Access to diagnosis using liquid biopsy (ADLiB): Identifying lymphoma in a tuberculosis‐endemic setting Authors: Katherine Antel, Nina Xiong, Mwanasha Merrill, Taiyang Chen, Lauren McGrath, Clara Droessler, Maxine van der Schyff, Meng‐Xuan Wu, Dylan Cheddie, Charlotte Slome, Liam McBride, Jonathan Day, Zainab Mohamed, Nazia Fakie, Nomonde Mbatani, Charlotte Shultz, Graeme Meintjes, Svitlana Tyekucheva, Elizabeth F. Cohen, Michael Y. Tolstorukov, Estelle Verburgh, Mark A. Murakami Published: 2026 ABSTRACT AbstractIn HIV‐ and tuberculosis (TB)‐endemic regions, lymphoma diagnosis is often delayed because symptoms can overlap with TB, and access to biopsy and specialized pathology is limited. To address this, we developed and internally evaluated the Access to Diagnosis using Liquid Biopsy (ADLiB) platform—a plasma cell‐free DNA (cfDNA)‐based approach capturing molecular features of lymphoma and key pathogens. Plasma cfDNA undergoes ultra‐low‐pass whole genome sequencing and deep targeted sequencing (50,000× coverage). The combined output—including estimated tumor fraction, mutations, copy number alterations, structural variants, B‐ and T‐cell receptor clonality, and pathogen detection—is integrated into a machine‐learning classifier developed within this study to differentiate malignant from benign lymphadenopathy. In a prospective cohort of 124 South African patients with lymphadenopathy due to lymphoma (76.6%), benign causes (18.6%; TB adenitis [9.7%] and reactive [8.9%]), and disseminated solid tumors (4.8%), patients with lymphoma had higher cfDNA concentrations (mean 248.3 ng/mL), tumor fraction (mean 11.6%), and more mutations (mean 9.5), frequently in KMT2C, SOCS1, TP53, STAT3, and CREBBP. Targeted structural variants were 100% specific for lymphoma. The ADLiB machine‐learning classifier distinguished malignant from benign lymphadenopathy within this cohort with 95% sensitivity, 75% specificity, and 92% accuracy (AUC 0.89). ADLiB provides a minimally invasive and feasible diagnostic approach to distinguish malignant from benign lymphadenopathy in resource‐limited settings with high pathogen prevalence that can complicate lymphoma diagnosis. With further optimization and clinical validation, this cfDNA‐based approach has the potential to accelerate and improve lymphoma diagnosis where traditional diagnostics are constrained. INTRODUCTION In South Africa, the prevalence of HIV among adults is 17%, 1 and the incidence of tuberculosis (TB) ranks among the highest globally. 2 The convergence of these epidemics makes lymphoma diagnosis uniquely challenging. People living with HIV (PWH) have a five‐ to 10‐fold higher lymphoma risk, 3 yet diagnosis is often delayed for several reasons. One major reason is frequent misdiagnosis of lymphoma as TB, since both can present with lymphadenopathy, weight loss, fever, cytopenias, and pleural effusions. Observational studies from Africa have shown that 40%–85% of lymphoma patients are initially misdiagnosed with TB. 4 , 5 Additional barriers include delayed lymph node biopsy, overestimation of the ability of fine needle aspiration (FNA) with cytology alone to exclude lymphoma, 6 and limited hematopathology capacity—only 12% of laboratories have basic immunohistochemistry (IHC), and fewer than 5% offer cytogenetics. 7 , 8 These constraints hinder accurate and timely diagnosis, underscoring the need for streamlined diagnostic pathways for lymphoma and other cancers presenting with lymphadenopathy. Published data from South Africa highlight how diagnostic delays increase mortality in patients suspected to have lymphoma, even before a definitive diagnosis is made. Reported median intervals from first healthcare visit to lymphoma diagnosis range from 7 to 12 weeks, 6 , 9 similar to delays reported in the United Kingdom (10–12 weeks). 10 However, these medians likely underestimate the problem because many patients never reach a confirmed diagnosis. In one South African study of patients who had FNAs, 48% (10/21) of those with FNA findings suggestive of lymphoma died while awaiting a diagnostic biopsy. 9 In another report, the attrition across diagnostic steps was striking: of 45 PWH with cytology suggestive of lymphoma, only 24 (53%) returned for confirmatory biopsy. Among those who did not return by 1 year, 57% had died, 29% were untraceable, and just 14% were confirmed alive. 11 These data highlight the urgent need for diagnostic tools deployable at first presentation to enable timely linkage to care and to mitigate the extreme vulnerability of this population. Emerging approaches for detecting cancer from plasma‐derived cell‐free DNA (cfDNA) offer such potential. These “liquid biopsies” can detect tumor‐derived single‐nucleotide variants (SNVs), copy number alterations (CNAs), and structural variants (SVs) 12 , 13 , 14 ; identify clonality through B‐cell receptor (BCR) and T‐cell receptor (TCR) sequencing 15 ; and identify infectious pathogens. 16 , 17 The primary benefit of cfDNA in this context is that it provides a minimally invasive diagnostic option, as blood can easily be drawn across clinical settings. This is particularly valuable where delays or barriers to tissue biopsy exist. A prior South African study tested cfDNA for lymphoma‐associated BCR rearrangements using a PCR‐based assay, but detection was inconsistent. 18 To our knowledge, no published studies have tested cfDNA for multiple classes of lymphoma genetic variants and metagenomic features to distinguish lymphoma from TB and other cancers in patients with lymphadenopathy. To address these gaps, we previously established a rapid‐access lymph node biopsy clinic in Cape Town, providing same‐visit core biopsies. 19 In a cohort of 277 patients, 43% of whom were PWH, TB was the most frequent cause of lymphadenopathy (34%), followed by lymphoma (27%) and other cancers (17%). Building on this work, we developed the Access to Diagnosis using Liquid Biopsy (ADLiB) platform, which integrates ultra‐low‐pass whole genome sequencing (ULP‐WGS) for genome‐wide copy number analysis and tumor fraction estimation with a targeted deep sequencing panel for (1) recurrent lymphoma alterations (SNVs, CNAs, and SVs); (2) BCR/TCR clonality; and (3) pathogens relevant to HIV‐associated cancer such as TB, Epstein‐Barr virus (EBV), human herpesvirus 8 (HHV8), and human papilloma virus (HPV). We analytically validated ADLiB in retrospective specimens and evaluated its performance in a prospective South African cohort of 124 patients with lymphadenopathy, demonstrating its potential to accelerate lymphoma diagnosis in HIV‐ and TB‐endemic regions. METHODS Patient enrollment We consecutively enrolled a cohort of patients from Groote Schuur Hospital (Cape Town, South Africa) in 2022–2023 via one of two clinical pathways: (1) presentation for a lymph node biopsy without a prior diagnosis or (2) known lymphoma (ambulatory or hospitalized) before or within 1 week of starting chemotherapy for newly diagnosed or relapsed disease. Patients in the first group