Diffuse large B-cell lymphoma (DLBCL) is the most common histological subtype of non-Hodgkin lymphoma, accounting for more than 30% of all NHL cases. The disease is clinically and molecularly heterogeneous, and up to 10-15% of DLBCL tumors carry a MYC rearrangement (MYC R) at chromosome band 8q24. MYC rearrangements result in deregulated MYC expression because transcriptional enhancer elements from immunoglobulin (IG) genes are juxtaposed to the MYC oncogene, driving overexpression that plays a central role in disease pathogenesis. Patients with MYC R, designated MYC+, consistently show shorter survival compared to those without (MYC-), and clinical evidence suggests they may benefit from more aggressive treatment regimens beyond standard R-CHOP.
The detection problem: Conventional methods for detecting MYC R include G-banded karyotype chromosome analysis and fluorescence in situ hybridization (FISH). However, both approaches have documented limitations. Cryptic rearrangements and wide variation in MYC breakpoints mean these translocations go undetected in an estimated 10-15% of cases. FISH probes, including both the dual-color dual-fusion probe set and the MYC break-apart probe, can miss rearrangements because breakpoints may lie more than 350-645 kb downstream from the 5' or 3' ends of MYC itself. Studies have confirmed that up to 11% of MYC breakpoints fall outside the probe detection windows, and FISH misclassification rates of up to 10% have been reported.
The translocation spectrum: Among MYC R cases, the most common translocation is t(8;14)(q24;q32), a MYC-IGH fusion occurring in approximately 70% of MYC+ tumors. The t(8;22)(q24;q11) MYC-IGL fusion accounts for 22%, and t(2;8)(p12;q24) MYC-IGK fusion accounts for roughly 8%. An additional 13% of MYC R cases involve non-IG gene partners (MYC/non-IG), including translocations such as t(3;8)(q27;q24) and t(8;9)(q24;p13). These non-IG and cryptic rearrangements carry expression signatures similar to double-hit DLBCL, making their identification clinically important even though conventional methods frequently miss them.
Next-generation sequencing (NGS) has been introduced into clinical settings for structural variant detection, but it faces the same breakpoint-variation limitation as FISH when MYC partner sequences lie far outside standard probe regions. Long-distance PCR approaches exist but are not routinely deployed because primer design cannot reliably cover all breakpoint possibilities in newly diagnosed cases. The diagnostic gap motivates the development of AI-based prediction tools that can flag MYC R from data already routinely collected in the cytogenetics laboratory.
The study employed a two-cohort architecture to develop and then independently validate the AI algorithm. Cohort 1 was drawn from the Mitelman Database of Chromosome Aberrations and Gene Fusions, a publicly accessible repository of published cytogenetic data from cancer cases. The database was searched for DLBCL cases diagnosed between 1983 and 2019 (accessed May 2020). After filtering for cases with documented karyotypes and 8q24 breakpoint status, a total of 474 DLBCL cases were identified: 108 MYC+ and 366 MYC-. These formed the primary training and testing pool for the AI algorithm. The class imbalance (roughly 1:3 ratio of MYC+ to MYC-) reflects the true prevalence in DLBCL and represents a realistic challenge for classifier development.
Karyotype processing pipeline: Raw karyotype strings from cohort 1 were first parsed using CytoGPS, a web-enabled software tool that converts ISCN (International System for Human Cytogenomic Nomenclature) karyotype strings into structured loss/gain/fusion annotations. Each case was then manually curated to verify accuracy. A Fisher Exact two-tailed test, chi-square test, and Bonferroni-adjusted p-values were applied to identify cytogenetic aberrations that differed significantly between MYC+ and MYC- groups. Evolutionary trajectory analysis was performed using the TRONCO package (R environment), which models the temporal ordering of aberrations to identify early versus late events in tumor progression.
Cohort 2 - institutional validation: Cohort 2 comprised 59 institutional cases from UT Southwestern Medical Center, all with both G-banded karyotype and FISH (MYC/IGH dual-fusion and MYC break-apart probes) performed on fresh clinical specimens. This cohort included 31 MYC+ cases (13 bone marrow, 7 lymph node, 9 other sites) and 28 MYC- cases (8 bone marrow, 11 lymph node, 8 other sites, with 3 cases having no available tumor site data). Specimens were processed into tissue culture within 4 hours of collection, and metaphase spreads were prepared from 24-hour short-term cultures using standard G-banding protocols. IRB approval was obtained for use of these institutional cases.
