Mantle cell lymphoma (MCL) is a mature B-cell non-Hodgkin lymphoma that represents approximately 7% of all adult NHL cases, with an incidence of 4 to 8 cases per million people per year. It predominantly affects white men, with a median age at diagnosis of 65 years, and its incidence is rising in Western countries. MCL carries a historically poor prognosis, with a median overall survival of just 3 to 5 years, and most patients are not cured even with contemporary therapeutic regimens including immunochemotherapy and autologous stem cell transplantation.
Molecular biology: The hallmark genetic event in classic MCL is the t(11;14)(q13;q32) translocation between the CCND1 gene and the IGH locus, leading to overexpression of cyclin D1, which drives uncontrolled G1/S cell cycle progression. The immunophenotype is typically CD19+, CD20+, CD5+, SOX11+, and cyclin D1+. A leukemic variant that is SOX11-negative follows a clinically indolent course, representing an important prognostic exception. Beyond CCND1, relevant pathogenic pathways include DNA damage response (ATM, TP53), chromatin modification (NSD2, KMT2C, KMT2D), apoptosis regulation (BCL2, BIRC3), and NOTCH, NF-kB, and PI3K/AKT signaling.
Current prognostic tools: The International Prognostic Index (IPI), the MCL-specific MIPI (Mantle Cell Lymphoma International Prognostic Index), and the proliferation index Ki67 are widely used for risk stratification. However, these tools rely on a small set of clinical and histological variables and cannot capture the full molecular heterogeneity of MCL. The LLMPP MCL35 proliferation assay, which is based on the gene expression of 35 genes weighted by their correlation with the proliferation signature, provides a more granular view but remains a research-grade tool not yet in routine clinical use.
This study by Carreras, Nakamura, and Hamoudi, published in Healthcare (MDPI) in 2022, set out to use artificial neural networks and complementary machine learning techniques to analyze genome-wide gene expression data from 123 MCL patients and identify novel prognostic markers for overall survival.
The gene expression data came from the publicly available GEO dataset GSE93291, curated by the Lymphoma/Leukemia Molecular Profiling Project (LLMPP) and last updated on 25 March 2019 under the direction of Professor Louis Staudt at the National Cancer Institute. The cohort comprised 123 untreated MCL patients with tumor biopsy content of at least 60%, profiled on the Affymetrix HG-U133 Plus 2.0 Array (GPL570), which covers over 47,000 transcripts across 20,862 unique genes after collapsing to gene symbols using max probe values. The patient cohort included a 3.8:1 male-to-female ratio and a median age of 62 years (range 38-93), with a median survival of 2.8 years. Most patients received multiagent chemotherapy (R-CHOP or R-CHOP-like regimens).
Software and hardware: All analyses were performed on a desktop workstation with an AMD Ryzen 7 3700X 8-core processor, 16 GB RAM, and an Nvidia GeForce GTX 1650 GPU. Software included IBM SPSS Statistics 26 and Modeler 18 for the neural networks and machine learning, Gene Set Enrichment Analysis (GSEA) 4.1.0 for pathway enrichment, R 3.6.3 for statistical analysis, JMP Pro 14 for visualization, and Morpheus for matrix visualization.
Neural network architecture: Two complementary architectures were used throughout the pipeline. The multilayer perceptron (MLP) is a feedforward network with one or two hidden layers using hyperbolic tangent or sigmoid activation functions; it can identify more complex relationships but is computationally slower. The radial basis function (RBF) network has a single hidden layer and is faster but less capable of capturing deep nonlinear patterns. For each neural network, cases were randomly split 70% training and 30% testing, with standardized rescaling of covariates (subtracting the mean and dividing by the standard deviation). Performance was evaluated using area under the ROC curve (AUC), percentage of incorrect predictions, and overall classification accuracy.
GSEA validation: Gene Set Enrichment Analysis used the GSE93291 gene cluster text file, phenotype class files (dead vs. alive), and gene set databases. Parameters included 1000 permutations, phenotype permutation type, the HG-U133 Plus 2 chip platform, and a weighted enrichment statistic with signal-to-noise gene ranking. This step confirmed that genes selected by the neural networks were biologically coherent with respect to the survival outcome.