had a range of conditions causing lymphadenopathy, including lymphoma, TB adenitis, and solid tumors. Participants from both clinical pathways comprised a noninterventional diagnostic cohort (n = 124) on which the platform was tested. Primary tissue biopsies and/or peripheral blood plasma specimens and clinical data were obtained from individuals from this cohort who provided informed consent to University of Cape Town Human Research Ethics Committee‐approved protocols (784/2022 and 057/2013), in accordance with the Declaration of Helsinki. Detailed clinical and pathological characteristics of patients whose biospecimens were evaluated in this study are provided in Supporting Information S2: Table 1. Reference standard clinical diagnosis Concurrent tissue specimens underwent standard morphological evaluation, IHC staining, and additional diagnostic studies for lymphoma subtyping according to the 2017 WHO classification by licensed histopathologists at the National Health Laboratory Service at the University of Cape Town. Fluorescence in situ hybridization (FISH) was not used to detect SVs in all cases of lymphoma, and was performed more restrictively than in well‐resourced settings, but results were included where testing was performed. TB diagnoses were confirmed using microbiologic testing appropriate to the disease site, including Xpert MTB/RIF (GeneXpert; Cepheid) testing on sputum or blood and/or positive mycobacterial cultures. HIV status was determined using standard serologic or molecular assays performed as part of routine clinical care. Detailed diagnostic criteria and testing methodologies are provided in the Supplemental Methods. Panel design and development The targeted panel for ADLiB was designed using baits from Twist Bioscience. This panel incorporated three modules: recurrent alterations, clonality, and pathogens (Figure 1A). For the recurrent alterations module, we targeted alterations reported in published studies on lymphoma, 20 , 21 , 22 , 23 , 24 , 25 , 26 , 27 which were accessed via cBioPortal 28 , 29 , 30 (with some datasets uploaded to a private instance when not publicly available). We weighted gene selection toward lymphoma subtypes most prevalent in South Africa. 31 We selected 172 genes (all exons), 93 regions for copy number analysis, and 11 large intergenic regions for lymphoma‐associated translocation analysis (Supporting Information Methods, Supporting Information S2: Table 2). For the clonality module, we incorporated probes targeting the complete set of functional V(D)J genes of the immunoglobulin heavy (IGH) and light chain (IGL, IGK) loci and TCR loci (TRA, TRB, TRD, TRG) based on the International ImMunoGeneTics Information System database, designed using the EuroClonality‐NGS DNA capture panel 32 (Supporting Information S2: Table 3). For the pathogen module, we targeted four pathogens that either confound lymphoma diagnosis (e.g., TB) or contribute to oncogenesis in PWH (EBV, HPV, HHV8) (Supporting Information S2: Table 4). ADLiB is a novel approach to detecting lymphoma in an HIV‐ and TB‐endemic setting using plasma cfDNA. (A) The ADLiB platform combines ULP‐WGS and a targeted panel designed to detect recurrent alterations, B‐ and T‐cell clonality, and pathogens. Created in https://BioRender.com. (B) Schematic of the ADLiB bioinformatics analysis pipeline. Created in https://BioRender.com. cfDNA extraction and sequencing Blood was collected in Streck™ Cell‐Free DNA BCT tubes (Streck) and processed within 24 h. Plasma was isolated by sequential centrifugation (1600g followed by 16,000g at room temperature), aliquoted, and stored at −80°C prior to cfDNA extraction. Samples were shipped on dry ice. cfDNA was extracted from plasma, followed by library preparation and sequencing on Illumina NovaSeq platforms (Supporting Information Methods). For ULP‐WGS, libraries were sequenced using the NovaSeq S Prime flow cell, aiming for ×0.1 mean target coverage. For targeted sequencing, libraries were sequenced using the NovaSeq X 25B flow cell to achieve a mean raw coverage of ×50,000 and duplex consensus coverage of at least ×2000. Bioinformatics analysis Raw sequencing reads underwent initial quality control and adapter trimming, followed by alignment to the GRCh37 human reference genome using BWA‐MEM, 33 demultiplexing using fgbio, 34 and preprocessing using GATK 35 and Picard. 36 Tumor fraction was estimated from ULP‐WGS using ichorCNA 37 with default parameters. Further details, including a description of the tools used to call recurrent genomic alterations, are provided in Supporting Information S2: Table 5 and Supporting Information Methods. Bioinformatics analysis pipelines were built using Workflow Description Language and run on DNAnexus®, a cloud‐based analysis platform (Figure 1B). Specimens for analytical validation We defined the preliminary performance of our panel using a combination of cell lines, commercial reagents, and patient specimens with gold‐standard reference comparisons (Supporting Information Methods, Supporting Information S2: Tables 6 and 7). Patient specimens were obtained via the Dana‐Farber Cancer Institute Mikaela McDonough Lymphoma Biobank with written consent under Dana‐Farber/Harvard Cancer Center Institutional Review Board‐approved protocols #01‐206 and #17‐561 and sequenced under protocol #21‐363. Sensitivity and specificity were measured against established reference tests meeting US Clinical Laboratory Improvement Amendments standards or the South African National Accreditation System according to ISO 15 189. The limit of detection (LOD) for SNVs was defined using a commercially available cfDNA pan‐cancer reference standard (#104 549; Twist Bioscience). Dilutions of 5%, 2%, 1%, 0.5%, 0.25%, 0.1%, and 0% variant allele fraction were used to determine the sensitivity at which 86 reference variants overlapping the ADLiB targeted genes were detected (Supporting Information S2: Table 8). Exploratory survival analysis Exploratory survival analyses were performed for the two largest lymphoma subgroups (diffuse large B‐cell lymphoma [DLBCL] and classical Hodgkin lymphoma [cHL]). Two‐year overall survival was evaluated in relation to plasma cfDNA concentration and estimated tumor fraction using nonparametric comparisons between surviving and deceased patients. Patients were treated according to local protocol with first‐line treatment for DLBCL with rituximab, cyclophosphamide, vincristine, and prednisone, and HL with adriamycin, bleomycin, vinblastine, and dacarbazine. ADLiB classifier Receiver operating characteristic (ROC) analysis We evaluated the ability of individual genomic features and a composite genomic score to distinguish malignant (lymphoma or other cancer) from benign lymphadenopathy. For each feature, ROC curves were generated by plotting the true positive rate (sensitivity) against the false positive rate (1−specificity) across a range of thresholds. The area under the ROC curve (AUC) was used to quantify diagnostic accuracy. ROC analyses were