Not Missing the Notch: Detection Challenges of Juxtamembrane NOTCH1 Variant Detection in T‐Cell Acute Lymphoblastic Leukemia Authors: Francisco Beas, Stephan Hutter, Ferran Balbastre‐Úbeda, Sandra Novoa‐Jáuregui, Maria Gabarrós‐Subirà, Eulalia Genescà, Petri Pölönen, Torsten Haferlach, Andrés Jerez Published: 2026 ABSTRACT ABSTRACTBackgroundActivating mutations in NOTCH1 are frequent in T‐cell acute lymphoblastic leukemia (T‐ALL) and, in the absence of alterations in RAS or PTEN, are associated with favorable prognosis. Besides classical heterodimerization and PEST domain mutations, juxtamembrane internal tandem duplications (JME‐ITDs) represent a third class of activating variants. Their intermediate size and sequence composition can challenge short‐read next‐generation sequencing and variant calling algorithms, leading to underdetection in routine diagnostics.MethodsAn index case of newly diagnosed T‐ALL harboring a NOTCH1 juxtamembrane insertion that was discordantly detected by variant callers during routine targeted NGS (Hematology OncoKitDx v2) prompted further analysis. The variant was evaluated using VarDict and VarScan2, followed by manual inspection with IGV and orthogonal confirmation by Sanger sequencing. To further assess variability in the detection of JME‐ITDs, six additional BAM files from patients with previously confirmed variants were reanalyzed using VarDict, VarScan2, Mutect2, and Pindel.ResultsIn the index case, VarDict detected a 51‐bp NOTCH1 JME‐ITD (c.5168‐2_5216dup) that was missed by VarScan2 and confirmed by Sanger sequencing. Reanalysis of the seven cases showed that Mutect2 and VarDict detected all JME‐ITDs (100%), VarScan2 identified only three (42.8%), and Pindel detected six mutations (85.7%). These results reflect differences in algorithmic strategies where haplotype‐based callers incorporate soft‐clipped and split reads, whereas pileup‐based discard them.ConclusionDetection of NOTCH1 JME‐ITDs depends strongly on variant‐calling strategy. Combining haplotype‐based callers with split‐read structural variant tools may reduce false negatives and improve detection of clinically relevant insertions in diagnostic NGS pipelines.Keywords: juxtamembrane mutation, NOTCH1, T‐lymphoblastic leukemia, variant detection The presence of NOTCH1 activating mutations in patients with T‐cell acute lymphoblastic leukemia (T‐ALL), in the absence of RAS/PTEN alterations, is associated with favorable prognosis and may allow selected patients who achieve minimal residual disease negativity after induction to spare allogeneic stem cell transplantation [1, 2]. In addition to well‐characterized heterodimerization and PEST domain mutations [3], a third class of NOTCH1 activating mutations—juxtamembrane internal tandem duplications (JME‐ITD, 33–180 base‐pairs)—have been reported [3, 4, 5]. However, their frequency might be underestimated in routine diagnostic testing given their intermediate size and complex sequence composition, features that challenge short‐read next‐generation sequencing (NGS) and variant calling algorithms. Similar limitations have been previously described for other ITD variants, such as CALR, representing a major technical challenge in hematological malignancies [6, 7, 8, 9]. We encountered this limitation when sequencing a bone marrow sample from a 33‐year‐old man recently diagnosed with T‐ALL who underwent routine targeted NGS (Hematology OncoKitDx V2) for mutational profiling. Initial analysis with VarDict revealed the presence of a 51‐bp juxtamembrane insertion in NOTCH1 (NOTCH1 c.5168‐2_5216dup, p.Val1739_Ala1740insGlyGluThrValGluProProProProAlaGlnLeuHisPheMetTyrVal), similar as the other JME‐ITD described in literature. To cross‐validate this finding, the sample was re‐analyzed using VarScan2, which failed to detect the insertion. To further investigate this discrepancy, manual inspection with the Integrative Genomics Viewer (IGV, v2.19.4) was performed, revealing multiple forward and reverse reads containing the insertion with consistent soft‐clipped bases at the breakpoints. The variant was subsequently confirmed by Sanger sequencing (Figure 1). No concurrent RAS or PTEN alterations were detected. (A) Sanger sequencing of NOTCH1 (NM_017617.5):C.5168‐2_5216dup (B) Reads containing the NOTCH1 variant in IGV. Given the implications of this finding and to assess the performance of different callers in detecting these variants, six additional BAM files from patients with confirmed NOTCH1 JME‐ITDs were obtained from the Munich Leukemia