The first analytical approach began with all 20,862 genes of the Affymetrix array as input predictors. Rather than targeting overall survival directly, the neural networks first correlated all genes with 28 known MCL prognostic genes (selected from an initial list of 86 literature-curated genes based on significant Kaplan-Meier log-rank p-values in the GSE93291 cohort). The most important established prognostic genes in this series by Cox regression hazard risk were MKI67 (HR = 4.4, p = 6.6 x 10-9), CDK4 (HR = 4.0, p = 3.2 x 10-8), CHEK1 (HR = 3.0, p = 0.2 x 10-5), CCND1 (HR = 3.1, p = 0.4 x 10-5), and CDKN2C (HR = 2.8, p = 0.8 x 10-5). All five belong to cell cycle and apoptosis pathways.
Step 1 - First reduction: For each target variable (overall survival outcome and each of the 28 prognostic genes), an MLP and an RBF neural network were run. Genes with normalized importance above 70% in any individual network were retained, and the top 1% of genes ranked by average normalized importance across all networks were selected. This step reduced 20,862 genes to 1,394 candidate genes. CCND1 was the best-predicted gene, with only 2.8% incorrect predictions in the testing set.
Step 2 - Second reduction to 58 genes: Using only the 1,394 genes as inputs, an MLP analysis was repeated 20 times, the top 4 runs by AUC were selected, and genes were ranked by their averaged normalized importance. A sequential forward selection process then identified the minimum number of genes that maximized the AUC, resulting in a set of 58 genes. The MLP using 58 genes achieved an AUC of 0.9, with training classification of 86% correct (93.8% alive, 82% dead) and testing classification of 77% correct (82% alive, 74% dead). GSEA confirmed that most of these 58 genes were enriched toward the dead survival outcome. The most important were KIF18A, FANCG, GCNA, YBX3, ZCCHC4, and DMTF1.
Step 3 - Cox regression and final 10 genes: A multivariate Cox regression (backward conditional) on the 58 genes identified 19 genes. Kaplan-Meier survival analysis with cut-off stratification then reduced this to a final set of 10 interpretable prognostic genes. KIF18A (HR = 3.5), YBX3 (HR = 2.3), PEMT (HR = 1.9), GCNA (HR = 1.8), and POGLUT3 (HR = 1.6) were associated with unfavorable overall survival (all log-rank p less than 0.04). SELENOP (HR = 0.6), AMOTL2 (HR = 0.5), IGFBP7 (HR = 0.5), KCTD12 (HR = 0.5), and ADGRG2 (HR = 0.3) were associated with favorable survival (all log-rank p less than 0.04).
The 10 genes identified through neural network-guided dimensionality reduction are not arbitrary statistical artifacts. Each has documented roles in cancer biology, and several have been independently implicated in hematological malignancies or solid tumors. KIF18A encodes a kinesin family microtubule motor protein involved in mitosis; it is overexpressed in multiple cancer types and is a candidate target for small-molecule inhibitors currently under investigation. YBX3 functions as a translation repressor and negative regulator of intrinsic apoptosis signaling, and has been linked to myelodysplastic syndromes and acute myeloid leukemia.
PEMT (phosphatidylethanolamine N-methyltransferase) negatively regulates cell proliferation and is involved in lipoprotein metabolic processes, with a documented critical role in breast cancer progression. GCNA (germ cell nuclear acidic repeat-containing protein) is expressed in germ cells and has been shown to regulate genome stability. POGLUT3 is a protein glucosyltransferase that specifically targets EGF repeats of NOTCH1 and NOTCH3, linking it directly to NOTCH pathway dysregulation, which is relevant in MCL; it has been associated with glioblastoma multiforme tumorigenesis.
Favorable-prognosis genes: SELENOP transports selenium and mediates oxidative stress response, and has been associated with prostate cancer recurrence. AMOTL2 is involved in actin cytoskeleton organization, angiogenesis, and Wnt signaling, with roles in pancreatic and lung cancer proliferation. IGFBP7 (insulin-like growth factor binding protein 7) participates in cell adhesion and growth regulation, and has prognostic relevance in acute lymphoblastic leukemia. KCTD12 functions as a GABA-B receptor auxiliary subunit and is linked to proliferation in breast cancer. ADGRG2 is a G protein-coupled receptor functioning as a tumor suppressor in endometrial cancer.
Correlation with Ki67: The 10 genes were further correlated with MKI67 expression, the gold-standard proliferation marker and most important gene in the LLMPP MCL35 assay. Cases with low MKI67 were associated with high expression of the favorable-prognosis genes KCTD12, ADGRG2, SELENOP, and IGFBP7. Conversely, high MKI67 was associated with high YBX3 expression, consistent with YBX3's role in promoting proliferation by repressing apoptosis signaling. This MKI67 correlation reinforces the biological plausibility of the AI-identified genes.