performed in Python (scikit‐learn). To identify optimal thresholds, we applied Youden's J statistic (sensitivity−(1−specificity)), which maximizes the combined sensitivity and specificity. Formulation of the composite genomic score A composite genomic score was generated by weighting classes of genomic alterations according to their biological plausibility and relative specificity for malignancy. This score was defined ahead of the analyses. SVs (5 points per event) received the highest weight because of their strong association with malignant lymphoid disease. Oncogenic and likely oncogenic SNVs were assigned 3 and 2 points, respectively, reflecting their established role in lymphomagenesis. Variants of uncertain or likely benign significance were assigned 1 point. CNAs were quantified across cytobands (−2 to +2 based on log2 copy ratio thresholds: ≥1 = 2, 0.1 to <1 = 1, >−0.1 to <0.1 = 0, >−1 to −0.1 = −1, ≤−1 = −2) to capture the broader genomic instability of cancer. These values were summed with the absolute tumor fraction to generate the composite genomic score (Supporting Information Methods). ROC analysis was performed to evaluate diagnostic performance, with the optimal threshold determined by Youden's J index. Posttest probabilities at this threshold were calculated using both the observed prevalence in our dataset (81%) and a clinically relevant prevalence of 44% from a prior study of a similar regional cohort. 19 Machine‐learning algorithm We trained a random forest classifier to differentiate cancer from benign samples using the molecular features detected by our panel (Supporting Information Methods). Samples with molecular results from the diagnostic cohort (n = 116) were used as input for the random forest classifier, with 80% used for training the model and 20% set aside for testing. Model performance was assessed using accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and AUC analysis. RESULTS Defining the preliminary performance of the targeted panel Recurrent alteration module Using the Twist cfDNA reference standard, we demonstrated high concordance between the reference variants and the variants called by ADLiB, with 94% sensitivity and 90% precision at ≥5% variant allele fraction (Figure 2A). Variant allele fraction measurements from the targeted panel correlated strongly with those expected from the dilution series (Supporting Information S1: Figure 1), with r = 0.75 (P < 0.01, Pearson correlation coefficient) and a mean squared error of 2 × 10−⁴. Analytical validation of ADLiB reveals high concordance with benchmarked specimens. (A) Sensitivity versus precision for detecting variants from the Twist cfDNA reference standard at various dilutions using LoFreq. (B) SVs detected in four patient specimens with diagnostic FISH and three lymphoma cell lines with known translocations. (C) Venn diagram showing overlap of CNA identified by ichorCNA and ADLiB within ADLiB baited regions for nine patient specimens. (D) Comparison of CNA identified by ichorCNA and ADLiB within ADLiB baited regions, showing only the genomic regions where either ADLiB or ichorCNA detected a copy number gain or loss. The scale is defined as follows: 2, log2 copy ratio > = 1; 1, log2 copy ratio > = 0.1 and <1; 0, log2 copy ratio >−0.1 and <0.1; −1, log2 copy ratio < = −0.1 and >−1; and −2, log2 copy ratio < = −1. (E) The top ten clones detected in four lymphoma cell lines with previously reported immunoglobulin gene rearrangements. Only the V gene associated with the clone is shown. Bars highlighted in red indicate clones that were previously reported. (F) Pathogen read counts for samples with gold standard pathogen measures, normalized per 100 million reads in the sample, and log‐transformed. Samples with zero reads have an undefined log and are not shown. The following log‐normalized read thresholds were chosen, and the confusion matrices generated accordingly: EBV, 7; HHV8, 5; HPV, 4; TB, 1. To detect SVs, we assessed four patient specimens with translocations previously identified by diagnostic FISH, in addition to three lymphoma cell lines with published translocations—Raji, Mino, and WSU‐NHL (Figure 2B). The targeted panel successfully detected the IGH::CCND1 fusion in three out of three mantle cell lymphoma (MCL) samples, the IGH::BCL2 fusion in three out of three follicular lymphoma (FL) samples, and the IGH::MYC fusion in one out of one Burkitt lymphoma sample, demonstrating 100% sensitivity for clinically relevant rearrangements within this cohort. For copy number analysis, we compared CNA between ADLiB and ichorCNA in the 93 genomic regions (focal and arm‐level) targeted by the panel for nine patient specimens (Figure 2C,D). We found a 43% concordance rate across samples and regions where a CNA (gain or loss) was called by either ADLiB or ichorCNA (24 concordant of 56 total calls). Twenty‐two CNA calls were specific to ADLiB, while 10 were specific to ichorCNA. This discrepancy could be explained by bait affinity and density for regions targeted by ADLiB, in contrast to unbiased sequencing without baits in ULP‐WGS used for ichorCNA, leading to deeper coverage in baited regions. Clonality module We evaluated BCR and TCR clonality using genomic DNA (gDNA) from four short tandem repeat (STR)‐authenticated lymphoma cell lines and paired tumor gDNA and plasma cfDNA from five independent patients. The most abundant clonal sequences detected by ADLiB matched the clonal sequence reported in reference datasets in two of the four cell lines and the most abundant clonal sequence of the same immunoglobulin locus in the remaining two (Figure 2E). However, in primary patient samples, dominant rearrangements in plasma cfDNA detected by ADLiB generally mapped to the TRA and TRB loci and did not recapitulate the dominant rearrangements detected in tumor gDNA by LymphoTrack (Supporting Information S1: Figure 2, Supporting Information Results). The estimated tumor fractions in these plasma specimens were modest (range, 3.2%–10.4%), raising the possibility that sensitivity for lymphoma‐specific BCR/TCR sequences was limited by the abundance of those from nonmalignant lymphocytes. Given these constraints, we focused on diversity‐based metrics rather than on detecting dominant rearrangements within the diagnostic cohort. Pathogen detection module To avoid misclassifying lymphoma as infection—a more consequential diagnostic error in this setting—we prioritized high specificity when setting thresholds for pathogen detection. This approach minimized false‐positive pathogen calls, reducing the risk of incorrectly attributing lymphadenopathy to infectious causes when lymphoma was present. Thresholds were defined to achieve >95% specificity, and sensitivity was reported at this level. To avoid overfitting, we applied five‐fold cross‐validation and evaluated performance on withheld test sets. Across folds, the average sensitivity and specificity exceeded 80% and 95%, respectively, for all pathogens (Supporting Information S1: Figure 3). To