The study design deliberately kept cohort 2 completely separate from algorithm development - no cohort 2 data were used for training. This separation is critical for evaluating generalizability, since models tested only on held-out splits from the same source dataset can still overfit to population-specific patterns in that source.
The core methodological contribution of this study is the construction of a two-component feature set feeding the AI algorithm: recurrent cytogenetic aberrations (RCAs) and a derived genetic progression score (GPS). RCAs are specific chromosomal gains and losses that appear at statistically significant frequencies in the dataset. Manual curation of the 474 cohort 1 karyotypes identified 22 RCAs across the full dataset (Figure 4 in the original paper). Statistical testing confirmed that gain of chromosome 1p34 (p = 0.003) and gain of 1q14 (p = 0.0008) were significantly more prevalent in MYC+ tumors, while chromosome losses were significantly more prevalent in MYC- tumors (141 MYC+ loss events vs. 611 MYC- loss events, p < 0.001).
Genetic Progression Score (GPS) construction: GPS was calculated using the Rtreemix package (R environment), which derives a weighted score from the number and timing of accumulated genetic aberrations in each tumor. The weighting scheme is temporally informed: late-occurring events in the tumor's evolutionary trajectory receive a higher weighted value than early events. For example, 1p36 loss (a late event) receives a higher weight than monosomy 13 (an earlier event). The GPS is thus a composite measure capturing both the cumulative mutational burden and the clonal progression stage of each tumor. Due to computational memory constraints, only 14 of the 22 RCAs could be included in GPS calculation: gains of 1p34, 1q14, 5, 7, 12, and 18, plus losses of 1p36, 2q, 4p, 13q, 17p, 19p13, and t(14;18).
GPS results: The GPS clearly separated MYC+ from MYC- tumors, with average GPS values of 1.27 versus 0.68 respectively (p < 0.0001). This statistically significant difference confirms that MYC+ tumors carry a heavier burden of late-stage cytogenetic alterations, consistent with the hypothesis that MYC R acts as a key driver producing a more aggressive evolutionary trajectory. Importantly, GPS emerged across multiple AI classifiers (see later cards) as the single most important predictive feature for MYC status.
Neural network architecture: The combined RCA + GPS feature set was fed into the primary AI algorithm, a neural network (NNET) with 15 input nodes and one output node. The NNET package in R was used for construction. A 10-fold cross-validation procedure was applied during training. Seventy percent of cohort 1 cases (332 cases) were used to train the model, and the remaining 30% (142 cases) formed the internal test set. A receiver operating characteristic (ROC) curve determined the optimal decision threshold, with MYC status classified as positive or negative based on the neural network output relative to this threshold.
On the internal test set (30% of cohort 1, 142 cases), the ROC curve for the RCA + GPS NNET algorithm showed an area under the curve (AUC) of 93.8%, with a sensitivity of 91.4% and specificity of 93.8% for predicting MYC R. These performance metrics represent strong discrimination for a binary classification task on cytogenetic data, particularly given the relatively small training size and the inherent heterogeneity of published karyotype data aggregated across multiple decades and institutions. The ROC-derived threshold of 0.89 was selected as the discriminator between MYC+ and MYC- classification.
Institutional validation (cohort 2): Applying the trained NNET algorithm to the 59 UT Southwestern cases, the model correctly classified 55 of 59 cases at the 0.89 threshold. Critically, all 31 MYC+ cases were correctly identified, including 2 cases with cryptic MYC rearrangements (cases 21 and 54) that represented complex chromosomal rearrangements not straightforwardly visible by standard karyotyping. Four MYC- cases were misclassified as MYC+ (cases 44, 49, 51, and 55), and subsequent analysis revealed these 4 cases had elevated GPS values, suggesting unusually high cytogenetic complexity for MYC- tumors - which may represent a biologically interesting subgroup warranting further study.
Cohort 2 performance metrics: Sensitivity 100%, specificity 87%, positive predictive value (PPV) 92%, and negative predictive value (NPV) 100%. The 100% sensitivity is of particular clinical relevance because failing to detect a MYC R carries direct treatment consequences - those patients may be undertreated if not identified. The NPV of 100% means the algorithm did not produce any false negatives: every MYC+ case was correctly flagged. The 4 false positives (cases incorrectly labeled MYC+) would result in additional FISH confirmation testing rather than missed diagnoses, representing a clinically acceptable error pattern.