Laboratory. All datasets were subsequently processed using VarDict (Perl implementation, release 2019.06.04) [10], MuTect2 (GATK v4.3) [11], and VarScan2 (v2.4.6) [12], three tools commonly implemented in routine diagnostic pipelines, as well as Pindel (v0.2.5b9) [13], included as a structural variant comparator optimized for complex insertions (Table 1). TABLE 1 NOTCH1 JME variants (left panel) tested for detection by different callers (right panel) in this study.Tumor burden (% by flow cytometry)NOTCH1 JME variantsVariant callersPosition (hg19)NOTCH1 inserted variant sequence (NM 017617.3)Size (bp)SNV callers (detection Y/N, VAF %)Structural caller (Pindel)Mutect2VarDictVarScan2Detection (Y/N)BreakpointSVLENSplit reads Sample 0 (UPN01) 769:139396891ACGTACATGAAGTGCAGCTGCGCCGGCGGGGGCGGCTCCACGGTCTCACCT51Y (44%)Y (47%)NY139,396,8915148 Sample 1 (UPN02) 949:139396874CAAAGGCGGCCGCCGCCACGTACATGAAGTGCAGCTGCGCCGGCGGGGGCGGCTCCACGGTCTCACCGCCCG72Y (26%)Y (11%)NY139,396,87472419 Sample 2 (UPN03) 859:139397633TCTGCACGGCCTCGATCTTGTAGGGGATGTTGAGGCTGCCCAGCGAGGCGAGCGCTCCCAGGAATGCGGCCACGG75Y (16%)Y (14%)NY139,397,63375541 Sample 3 (UPN04) 779:139396913CCGGCGGGGGCGGCTCCACGGCCCTTGCAC30Y (38%)Y (26%)Y (9%)Y139,396,913305051 Sample 4 (UPN05) 969:139397636GCACGGCCTCGATCTTGTAGGGGATGTTGAGGCTGCCCAGCGAGGCGAGCGCTCCCA57Y (34%)Y (38%)Y (2%)Y139,397,636573035 Sample 5 (UPN06) NR9:139396887CGCCACGTACATGAAGTGCAGCTGCGCCGGCGGGGGCGGCTCCACGGCCTG51Y (34%)Y* (1.7%)Y (1.5%)Y139,396,887515074 Sample 6 (UPN07) 239:139396900CGAGGCTTCGGTGGTGGGAAGCCATCCGCAATATTA36Y* (8%)Y (11%)NNNANANAAbbreviations: Bp, base pair; JME, juxtamembrane; N, no; NA, not applicable; NR, not reported; SNV, single nucleotide variant; SVLEN, structural variant length; VAF, variant allele frequency; Y, yes.*Variants identified at the same locus but that slightly differed by 1–2 nucleotides upstream of the common inserted sequence. NOTCH1 JME variants (left panel) tested for detection by different callers (right panel) in this study. Sample 0 (UPN01) Sample 1 (UPN02) Sample 2 (UPN03) Sample 3 (UPN04) Sample 4 (UPN05) Sample 5 (UPN06) Sample 6 (UPN07) Abbreviations: Bp, base pair; JME, juxtamembrane; N, no; NA, not applicable; NR, not reported; SNV, single nucleotide variant; SVLEN, structural variant length; VAF, variant allele frequency; Y, yes. Variants identified at the same locus but that slightly differed by 1–2 nucleotides upstream of the common inserted sequence. Both Mutect2 and VarDict correctly identified all seven insertions (100%). However, for each caller, one reported variant differed by 1–2 nucleotides upstream of the common inserted sequence (Table 1). Quantitative metrics (Tables S1 and S2) were otherwise largely concordant between both callers, except in one sample (UPN6) in which VarDict reported a substantially lower variant allele frequency (VAF) than Mutect2. In contrast, VarScan2 identified only three (42.8%) insertions. Moreover, examination of the associated metrics revealed a substantial underestimation of the VAF in the detected samples. A finding consistent with an overall reduced sensitivity for ITD detection (Table 1, Table S3). Finally, Pindel detected all insertions except the last one, which was correctly identified but did not pass the filtering thresholds due to low supporting split‐read count (Table 1, Table S4). To better understand the reasons underlying these differences, we next examined the algorithmic characteristics of each variant caller. When complex insertions occur, reads spanning their breakpoints often only partially align to the reference genome, leaving unaligned bases at their ends (soft‐clipping) or multiple supplementary alignments (split reads). These alignment features remain preserved in the BAM file but are interpreted differently by each tool. Callers such as Mutect2 and VarDict, which implement local realignment or haplotype‐based assembly, actively integrate soft‐clipped and split reads to reconstruct the underlying sequence, increasing sensitivity for complex indels. In contrast, VarScan2 or other strictly pileup‐based approaches discard them, only using fully aligned reads to minimize false positives at the expense of missing relevant events. Pindel, on the other hand, adopts a different strategy, searching for pairs of reads with split alignments to identify breakpoint junctions. When comparing these algorithms, Pindel complements haplotype assembly or local alignment callers rather than replacing them. While Mutect2 and VarDict provide better performance for small to moderate indels that