detect TB, we analyzed plasma from 12 HIV‐positive patients with confirmed disseminated TB (all blood culture positive) and nine negative controls. A log‐normalized read threshold of 1 achieved 100% sensitivity and specificity (Figure 2F). EBV detection was benchmarked against the Cobas EBV test (Roche Diagnostics) on plasma samples, with a log‐normalized read count threshold of 7, yielding 82% sensitivity and 100% specificity. For HHV8, validated in six HIV‐positive patients with Kaposi's sarcoma confirmed by LANA‐positive IHC, a log‐normalized read threshold of 5 achieved 100% sensitivity and specificity. For HPV16 and HPV18, tested in ten cervical cancer patients, a log‐normalized read threshold of 4 provided 90% sensitivity and 100% specificity. For EBV, we used cases tested as negative by the Cobas assay for negative controls; for TB, HHV8, and HPV16/18, we used tumor DNA from lymphoma specimens at Dana‐Farber Cancer Institute (DFCI) for negative controls. Clinical and demographic characteristics of the diagnostic cohort We prospectively enrolled 124 patients with lymphadenopathy, with a median age of 44 years (IQR: 30–58) (Table 1). The population represented the expected demographics of patients accessing public health care in Cape Town, South Africa. Using self‐identified ethnicity, most patients were mixed race (50%) or Black African (44%), and a minority were White (6%). HIV‐positive patients comprised 34% of the cohort. The median age in the cohort was 45 years, with HIV‐positive patients having a median age of 39 years and HIV‐negative patients having a median age of 51 years (P = 0.001). HIV‐positive patients had a low median CD4 count (218 cells/mm3), although 50% had a suppressed viral load below the detectable limit and 88% were on antiretroviral therapy at the time of enrollment. Table 1Baseline characteristics of participants.CharacteristicHIV– (n = 82)HIV+ (n = 42)P valueOverall (n = 124)Age, years (median [IQR])50.5 [31.0–64.0]38.5 [29.3–44.8]0.00144.5 [30.0–58.3]Male sex, n (%)46 (56.1)27 (64.3)0.49473 (58.9)Race, n (%)<0.001Black African17 (20.7)37 (88.1)54 (43.5)Mixed race58 (70.7)5 (11.9)63 (50.8)White7 (8.5)0 (0.0)7 (5.6)Diagnosis, n (%)0.252B‐cell lymphoma58 (70.7)31 (73.8)89 (71.8)Aggressive subtype48 (82.8)30 (96.7)78 (87.6)Indolent subtype10 (17.2)1 (3.3)11 (12.4)T‐cell lymphoma6 (7.0)0 (0.0)6 (4.8)Non‐lymphoma18 (22.0)11 (26.2)29 (23.4)Benign (other)7 (38.9)4 (36.4)11 (8.9)TB lymphadenitis6 (33.3)6 (54.5)12 (9.7)Other cancer5 (27.8)1 (9.1)6 (4.8)Lymphoma stage, n (% of lymphoma cases)0.835Stage I–II17 (27.0)7 (22.6)24 (19.4)Stage III–IV46 (73.0)24 (77.4)70 (74.5)B symptoms present, n (% of lymphoma cases)35 (35.6)20 (47.6)0.61055 (57.3)CD4 count, cells/mm³ (median [IQR])–218 [97–379]–218 [97–379]HIV viral load, copies/mL, n (%)––Lower than detectable limit (<50)–19 (52.8)19 (52.8)50–250–5 (13.9) 5 (13.9) >250–12 (33.3)12 (33.3)On ART (HAART), n (%)–37 (88.1)–37 (88.1)TB positive (any site), n (%)9 (11.0)12 (28.6)0.02621 (16.9)ECOG performance status, n (%)0.4380–153 (79.1)25 (73.5)78 (77.2)2–414 (20.9)9 (26.5)23 (22.8)Enrollment status, n (%)0.579Known lymphoma49 (59.8)28 (66.7)77 (62.0)Undiagnosed at enrollment33 (40.2)14 (33.3)47 (38.0)Abbreviations: ART, antiretroviral therapy; CD4, cluster of differentiation 4; ECOG, Eastern Cooperative Oncology Group performance status; HAART, highly active antiretroviral therapy; HIV, human immunodeficiency virus; IQR, interquartile range; TB, tuberculosis.John Wiley & Sons, Ltd. Baseline characteristics of participants. 5 (13.9) Abbreviations: ART, antiretroviral therapy; CD4, cluster of differentiation 4; ECOG, Eastern Cooperative Oncology Group performance status; HAART, highly active antiretroviral therapy; HIV, human immunodeficiency virus; IQR, interquartile range; TB, tuberculosis. Most patients were enrolled in the outpatient setting, reflected in the lower European Cooperative Oncology Group Performance Status scale 38 (ECOG) scores in the population (77% ECOG 0 or 1). We accrued slightly more patients from the lymphoma clinic, where participants had a known lymphoma diagnosis (n = 77, 62%), than from the lymph node biopsy clinic, where patients were enrolled prior to diagnosis (n = 47, 38%). This enriched for lymphoma cases, resulting in 95 lymphoma cases in the cohort (77%). Other causes of lymphadenopathy included TB adenitis (n = 12, 10%), other benign etiologies like reactive lymphadenopathy (n = 11, 9%) and disseminated other cancers (n = 6, 5%). As expected, most lymphoma patients had B‐cell lymphoma (93.7%, 89/95), the majority of which were aggressive subtypes (87.6%, 78/89), and 6.3% (6/95) had T‐cell lymphoma. A high proportion of lymphoma cases were advanced‐stage disease (stage III/IV: 73.7%, 70/95), with similar stages across HIV‐positive and HIV‐negative groups (P = 0.83). In total, 77.9% (74/95) of lymphoma patients were newly diagnosed (treatment‐naïve), 14.7% (14/95) had initiated treatment with plasma collected within 1 week of treatment initiation, and 7.4% (7/95) had relapsed disease with plasma collected prior to retreatment. Nine lymphoma patients had concurrent active TB at the time of lymphoma diagnosis, with TB diagnosed at sites other than lymph nodes (pulmonary n = 5; abdominal n = 2; disseminated n = 2). All were diagnosed with TB prior to lymphoma and were receiving antituberculous therapy at the time of sampling. No cases demonstrated co‐localization of TB and lymphoma within the same lymph node biopsy. Of these nine cases, six were HIV‐positive, and three were HIV‐negative. The patients had classic HL (n = 4), DLBCL (n = 3), plasmablastic lymphoma (n = 1), and Burkitt lymphoma (n = 1) (Supporting Information S2: Table 1). cfDNA concentration and tumor fraction estimation We used the panel to molecularly profile cfDNA samples from our prospectively enrolled diagnostic cohort. Ninety‐four percent of participants (n = 116) successfully underwent sequencing and bioinformatic analysis, with 8 excluded (Figure 3A). Plasma cfDNA concentrations from patients with lymphoma (mean: 248.3 ng/mL) were nominally higher than those with solid tumors (mean: 179.7 ng/mL; P = 0.42, Wilcoxon rank‐sum test) and TB (mean: 119.7 ng/mL; P = 0.06, Wilcoxon rank‐sum test), and significantly higher than those with other benign causes of lymphadenopathy (mean: 33.6 ng/mL; P < 0.01, Wilcoxon rank‐sum test) (Supporting Information S1: Figure 4). ADLiB detects clinically relevant molecular features and pathogens in the diagnostic cohort. (A) Consort diagram for the diagnostic cohort showing samples excluded from each step. (B) ichorCNA estimated tumor fraction by disease type for the diagnostic cohort. “Other lymphoma” refers to lymphoma subtypes with fewer than two cases in the cohort, namely small lymphocytic lymphoma (n = 1), mantle cell lymphoma (n = 1), and mediastinal gray zone lymphoma (n = 1). The other abbreviations are as follows: BL, Burkitt lymphoma; DLBCL, diffuse large B‐cell lymphoma; FL, follicular lymphoma; HGBL, high‐grade