The algorithm's ability to correctly identify both cryptic MYC R cases in cohort 2 is a particularly notable finding. These cases had complex karyotypes where the MYC rearrangement was not immediately apparent from standard karyotype inspection, yet the elevated GPS and aberration pattern led the neural network to correctly assign them as MYC+. This demonstrates that the AI is capturing cytogenetic complexity patterns that serve as surrogate markers for MYC R even when the direct chromosomal evidence is ambiguous.
To assess whether the NNET approach was optimal relative to other machine learning architectures, the authors employed six additional classifiers from the open-source Microsoft Azure Machine Learning Studio platform. The classifiers tested were: Gradient Boosting (GBoost), MaxAbsScaler with Light GBM, Support Vector Machine (SVM), Random Forest Tree, SparseNormalizer with K-nearest centroid neural network (KCNN), and Standard Scan Wrapper Logistic Regression. All six were applied to the same RCA + GPS feature set used by the NNET model.
Comparative performance: Five of the six Azure classifiers outperformed the NNET model on the internal test data, with GBoost and Light GBM showing particularly strong AUC values. However, a critical reason for selecting NNET as the primary reported model is its transparency: the NNET package in R produces output probabilities for each case relative to the ROC-derived threshold, making it possible to explicitly classify individual cases as MYC+ or MYC- with an interpretable confidence score. The Azure "black box" classifiers produce predictions without readily accessible per-case probability outputs that can be mapped to a clinical decision threshold, limiting their direct clinical applicability.
GPS as the dominant feature: Across all six Azure classifiers, GPS consistently ranked as the single most important predictor of MYC status. This finding has important implications: it suggests that a relatively simple summary statistic (the weighted accumulation of cytogenetic events over the tumor's evolutionary trajectory) captures most of the predictive signal, and that individual specific aberrations are secondary contributors. This also means that even simpler models using GPS alone might retain substantial predictive power, though the combination with RCAs in the full model clearly improves performance.
The convergence of multiple independent classifier architectures on the same result - that GPS-based cytogenetic complexity predicts MYC R - provides internal consistency evidence that the signal being captured is biologically real rather than an artifact of the NNET architecture or its specific hyperparameters. This robustness across classifiers strengthens confidence in the overall analytical framework.
Beyond algorithmic prediction, the study used TRONCO-based evolutionary trajectory analysis to characterize the temporal ordering of cytogenetic events in MYC+ versus MYC- tumors. In MYC+ tumors, MYC R itself emerged as the sole driver alteration - the primary event around which other chromosomal changes accumulate as passenger events. This finding contrasts with MYC- tumors, which show a more distributed pattern of driver and passenger aberrations. The identification of MYC R as the single driver in MYC+ DLBCL is consistent with experimental models showing that MYC overexpression is sufficient to initiate lymphomagenesis in B cells when combined with constitutive BCL2 expression.
Chromosome 1 gains: Partial gain of chromosome 1 (specifically gains at 1p34 and 1q14) was significantly enriched in MYC+ tumors. Chromosome 1q gains are among the most commonly reported secondary aberrations in aggressive B-cell lymphomas and are associated with copy number increases in genes including MCL1 and AKT3 that promote cell survival. The 1p34 gain association is less well characterized but may reflect amplification of genes involved in DNA damage response and ribosome biogenesis, consistent with the proliferative phenotype driven by MYC overexpression.
Gene expression signature correlates: Evolution patterns revealed specific RCAs in MYC+ tumors that correlate with established gene expression profiles. Trisomy of chromosome 3 was associated with altered expression of MYD88 and FOXP1, both of which are recurrently mutated in the activated B-cell (ABC) subtype of DLBCL. Loss at 7p15 correlated with CARD11 expression changes, and loss at 6p21 correlated with altered PIM1 expression. Loss at 17q23 correlated with CD79B. These specific gene-expression connections provide a mechanistic framework connecting the cytogenetic aberration pattern to the molecular phenotype of MYC+ tumors, explaining why chromosomal complexity as a whole (captured in GPS) is such a strong predictor of MYC R status.
Late-appearing events in MYC+ tumor progression included gains of chromosomes 13, 15, 16, and 20, and losses of chromosomes 17 and 19. Gains at 19p13 and 19q13 also appeared late. These late events may represent secondary adaptations that enhance the fitness of MYC-driven clones and are reflected in the higher GPS weighting assigned to late-occurring aberrations, providing mechanistic justification for the GPS construction approach.