In parallel to the whole-genome approach, a second independent analytical strategy used 9 curated immune oncology pathway panels containing 7,817 genes in total. These panels covered transcriptome-wide cancer biology, cancer progression and pathways, metabolic pathways, immuno-oncology, and host response gene sets. For each of the 9 panels, an individual MLP neural network predicted overall survival (dead vs. alive), and genes with normalized importance above 70% were extracted and pooled, yielding 125 candidate genes.
Panel-level performance: The predictive accuracy varied by panel. The autoimmune panel achieved the highest AUC (0.98), followed by the pan-cancer human IO360 panel (AUC = 0.94), human inflammation (AUC = 0.89), pan cancer (AUC = 0.89), and metabolic pathways (AUC = 0.87). Notably, some panels showed stronger prediction toward the dead outcome than the alive outcome, suggesting asymmetric gene expression signatures between poor and favorable survival groups. GSEA on the pooled 125 genes produced a sinusoidal-like enrichment pattern, with some genes enriched toward dead and others toward alive outcomes.
Top 15 genes from Method 2: A subsequent MLP analysis ranked the 125 genes by normalized importance, and the top 15 with normalized importance greater than 0.68 were CD8B, CEACAM6, FABP5, CFB, IL6ST, AHR, BST2, ROBO4, AR, ID1, PIK3CD, ITGAX, TYMS, CSF1, and PCK2. Among these, TYMS (thymidylate synthase) was highlighted as a particularly strong individual predictor. TYMS alone predicted overall survival with a hazard risk of 3.2 (95% CI 2.0-5.0, p = 8.9 x 10-7), and high TYMS expression also correlated with high MKI67 by Fisher's exact test (p = 0.001).
Multivariate analysis: In a multivariate Cox regression including all 15 top genes as quantitative variables (backward conditional method), the final step retained three significant independent predictors: TYMS (p less than 0.001, HR = 2.6), AR (androgen receptor, p = 0.012, HR = 1.5, poor prognosis), and CSF1 (colony-stimulating factor 1, p = 0.049, HR = 0.6, favorable prognosis). TYMS encodes thymidylate synthase, a critical enzyme in the folate-dependent de novo thymidine synthesis pathway and the target of fluoropyrimidine chemotherapy agents such as 5-fluorouracil. Its association with MCL proliferation and poor survival aligns with its established role in cancer cell division.
A critical test of the biological relevance of AI-identified prognostic genes is whether they generalize beyond the discovery cohort. The 19 genes identified from Method 1 were applied to two external datasets: the GSE10846 dataset for diffuse large B-cell lymphoma (DLBCL, n = 414), and a pan-cancer series from The Cancer Genome Atlas (TCGA, n = 7,289), comprising the most frequent human cancers including lung, breast, colorectal, prostate, stomach, and liver. A risk-score formula was applied in which beta coefficients from the MCL multivariate Cox regression were multiplied by the corresponding gene expression values to generate patient-level risk scores.
Pan-cancer results: High-risk versus low-risk stratification using the 19-gene MCL signature in the TCGA pan-cancer series yielded a Cox regression hazard risk of 3.3 (95% CI 2.9-3.6, p less than 0.0001). This is a striking finding, indicating that genes identified through AI analysis of a single MCL cohort have prognostic relevance across multiple cancer types. The weight and direction of the survival association varied by cancer subtype, consistent with the context-dependent roles of genes like AMOTL2 (angiogenesis in pancreatic cancer, proliferation in lung cancer) and KCTD12 (breast cancer proliferation). The genes were also validated in the DLBCL series (GSE10846), confirming their cross-lymphoma prognostic applicability.
Machine learning comparison: In addition to neural networks, 11 additional machine learning algorithms were benchmarked against the 19-gene set: logistic regression, Bayesian network, discriminant analysis, KNN algorithm, linear SVM (LSVM), Tree-AS, C5 tree, CHAID, Quest, random trees, and C&R tree. The C5 tree algorithm (C5.0) achieved 95% overall accuracy using only 9 of the 19 genes, making it the most parsimonious and interpretable model. The Bayesian network achieved 85% accuracy using all 19 genes, and enables probability estimation for individual patients based on conditional independencies between gene expression variables and the survival outcome.