B‐cell lymphoma; HL, Hodgkin lymphoma; PBL, plasmablastic lymphoma; PMBCL, primary mediastinal large B‐cell lymphoma; PTCL, peripheral T‐cell lymphoma; TB, tuberculosis. The dashed red line indicates 3% tumor fraction (validated lower LOD for ichorCNA). (C) Average number of SNVs detected per sample, by sample type. Variants were annotated for oncogenicity using OncoKB. (D) OncoPrint of SNVs detected in the diagnostic cohort. Samples with no detected variants were excluded from the diagram. (E) Circos plot of SVs detected in the diagnostic cohort. (F) CNA calls for Hodgkin lymphoma (HL), diffuse large B‐cell lymphoma (DLBCL), plasmablastic lymphoma (PBL), and Burkitt lymphoma (BL) cases in the diagnostic cohort. Log2 copy ratio thresholds of 0.1 and −0.1 were defined as gains and losses, respectively. (G) Samples with predicted pathogen status as determined by the log‐normalized read thresholds set by the validation specimens. (H) B‐ and T‐cell receptor clonality metrics for the diagnostic cohort. The Gini coefficient was used as a metric of clonal diversity, ranging from 0 (equal abundance of each sequence) to 1 (high clonality toward one sequence). All patients underwent ULP‐WGS. The estimated tumor fraction in patients with lymphoma (mean: 11.6%) was nominally higher than in those with solid tumors (mean: 8.8%; P = 0.50, Wilcoxon rank‐sum test) and significantly higher than the low values in those with TB (mean: 1.4%; P < 0.01, Wilcoxon rank‐sum test) and benign causes of lymphadenopathy (mean: 1.3%; P < 0.01, Wilcoxon rank‐sum test) (Figure 3B). Most lymphoma subtypes (except peripheral T‐cell lymphoma) had a median tumor fraction greater than 3%, the validated LOD of ichorCNA, 37 while patients with TB and benign cases had a median tumor fraction less than 3%. Given that ichorCNA has limited specificity below its 3% LOD, the low‐level copy number instability detected in TB and benign samples is most consistent with background noise and false‐positive tumor fraction calls rather than true circulating tumor DNA. Eighteen patients (17%) with lymphoma had a tumor fraction lower than 3%. This low tumor fraction did not correlate with the stage of disease, timing of plasma collection, or lymphoma subtype (Supporting Information S2: Table 9). Importantly, these lymphoma cases met quantitative cfDNA input requirements for ADLiB and generated adequate sequencing metrics for variant calling, despite their reduced tumor fraction. Exploratory analyses were performed to evaluate associations between cfDNA‐derived metrics and overall survival at 2 years within the two largest lymphoma subgroups, DLBCL (n = 23) and cHL (n = 23). In DLBCL, higher plasma cfDNA concentration was significantly associated with mortality (P = 0.033), while tumor fraction was not associated with survival (P = 0.87). In cHL, a similar trend was observed for cfDNA concentration (P = 0.086), whereas tumor fraction again showed no significant association with survival (P = 0.35). Any potential prognostic implications, while intriguing, should be interpreted cautiously given the limited cohort size and follow‐up duration (Supporting Information S1: Figure 5). Molecular findings in the diagnostic cohort The average number of variants identified per patient was significantly higher in those with lymphoma (mean: 9.5; P < 0.01, Wilcoxon rank‐sum test) and solid tumors (mean: 6.3; P < 0.01, Wilcoxon rank‐sum test) than in those with benign causes of lymphadenopathy, including TB (mean 1.2; Figure 3C). Most detected variants had unknown oncogenicity per the publicly available genetic variant annotation tool OncoKB, 39 , 40 and no known oncogenic mutations were detected in any benign cases. Overall, the mutational pattern reflected the recurrent alterations previously reported in the lymphoma subtypes that were analyzed (Figure 3D). The 50 most frequently mutated genes included tumor suppressors (TP53, ARID1A, BCL7A), JAK/STAT pathway members (SOCS1, STAT3, STAT6), targets of aberrant somatic hypermutation (BCL2, MYC, PIM1, SGK1), and chromatin modifiers (KMT2C, TET2, CREBBP). Mutations were also seen in known lymphoma drivers such as NRAS, KRAS, and CDKN2A. We additionally provide mutation types (e.g., missense, nonsense) and indicate whether variants have previously been reported in COSMIC, a comprehensive catalog of somatic mutations in human cancer, in Supporting Information S1: Figure 6. We identified several clinically relevant SVs, including IGH::BCL2 in six of seven follicular lymphoma cases and four of 26 DLBCL cases; all four DLBCL cases were diagnosed as de novo germinal center B‐cell (GCB)‐subtype DLBCL without known antecedent FL. We identified CCND1::IGH in the one mantle cell lymphoma case, and MYC::IGH in three of five Burkitt lymphoma cases (Figure 3E). We also detected BCL6::CIITA in one of 28 Hodgkin lymphoma cases and BCL6::IGH in two of 26 DLBCL cases. Among five lymphoma patients with diagnostic FISH, ADLiB correctly identified 6/6 known SVs, including two IGH::BCL2 events and four MYC breakpoints where the partner gene was not identified. The CNA analysis revealed expected alterations, including 2p and 9p gains in Hodgkin lymphoma, 1q and 11q gains in plasmablastic lymphoma, 7q and 12q gains in Burkitt lymphoma, and recurrent DLBCL alterations such as gains in 1q, 6p, 7p, and 7q and losses in 6q and 17p (Figure 3F). In Hodgkin lymphoma, altered genes included XPO1 and CD274 gains. In DLBCL, we detected gains in NOTCH2, CCND3, CD274, and BCL2 and losses in TNFAIP3, PRDM1, PTEN, CD70, and TMEM30A. Patients with benign conditions, including TB, had relatively flat copy number profiles (Supporting Information S1: Figure 7). We next assessed the detection of TB by the cfDNA assay under three case definitions (Figure 3G and Supporting Information S2: Table 10). When TB was defined as positivity at any site, including pulmonary TB with GeneXpert MTB/RIF Ultra positivity on sputum, the assay detected TB with 40.0% (8/20) sensitivity and 85.4% (82/96) specificity. The PPV in this case was 36.4% and the NPV 87.2%. When TB was defined only as lymph node disease (GeneXpert MTB/RIF Ultra test positive or culture positive on lymph node tissue), the assay detected TB with 66.7% (8/12) sensitivity, and when TB was defined as disseminated TB, requiring either blood culture or urinary lipoarabinomannan (LAM) positivity, the assay detected TB with 50.0% (2/4) sensitivity. We only report specificity for the broadest definition of true positive TB, since the negative cases by site cannot be considered true/false negatives. Among the proven lymphoma cases with confirmed TB already on treatment (n = 9), none had detectable TB cfDNA reads in plasma. ADLiB detected EBV only in patients with lymphoma, including four with Hodgkin lymphoma, three with Burkitt lymphoma, four with DLBCL, and six with plasmablastic lymphoma. When benchmarked against EBV PCR (n = 28 with results for both tests), ADLiB detected EBV with 82% sensitivity and 100% specificity, with a PPV of 100% and