

DISCUSSION

BM‐based assessment of MRD is well established in MM, but local sampling of spatially heterogeneous disease and the requisite for invasive procedure remain intrinsic limitations. Thus, the current paradigm may either eventually be replaced or improved upon by the addition of systemic, peripheral MRD assessment. Profiling of plasma cfDNA stands to deliver informative prognostic data on tumor genomics, as well to serve as a direct subject of residual disease assessment in low‐TF settings. While various technologies aim to leverage ctDNA, each has limitations primarily related to sensitivity. ¹⁸ , ¹⁹ , ²⁰ , ²¹ Furthermore, the capabilities of deep targeted sequencing approaches (i.e., B‐cell Receptor Sequencing) on which current MM MRD assessments are based in BM are generally limited in the context of low tumor burden and by the ceiling on available cfDNA fragments and genomic elements available in plasma. ³² While the limits of target‐seq approaches may be improved upon with disease‐specific modifications and tumor‐informed tracking, ²³ , ²⁴ still, these techniques do not capture the full scope of prognostic information that can be delivered with more comprehensive WGS. We argue that an approach encouraging tumor and plasma WGS stands to offer value in terms of richer data generation at an ever‐decreasing cost. ⁵³

We performed a study of longitudinal cfDNA‐based tracking of somatic variants identified via WGS of primary HR‐SMM from CD138‐selected BM tumors. A genome‐wide mutational integration enabled sensitive, accurate calling of residual disease with individualized detection rate LODs in the 10⁻⁴ range, corresponding to TFs as low as 6.8 × 10⁻⁵, utilizing plasma volumes of only 1 mL input. A tumor‐informed interrogation, utilizing the MRDetect framework with built‐in error suppression and de‐noising, allows for confident somatic variant calling in plasma at low coverage even when few supporting reads are present. Utilization of the approach in this HR‐SMM setting revealed high concordance with conventional flow cytometry‐based BM MRD positivity in that cfDNA WGS was able to recapitulate almost all cases of BM‐positive disease. The one notable exception was in a case with exceptionally low mutation burden, a known limitation with this approach, an infrequent occurrence even in SMM, as compared to later disease stages. ³² Conversely, peripheral assessment of residual disease revealed a substantial fraction of cases that were MRD‐negative by BM assessment but had detectable disease in the blood. Extensive quality control utilizing fragmentomic analysis, clonal dynamics, and mutational signatures analysis suggests that the majority of BM−/cfDNA+ cases are true‐positives, representing increased sensitivity for systemic disease over conventional single‐site BM‐based disease assessment. In comparing plasma cfDNA to BM flow MRD, factors such as hemodilution or compartment‐dependent disease clearance dynamics might also play a role in discordant calls, though we noted no clear temporal trend in disagreements between BM and plasma. These observations suggest that peripheral MRD assessment offers a more representative assay of systemic residual disease and can refine MRD assessment when added to traditional single‐site assays, which may be falsely negative in case of spatially patchy or predominantly extramedullary disease. Finally, the quantification of TF by this method bears both prognostic and predictive utility. First, cfDNA TF can be monitored with greater frequency, and growing values are observed to precede disease progression or BM MRD conversion, suggesting its value as a biomarker for further investigation. Second, the amount of ctDNA shedding appears to be linked to genomic complexity and is associated with poor outcomes in this cohort. This is a point of interest, considering that this same genomic complexity is also linked to proliferative capacity and egression of circulating tumor cells, suggesting that higher cell turnover is also linked to shedding of ctDNA. ⁵⁴ We note that the data set used here derives from an interventional study of HR‐SMM, where MRD is less established than in MM. However, these cases are universally classified as genomic MM (i.e., have already undergone malignant transformation) and were treated with potent triplet therapy and maintenance as in MM, giving context to their study for MRD dynamics. The performance of cfDNA WGS in this precursor disease state reinforces the potential utility in more advanced, relevant disease states, where prospective investigation will be needed.

The approach described here has multiple clinical applications and represents an incremental advantage over existing techniques for peripheral MRD assessment. However, it remains limited in delivering information on specific genomic regions of interest. For example, the relative low coverage and baseline‐tumor‐informed nature of the genome‐wide approach is not suited to inform on the emergence of acquired resistance variants (i.e., antigen escape variants to dominant immunotherapies). ⁵⁵ , ⁵⁶ For targeted assessment of acquired variants in key genomic regions, a combined approach utilizing targeted sequencing or error‐corrected high‐coverage WGS would be required, but these are still limited in low TF (i.e., MRD setting) due to the sparse representation of genomic elements of interest. ⁵³ The LODs achievable here with the cfDNA WGS approach also remain higher than conventional BM‐based flow cytometry or VDJ sequencing. However, the LODs achieved here are not a ceiling and can be improved upon by increasing depth of sequencing, which will be more feasible as associated costs continue to fall, ³² or by augmenting sequencing approaches with improved error suppression platforms. ⁵⁷ Additionally, the LODs may improve in advanced disease settings with higher mutational burdens. Our TF LOD in the 10⁻⁴–10⁻⁵ range was expected for our disease setting (HR‐SMM; median mutational burden, 3636; range 1795–7456) and WGS cfDNA coverage (×40), but the method could achieve LOD of 10⁻⁶ in settings of high tumor mutational burden (e.g., 60,000+ SNVs) and high coverage (e.g., ×120 WGS). ³⁰ We expect augmented performance of this technique in the newly diagnosed and relapsed/refractory MM disease states where mutational burden is notably higher (~5000 with ranges up to 70,000 ²⁶ , ⁴³ ), proliferative capacity and genomic complexity are enhanced (i.e., more ctDNA shedding), and MRD is a more established clinical endpoint. Additionally, though the reliance on baseline tumor WGS is a strength of the approach, it remains a limitation for those patients who do not have access or material available for baseline WGS. To this end, tumor agnostic strategies such as the development of deep‐learning classifiers for cfDNA WGS based on specific MM features (i.e., APOBEC and poly‐eta mutational signatures, somatic hypermutation of immunoglobulin loci) or the implementation of high‐specificity error‐corrected sequencing will need to be developed. ⁵³ , ⁵⁸