remain largely contained within individual reads, Pindel excels in detecting medium to large‐sized insertions that disrupt read anchoring and generate split‐read signals. This complementary behavior is particularly relevant for NOTCH1 juxtamembrane insertions, which often yield both soft‐clipped and split reads. In this setting, the integration of both approaches minimizes the risk of false negatives in the clinical sequencing context. Importantly, our observations are consistent with previous findings reported by Pölönen et al. [14], who analyzed a large cohort of 1300 T‐ALL patients and identified four cases with NOTCH1 JME variants. In their study, all four events were detected using Pindel, supporting the necessity for split‐read approaches. In contrast, the performance of conventional somatic callers (Strelka, Mutect2, Varscan2 and SomaticSniper) was more variable, with haplotype‐based approaches appearing to have a better overall performance. We acknowledge that while a formal limit‐of‐detection study using dilution series would have been appropriate, DNA availability precluded it. However, our cohort included variants with VAFs as low as 8%, which were reliably detected by Mutect2, VarDict, and Pindel but missed by VarScan2. Together with our previous observations, this suggests that algorithmic design, rather than allele frequency alone, is a major determinant of sensitivity for JME‐ITDs in routine pipelines. Based on our results, we highlight that diagnostic pipelines analyzing genes prone to internal tandem duplications should integrate: (i) a caller implementing haplotype assembly or local realignment (e.g., Mutect2 or VarDict) and (ii) a structural‐variant‐oriented caller leveraging split‐read evidence (e.g., Pindel). In cases of discordance between algorithms, results should be resolved through orthogonal validation such as Sanger sequencing. This combined strategy minimizes false negatives and provides internal cross‐validation for variants with prognostic or therapeutic relevance. Author Contributions Francisco Beas, Stephan Hutter and Andrés Jerez designed the study, collected, assembled, analyzed, interpreted the data; performed statistical analyses and wrote the manuscript. Ferran Balbastre‐Úbeda, Petri Pölönen, Sandra Novoa‐Jáuregui, Maria Gabarrós‐Subirà, Petri Pölönen, and Eulalia Genescà collected, assembled and interpreted data, and all authors revised the manuscript and approved the final version. Conflicts of Interest The authors declare no conflicts of interest. Supporting information Table S1: Mutect2 metrics for each variant. Abbreviations: AD, allele depth (alternative read); DP, depth of coverage; VAF, variant allele frequency. * denotes variants identified at the same locus but that slightly differed by 1–2 nucleotides upstream of the common inserted sequence. Table S2:. VarDict metrics for each variant. Abbreviations: AD, allele depth (alternative read); DP, depth of coverage; VAF, variant allele frequency. * denotes variants identified at the same locus but that slightly differed by 1–2 nucleotides upstream of the common inserted sequence. Table S3:. VarScan2 metrics for each variant. AD, allele depth (alternative read); DP, depth of coverage; VAF, variant allele frequency; NA, not applicable. Table S4:. Pindel metrics for each variant. SVLEN, structural variant length; NA, not applicable. Data Availability Statement The data that support the findings of this study are available from the corresponding author upon reasonable request. References Associated Data Supplementary Materials Table S1: Mutect2 metrics for each variant. Abbreviations: AD, allele depth (alternative read); DP, depth of coverage; VAF, variant allele frequency. * denotes variants identified at the same locus but that slightly differed by 1–2 nucleotides upstream of the common inserted sequence. Table S2:. VarDict metrics for each variant. Abbreviations: AD, allele depth (alternative read); DP, depth of coverage; VAF, variant allele frequency. * denotes variants identified at the same locus but that slightly differed by 1–2 nucleotides upstream of the common inserted sequence. Table S3:. VarScan2 metrics for each variant. AD, allele depth (alternative read); DP, depth of coverage; VAF, variant allele frequency; NA, not applicable. Table S4:. Pindel metrics for each variant. SVLEN, structural variant length; NA, not applicable. Data Availability Statement The data that support the findings of this study are available from the corresponding author upon reasonable request.