an NPV of 78.6% (Supporting Information S2: Table 11). Against EBV‐encoded ribonucleic acid (EBER) in situ hybridization‐ or latent membrane protein 1 (LMP1)‐positive tumor tissue (n = 69), sensitivity increased to 90.9% but specificity reduced to 91.5%, yielding a PPV of 83.3% and an NPV of 95.6%. Detection of HHV8 and HPV was exploratory in this version of the assay because the diagnostic cohort did not include patients with Kaposi sarcoma or cervical cancer. We detected HHV8 in four patients, all of whom were plausible HHV8‐positive cases, having advanced HIV and lymphoma. HHV8 was routinely tested by latency‐associated nuclear antigen (LANA) IHC in only one of these patients, who was positive. We detected HPV in four patients (two male and two female). We were unable to consistently detect dominant clones in cfDNA or distinguish tumor‐derived clones from peripheral blood mononuclear cell‐derived clones without prior tumor DNA sequencing. Instead, BCR and TCR clonality metrics were used to assess clonal skewing (Figure 3H). The Gini coefficient, ranging from 0 (equal abundance of all sequences) to 1 (dominance of a single sequence), was used as a diversity metric. Patients with B‐cell lymphoma had nominally higher BCR clonality (mean: 0.17) than those with T‐cell lymphoma (mean: 0.08; P = 0.62), solid tumors (mean: 0; P = 0.54), TB (mean: 0.05; P = 0.06), and benign lymphadenopathy (mean: 0; P = 0.07). All pairwise comparisons were performed using the Wilcoxon rank‐sum test. TCR clonal diversity was modestly elevated in both T‐cell and B‐cell lymphoma cases compared to benign, but showed substantial overlap with other diagnostic groups and limited discriminatory value (Supporting Information S2: Table 12). ADLiB classifier We distinguished benign from malignant causes of lymphadenopathy using two methods—a directed method that calculated a composite genomic score that can be prespecified based on biological knowledge of lymphoma and preliminary data, and an agnostic machine‐learning method. Before these combined methods, we performed ROC analysis for each feature to quantify the discriminatory performance between benign cases and malignant cases (lymphoma and other cancers combined) (Table 2 and Figure 4A, input features and data shown in Supporting Information S2: Tables 13 and 14). The mutation score (derived from SNVs) demonstrated the strongest diagnostic accuracy (AUC = 0.89). At an optimal threshold of 4.0, this score detected lymphoma or other cancers with 74.5% sensitivity and 90.9% specificity. The tumor fraction also performed well (AUC = 0.86), with a threshold of 3.24% yielding 79.8% sensitivity and 86.4% specificity. Copy number alterations performed more modestly (AUC = 0.70), achieving 51.1% sensitivity and 90.9% specificity at a threshold of 3.0. BCR clonality, TCR clonality, and the SV score were less able to discriminate between benign and malignant lymphadenopathy (AUCs 0.59–0.62). Thresholds for BCR clonality and the SV score reached 100% specificity, but these were offset by very low sensitivities (24.5% and 18.1%, respectively). TCR clonality performed poorly overall, with 51.1% sensitivity and 68.2% specificity. Taken together, these results indicate that mutation scores and tumor fraction are the most robust individual predictors of malignancy. Table 2Performance metrics of ROC analysis.FeatureAUCOptimal thresholdSensitivitySpecificityMutation score0.8853.500.7450.909Tumor fraction0.8610.0320.7980.864CNA0.7052.500.5110.909EBV0.6650.500.3301.000BCR clonality0.6220.0620.2451.000SV score0.6060.500.2131.000TCR clonality0.5870.0240.5110.682TB*0.337N/A1.0000.000*The apparently extreme coordinates (e.g., 100% sensitivity/0% specificity) reflect ROC boundary artifacts rather than meaningful diagnostic performance, see Supporting Information Methods for handling of binary variables.Abbreviations: AUC, area under the curve; BCR, B‐cell receptor; CNA, copy number alterations; EBV, Epstein‐Barr virus; ROC, receiver operating characteristics; SV, structural variant; TB, tuberculosis; TCR, T‐cell receptor.John Wiley & Sons, Ltd. Performance metrics of ROC analysis. *The apparently extreme coordinates (e.g., 100% sensitivity/0% specificity) reflect ROC boundary artifacts rather than meaningful diagnostic performance, see Supporting Information Methods for handling of binary variables. Abbreviations: AUC, area under the curve; BCR, B‐cell receptor; CNA, copy number alterations; EBV, Epstein‐Barr virus; ROC, receiver operating characteristics; SV, structural variant; TB, tuberculosis; TCR, T‐cell receptor. ADLiB classifier discriminates cancer from benign causes of lymphadenopathy. (A) ROC curves for individual ADLiB features distinguishing cancers (lymphoma and other cancers) from benign conditions. (B) Composite genomic scores across benign, lymphoma, and other cancer cases. Lymphoma samples demonstrated significantly higher scores compared with benign and other cancers, consistent with the accumulation of lymphoma‐associated mutations and genomic alterations. Boxplots show medians and interquartile ranges, with individual patient values overlaid; P‐values are from pairwise Wilcoxon rank‐sum tests. (C) ROC curve for the composite genomic score integrating multiple features, demonstrating overall diagnostic accuracy with the optimal threshold (Youden's J statistic) marked on the curve. (D) Schematic of the machine‐learning classification workflow. The input data consisted of the 116 samples for which molecular results were reported in the diagnostic cohort. Created in https://BioRender.com. (E) ROC curve and confusion matrix for distinguishing cancer from benign samples. (F) Accuracy by disease type for the machine‐learning classifier, including the number of correctly classified samples out of the total number of samples. BL, Burkitt lymphoma; DLBCL, diffuse large B‐cell lymphoma; FL, follicular lymphoma; HL, Hodgkin lymphoma; PBL, plasmablastic lymphoma; PTCL, peripheral T‐cell lymphoma; SLL, small cell lymphoma; TB, tuberculosis. Next, we calculated the composite genomic score, incorporating mutations, SVs, CNAs, and tumor fraction. The median composite scores differed significantly across diagnostic categories (Kruskal–Wallis P < 0.0001, Figure 4B). Benign samples had a lower median score (3.4, IQR 2.0–5.2) than lymphoma (23, IQR 8.9–37.8) and other cancers (18, IQR 16.3–30.7) (Figure 4B). Pairwise comparisons showed that both lymphoma (P = 5.3 × 10−8) and other cancers (P = 3.6 × 10−⁴) had significantly higher composite genomic scores than benign cases, while the scores did not differ between lymphoma and other cancers (P = 0.96). ROC analysis of the composite genomic score yielded an AUC of 0.90, indicating strong discriminatory ability. The optimal threshold by Youden's index was 7.8, which was rounded to 8 for clinical applicability (Figure 4C). At this cutoff, the assay identified 74 of 94 cancer cases, with a sensitivity of 78.7% and specificity of 90.9%. The PPV was high at 97.4%, while the NPV