Altogether, with incremental improvements, WGS holds promise for being a dominant method for assessment of residual disease in low‐TF settings in MM, worthy of further prospective investigation. We observe here that tumor‐informed, genome‐wide WGS of cfDNA provides prognostic data, highly sensitive assessment of MRD, and informs on dynamic changes in tumor burden that may facilitate personalized approaches for patients with MM and its precursors.

AUTHOR CONTRIBUTIONS

Chrissy Baker: Formal analysis; data curation; writing—original draft; visualization. Elizabeth Hill: Writing—original draft; formal analysis; data curation. Dickran Kazandjian: Project administration; resources; writing—review and editing. Marios Papadimitriou: Writing—original draft; formal analysis; data curation. Michael Durante: Writing—original draft; formal analysis; data curation. Abhishek Pandey: Formal analysis; writing—review and editing. Bachisio Ziccheddu: Writing—review and editing; formal analysis. Tomas Jelinek: Writing—review and editing; formal analysis. David Coffey: Writing—review and editing; formal analysis. Brian Walker: Writing—review and editing; formal analysis. Ryan Young: Writing—review and editing; data curation. Kylee Maclachlan: Writing—review and editing; formal analysis. Neha Korde: Writing—review and editing; project administration; resources. Nickoli Parkinson: Writing—review and editing; formal analysis. Zoe R. Goldstein: Writing—review and editing; project administration. Alexi Runnels: Writing—review and editing; formal analysis; data curation. William F. Hooper: Writing—review and editing; formal analysis; data curation. Dan Landau: Writing—review and editing; formal analysis. Nicolas Robine: Writing—review and editing; formal analysis. Francesco Maura: Formal analysis; writing—original draft; data curation. Ola Landgren: Project administration; resources; writing—original draft. Benjamin Diamond: Conceptualization; supervision; data curation; formal analysis; project administration; methodology; visualization; writing—original draft.

CONFLICT OF INTEREST STATEMENT

B.D. has received honoraria from Janssen and Sanofi for ad hoc advisory boards and independent data review committee for Janssen. F.M. has received honoraria from Medidata. O.L. has received research funding from the National Institutes of Health (NIH), NCI, US Food and Drug Administration, MMRF, International Myeloma Foundation, Leukemia and Lymphoma Society, the Paula and Rodger Riney Myeloma Foundation, Perelman Family Foundation, Rising Tide Foundation, Amgen, Celgene, Janssen, Takeda, Glenmark, Seattle Genetics, and Karyopharm; received honoraria and is on advisory boards for Adaptive, Amgen, Binding Site, BMS, Celgene, Cellectis, Glenmark, Janssen, Juno, and Pfizer; and serves on independent data monitoring committees for clinical trials led by Takeda, Merck, Janssen, and Theradex. T.J. received research funding from Janssen and Sanofi, and has an honoraria and consultancy/advisory role for Bristol Myers Squibb, GlaxoSmithKline, Janssen, Pfizer, and Sanofi.

ETHIC STATEMENT

Samples and data were obtained and managed in accordance with the Declaration of Helsinki. All patients reported here provided written informed consent for the respective clinical interventional study, and investigations on human samples were approved by the National Cancer Institute institutional review board.

FUNDING

This work was supported by the Sylvester Comprehensive Cancer Center National Cancer Institute (NCI) Core Grant (P30 CA 240139). B.D. is supported by the American Cancer Society and is a K12 Scholar supported by the National Cancer Institute of the National Institutes of Health under Award Number K12CA226330. This work was supported by the OPJAK SALVAGE project (No. CZ.02.01.01/00/22_008/0004644) under the Ministry of Education, Youth and Sports of the Czech Republic.

Supporting information

Supporting File 1: hem370367‐sup‐0001‐supplemental_Figures.docx.

Supporting File 2: hem370367‐sup‐0002‐Supplemental_Tables_v2.xlsx.

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are openly available in EGAS00001007404 at https://ega-archive.org/studies/EGAS00001007404. All required code used for the WGS analyses has been uploaded to https://github.com/UM-Myeloma-Genomics

• 25 CD138+ bone marrow‐derived high‐risk smoldering multiple myeloma WGS are publicly available at EGAS00001007404.

• 87 plasma‐derived cfDNA WGS will be uploaded to the same repository upon acceptance of the study.