Clinical outcome data were available for 44 of the 59 cohort 2 patients. Kaplan-Meier survival analysis with log-rank testing confirmed that MYC R was associated with significantly shorter survival in this institutional cohort (Figure 10), consistent with the extensive published literature on this topic. In prior studies, MYC+ DLBCL patients treated with R-CHOP showed 5-year overall survival of approximately 44% compared to 67% for MYC- patients (p = 0.001). In pediatric populations, event-free survival was approximately six-fold lower in MYC+ cases. Even within germinal center (GCB) DLBCL, which typically carries a favorable prognosis, the presence of MYC R negates the positive outcome benefit.
Treatment implications: The clinical significance of reliably detecting MYC R extends directly to first-line treatment selection. Standard R-CHOP (rituximab, cyclophosphamide, doxorubicin, vincristine, prednisolone) is inadequate for many MYC+ patients. Dose-adjusted EPOCH-R (etoposide, prednisone, vincristine, cyclophosphamide, doxorubicin, rituximab) has shown more promising outcomes in prospective phase 2 studies specifically enrolling MYC R-positive patients. A multicenter HOVON phase 2 trial combining R-CHOP with lenalidomide also showed positive outcomes in MYC+ patients. The implication is clear: identifying MYC R before initiating therapy enables treatment intensification decisions that may improve survival.
Classification of cryptic and non-IG cases: To specifically evaluate the algorithm's performance on the most diagnostically challenging cases, the authors constructed an expanded test set by mixing 5 cryptic MYC+ cases (cases 60-64 from cohort 1) and 9 MYC/non-IG cases (cases 65-73, covering inv(8)(p21q24), t(4;8), t(8;18), t(3;8), t(8;9), t(8;16)) with the cohort 2 cases. The NNET correctly assigned all of these cases to their expected MYC category. This result is particularly important because cryptic and non-IG rearrangements are the cases most likely to be missed by current FISH approaches, representing the exact clinical scenario where AI augmentation would add the most value.
The authors note that 4 MYC- cases misclassified as MYC+ by the algorithm (cases 44, 49, 51, 55) had elevated GPS values, raising the possibility that these tumors - though technically MYC- by current chromosome and FISH criteria - may harbor undetected cryptic MYC alterations or share the high-complexity cytogenetic phenotype associated with MYC-driven biology. Prospective studies with comprehensive genomic profiling of such "GPS-high MYC-" cases would be of significant interest.
Dataset size and retrospective design: The authors candidly acknowledge that this is a proof-of-concept retrospective study. The institutional validation cohort of 59 cases is sufficient to demonstrate feasibility but is too small to draw definitive conclusions about generalizability across different patient populations, cytogenetics laboratory protocols, and FISH probe sets. Cohort 1, while containing 474 cases drawn from 36 years of published data, reflects the sampling biases inherent in the published cytogenetics literature, which overrepresents complex and unusual karyotypes. The authors call for larger datasets from major centers or multi-center collaborations to robustly validate performance.
MYC- case limitations: Because the study was retrospective, it was not possible to perform comprehensive dual-color dual-fusion and break-apart FISH on all MYC- cases to definitively rule out cryptic MYC R in metaphase cells. This means some MYC- cases in the training data may have been misclassified at the source, introducing label noise that could partially explain algorithm errors. The 4 false-positive cases with elevated GPS may represent this phenomemon, but confirming this would require genome-wide sequencing of residual material.
Simple karyotype cases: The authors acknowledge that tumors presenting with very simple karyotypes - for example, a MYC R with only one or two additional aberrations - may not generate sufficient RCA and GPS signal for the AI to make a confident prediction. This is a genuine limitation because simple karyotypes do occur, particularly in some MYC/IG cases. However, the authors note that DLBCL and large B-cell lymphomas with MYC R typically present with complex karyotypes, limiting the clinical frequency of this scenario.
Future directions: The paper calls for development of web-portal or standalone software tools that automate GPS calculation and AI prediction from standard karyotype input, making the algorithm accessible to clinical cytogenetics laboratories without bioinformatics infrastructure. Prospective studies are needed to evaluate whether algorithm-guided screening changes clinical decision-making and patient outcomes. Extension to larger non-IG and cryptic MYC R case series would also test the robustness of the current results on the most diagnostically challenging cases. Finally, integration with NGS-based karyotyping pipelines could enable GPS computation from sequencing data, expanding the applicable patient population beyond those with successful metaphase preparations.