was more modest at 50.0%. The composite genomic score performed well within HIV‐positive and HIV‐negative subgroups (AUC 0.79 and 0.93, respectively), with performance in the HIV‐positive group likely limited by a smaller sample size (n = 39; Supporting Information S1: Figure 8). Next, we assessed the ability of ADLiB features to distinguish lymphoma from other cancers by restricting the analysis to malignant cases (n = 94; 88 lymphoma, 6 other cancers, Supporting Information S2: Table 15). Most features could only modestly distinguish lymphoma from other cancers, with AUCs between 0.55 and 0.66. However, three features detected lymphoma with 100% specificity, despite low sensitivity. These were EBV positivity, BCR clonality, and the presence of an SV. The high specificity of BCR clonality for B‐cell lymphomas likely reflects its role as a virtually defining feature of mature B‐cell neoplasms in contrast to the immunological diversity of those with nonmalignant reactive processes and healthy individuals. The specificity of SVs reflects the rarity of these genetic aberrations in healthy individuals. The mutation score had an AUC of 0.59, with a relatively high threshold (10.5) required to optimize specificity (67%), consistent with the panel's prioritization of lymphoma‐associated mutations. The composite genomic score performed poorly in this restricted analysis (AUC 0.51). To further improve diagnostic accuracy by incorporating all ADLiB outputs, we next tested machine‐learning classifiers that integrated pathogen data and BCR/TCR clonality alongside genomic variation. A random forest classifier was trained to distinguish cancers (lymphoma and solid tumors) from benign conditions, using the same inputs as the ROC analysis together with BCR/TCR clonality, EBV positivity, and TB status, with the dataset randomly split into training and testing sets (Figure 4D). This model achieved an AUC of 0.89, with 95% sensitivity, 75% specificity, and an overall accuracy of 92%, which in this cohort corresponds to an NPV of 77% for malignancy (lymphoma or solid tumor) vs benign adenopathy (Figure 4E). Two cases were misclassified (Figure 4F): one benign case was incorrectly labeled as cancer (tumor fraction 4.6%, no evidence of malignancy after 18 months of follow‐up), and one relapsed Hodgkin lymphoma case (tumor fraction 3.9%) was misclassified as benign because of low mutation, SV, and CNA encodings (2, 0, and 0, respectively). In this cohort, where the prevalence of cancer was 101/124 (82%), the classifier's positive likelihood ratio of 3.8 corresponds to a post‐test probability of 94.4%. But because our cohort was enriched for lymphoma through enrollment of known lymphoma cases, this post‐test probability is less than would be expected in a similar regional cohort with undiagnosed lymphadenopathy, where the cancer prevalence is 44%, 19 which would correspond to a post‐test probability of 75%. DISCUSSION Here, we present the development, optimization, and preliminary analytical validation of the ADLiB platform. These data provide proof‐of‐concept that malignant lymphoma can be detected in TB‐ and HIV‐endemic settings through integrated molecular interrogation of a single standard phlebotomy specimen, without paired tumor tissue or matched germline material. This represents a step toward addressing critical delays in lymphoma diagnosis in HIV‐ and TB‐endemic regions. ADLiB distinguished cancer from benign causes of lymphadenopathy by estimating tumor fraction, detecting recurrent genetic alterations, assessing BCR and TCR clonality, and identifying pathogens through a targeted cfDNA sequencing panel. When evaluated against the reference standard of clinical–pathologic diagnosis in our prospective cohort, ADLiB distinguished malignant from benign lymphadenopathy with 95% sensitivity. The ability to detect lymphoma‐defining alterations such as IGH::BCL2 and IGH::MYC in a minimally invasive manner could enable accurate lymphoma diagnosis and subtyping in settings where FISH and other molecular assays are unavailable. The ADLiB workflow—from sample collection to cloud‐based analysis—was technically feasible, with cfDNA successfully extracted and analyzed in 94% of cases. To ensure real‐world impact, ADLiB would require local or regional implementation to optimize turnaround time and batching efficiency, although current sequencing and reagent costs remain a major barrier to widespread deployment. In South Africa, results could be returned within 14 days if blood is collected during the initial healthcare encounter and the sequencing infrastructure supports weekly batch runs. Unlike traditional diagnostic workflows, genetic alterations were not reviewed by a molecular pathologist. Although this would likely have improved diagnostic accuracy, trained molecular pathologists are scarce and therefore unrealistic in low‐resource settings. High‐throughput sequencing platforms were installed for COVID‐19 surveillance and could be repurposed for lymphoma diagnosis. Given current cost constraints, ADLiB is best viewed as a proof‐of‐concept and assay development platform rather than a near‐term replacement for existing diagnostic pathways, although regional batching—potentially across sub‐Saharan Africa—may help reduce expenses. At present, ADLiB is best positioned as a potential triage tool. Lymphoma subtyping to guide treatment still relies on morphological analysis, IHC, and, in some cases, molecular profiling. A larger study cohort could help refine machine‐learning approaches to classify subtypes based on cfDNA data; however, the precision will be challenged by genetic overlap among entities. Additional cfDNA modalities, such as methylome profiling 41 or computational inference of gene expression 42 may be required for subtyping. Nonetheless, the platform's ability to detect SVs and biologically relevant mutations represents a meaningful advance for resource‐limited settings, enabling diagnostic subclassification and identification of therapeutic targets without invasive tissue sampling. Accordingly, the current study does not represent clinical validation of a finalized diagnostic algorithm but rather assay development and internal evaluation intended to define feasibility and analytical performance. Tumor fraction estimation is central to ADLiB performance and influences several downstream analyses. Tumor fraction was calculated using ichorCNA. 37 This software, which uses a hidden Markov model, has been validated for WGS data with coverage as low as 0.1× including from lymphoma studies. 