C5.0 tree structure: The C5.0 algorithm builds decision trees by splitting samples at the field providing maximum information gain, then recursively splitting subsamples until further splitting is uninformative. Final low-information splits are pruned. In this analysis, the C5.0 model demonstrates that using only 9 genes' expression values, the dead vs. alive outcome can be predicted with 95% accuracy, providing a highly compact and clinically implementable rule set. The Bayesian network, structured as a Tree Augmented Naive Bayes (TAN) model, offers complementary value by handling missing data robustly and quantifying the probability of each survival outcome given any observed gene expression pattern.
The final analytical step combined the 19 genes from Method 1, the 15 genes from Method 2, and the 17 genes of the LLMPP MCL35 proliferation assay signature into a unified 50-gene panel (with overlaps removed). All machine learning and neural network methods were retested using this composite panel as input. The results demonstrated substantially improved predictive accuracy compared to any individual approach. Logistic regression achieved 100% overall accuracy, Bayesian network 92%, discriminant analysis 86%, CHAID 85%, C&R tree 85%, and SVM 81%.
Gene parsimony in combined models: Not all 50 genes were required in every model. CHAID used only 6 genes (E2F2, GCNA, FMNL3, POGLUT3, SELENOP, ZDHHC21) to achieve 85% accuracy, illustrating that a small, biologically interpretable gene panel can capture most of the prognostic information. The C&R tree used 21 genes, including several from both Method 1 (ADGRG2, GCNA, SELENOP) and the MCL35 signature (MKI67, RGS1, RPGRIP1L, TAMM41), demonstrating complementarity between AI-discovered genes and established proliferation markers.
RGS1 immunohistochemistry: Among the 19 genes from Method 1, RGS1 (regulator of G-protein signaling 1) was selected for protein-level validation because a reliable rabbit polyclonal primary antibody was available and because the research group had previously demonstrated that high RGS1 protein expression correlates with poor prognosis in DLBCL. An independent series of 11 MCL cases from Tokai University was stained using automated immunohistochemistry (Leica BOND-MAX stainer), scanned with a Hamamatsu NanoZoomer S360, and visualized with NDP.view2 software.
RGS1 results: The clinicopathological characteristics of the validation series included median age of 72 years (range 41-82), 82% male, 100% CD20 and CCND1 positive, 91% CD5 positive, and Ki67 proliferation index of 10-50%. High RGS1 protein expression correlated with significantly worse overall survival (p = 0.048). No correlation was found between RGS1 expression and other clinicopathological variables in this small series. While the sample size of 11 cases limits the strength of the conclusion, this immunohistochemical result provides orthogonal, protein-level evidence supporting the AI-generated gene expression finding.
Cohort size and single-dataset derivation: The primary discovery cohort comprised 123 MCL cases from the LLMPP GSE93291 dataset. Although this is a well-curated, expert-reviewed series, it remains a single retrospective dataset with relatively modest case numbers for genome-wide analysis. The risk of overfitting is real in any supervised machine learning approach applied to high-dimensional data with a limited number of outcome events. While the 70/30 training-testing split provides internal validation, it does not substitute for independent external cohort validation using a fully separate patient series.
Absence of clinical covariates: The neural network analysis used gene expression data exclusively as predictors. Patient-level background information such as age, sex, comorbidities, treatment regimen details, and clinical stage was not integrated into the primary AI models. The authors acknowledge this as a deliberate proof-of-concept choice, noting that incorporating clinical variables into the neural networks (as they have done in complementary DLBCL analyses) would likely further improve predictive accuracy and yield more clinically contextualized models.
Interpretability and explainability: The multilayer perceptron neural networks function as "black box" models, ranking genes by normalized importance without providing mechanistic explanations for their prognostic associations. While the authors addressed this limitation by pairing each ANN result with conventional Kaplan-Meier analysis, Cox regression, GSEA, and multiple interpretable machine learning models (C5 tree, Bayesian network), the underlying neural network decision logic remains opaque. The authors explicitly note that explainable AI (XAI) development will be a focus of future work, and they highlight that in medicine, AI tools can be clinically validated even when the mechanism of their decision-making cannot be fully understood.
Future directions: The study identifies three main future research priorities. First, external validation of the highlighted genes in independent MCL cohorts and in other NHL subtypes. Second, integration of clinical background information (age, stage, treatment data) into the neural network input layers to create multimodal prognostic models. Third, functional validation of genes such as TYMS and KIF18A as therapeutic targets in MCL, given that TYMS is already a target of fluoropyrimidine agents and KIF18A inhibitors are under development. The pan-cancer generalizability of the 19-gene MCL signature to the TCGA series (n = 7,289) further motivates investigating these genes as cross-cancer biomarkers.