43 , 44 ichorCNA outputs two potential tumor fraction estimates; while other studies often select the estimate most consistent with known pathology and expected CNA, we deliberately remained blinded and accepted only the top‐ranked solution. This conservative approach—intended to reflect real‐world diagnostic conditions—may have reduced sensitivity. The tumor fraction may also have been affected by low tumor burden, subtype heterogeneity, benign cases, and cfDNA concentration variability. Future work could refine detection thresholds for low‐tumor fraction cases, optimize sequencing for minimal tumor burden, or integrate additional biomarkers to improve estimation of the non‐invasive tumor fraction. The ADLiB platform was designed primarily to distinguish malignant from benign lymphadenopathy in TB‐endemic settings, rather than to serve as a standalone diagnostic for TB. The TB component demonstrated high specificity and negative predictive value, supporting its utility in ruling out TB as a mimic of lymphoma. Positive calls, however, should still be interpreted alongside standard microbiologic confirmation. We expect the sensitivity to be limited by the low abundance of TB cfDNA levels in plasma. Improving sensitivity may require targeted amplification 45 or presequencing enrichment 46 rather than adjustment of calling thresholds. Fragmentomic methods are unlikely to improve the distinction between cancer and TB, as both yield short cfDNA fragments, and our ×0.1 WGS depth was too shallow to support such approaches. 47 Several key limitations should be acknowledged. The first relates to the detection of dominant BCR/TCR rearranged sequences. While expected dominant sequences were identified in gDNA from cell lines, dominant sequences were not consistently detected in cfDNA, where circulating tumor DNA was highly diluted. In agreement with these findings, inconsistent PCR‐based assay performance has been reported in similar populations. 18 We did not have access to matching tumor and cfDNA from our prospective cohort that we could have used to test the ability of ADLiB to detect a known dominant sequence at a low sequence abundance. Although this analysis would have helped us define our panel capability, it would have had limited practical utility, given the challenges of obtaining and molecularly profiling tumor tissue that this blood‐based assay is specifically designed to circumvent. Importantly, receptor diversity metrics improved classifier sensitivity, with oligoclonality as indicated by low clonal diversity correlating with lymphoma. Future iterations of the assay will focus on optimization of BCR/TCR rearrangement detection, which could further enhance its ability to distinguish lymphoma from other causes of lymphadenopathy. Copy number calling was further limited by the absence of a gold‐standard reference, reflecting the limited use of clinical FISH testing in this cohort. In future studies, prioritizing cases with known clinically significant copy number alterations, such as del(17p), would enable more robust benchmarking. Although the assay showed high concordance with translocations identified by FISH, we were not always able to identify the partner gene in cases involving MYC rearrangements. This likely reflects both the diversity of MYC translocation partners and limitations inherent to targeted breakpoint coverage. A final limitation was that, although the ADLiB classifier showed promise in distinguishing cancer from benign causes, our analysis was constrained by the small sample size, particularly when distinguishing lymphoma from other cancers (n = 6). In addition, our machine‐learning model was trained on mutually exclusive diagnoses as the final cause for lymphadenopathy, whereas in reality—especially in HIV‐ and TB‐endemic settings—lymphoma and TB often coexist. Larger studies enriched for untreated and extrapulmonary TB will be required to further define the performance of plasma‐based TB detection. With a larger validation cohort, alternative modeling strategies could be explored, including threshold optimization and predictive approaches that prioritize sensitivity without disproportionately compromising specificity. In conclusion, this pilot study has established that the ADLiB platform can detect clinically relevant molecular features from a single plasma specimen, without requiring a paired germline specimen, thereby offering a minimally invasive approach that could accelerate lymphoma diagnosis in HIV‐ and TB‐endemic regions. The platform is compatible with existing next‐generation sequencing infrastructure in Africa and could be deployed with minimal additional technical investment. While cost and technical hurdles remain, the potential to reduce diagnostic delays and identify clinically relevant genomic alterations supports further technical optimization, clinical validation, and implementation‐focused studies. AUTHOR CONTRIBUTIONS Katherine Antel: Conceptualization; investigation; funding acquisition; writing—original draft; methodology; validation; visualization; writing—review and editing; formal analysis; project administration; supervision; resources. Nina Xiong: Investigation; writing—original draft; methodology; validation; visualization; writing—review and editing; formal analysis; conceptualization. Mwanasha Merrill: Investigation; writing—review and editing; writing—original draft; funding acquisition; project administration; conceptualization. Taiyang Chen: Methodology; formal analysis; visualization. Lauren McGrath: Project administration; methodology; investigation. Clara Droessler: Investigation. Maxine van der Schyff: Investigation; data curation. Meng‐Xuan Wu: Data curation. Dylan Cheddie: Data curation; investigation. Charlotte Slome: Project administration; data curation. Liam McBride: Formal analysis; investigation. Jonathan Day: Project administration; data curation. Zainab Mohamed: Data curation. Nazia Fakie: Data curation. Nomonde Mbatani: Data curation. Charlotte Shultz: Data curation. Graeme Meintjes: Data curation. Svitlana Tyekucheva: Formal analysis. Elizabeth F. Cohen: Methodology; formal analysis. Michael Y. Tolstorukov: Methodology; formal analysis. Estelle Verburgh: Conceptualization; funding acquisition; supervision; resources; project administration. Mark A. Murakami: Conceptualization; investigation; funding acquisition; writing—original draft; methodology; validation; visualization; writing—review and editing; supervision; resources. CONFLICT OF INTEREST STATEMENT The authors declare no conflicts of interest. FUNDING We are grateful for support from our funding sources, including the ASH Global Research Award (KA and MMw), AACR Beginning Investigator Grant for Catalytic Research (KA), PMC FLAMES F.L.A.I.R. Award (KA), and Harold and Virginia Lash Gift Fund at Dana‐Farber Cancer Institute (KA, MM, MAM). Research reported in this publication was supported by the Fogarty International Center and National Heart, Lung & Blood Institute of the National Institutes of Health under Award Number D43 TW010345 (EV); the Fogarty International Center and the South African Medical Research Council under Award Number D43 TW010345 (EV) and by the Cancer Association of South Africa (CANSA) (EV). Any opinion, findings, and conclusions or recommendations expressed in this material are those of the author(s), and CANSA does not accept liability in regard thereto. Supporting information supplemental_data_clean. supplemental_tables.