2.1 Clinical and genetic data

In this research, we conducted an examination of both clinical and genetic data derived from three distinct cohorts of patients diagnosed with DLBCL. The first cohort (JSPH cohort) included 69 newly diagnosed less-defined subtype DLBCL patients diagnosed in the First Affiliated Hospital of Nanjing Medical University (Jiangsu Province Hospital) from April 2021 to September 2023 (Table 1). The first-line treatments of all the newly diagnosed DLBCL patients were standard R-CHOP regimens.

The analysis also encompassed two additional published cohorts of DLBCL that provided accessible clinical data, DNA mutation profiles, and gene expression data^{7, 8} (Table S1). All patients received the standard R-CHOP regimen. Clinical data were gathered from the primary sources, and details regarding LymphGen subtypes were accessible within the GSE181063. The LymphGen subtype of GSE117556 was determined by employing the LymphGen implement (https://llmpp.nih.gov/lymphgen/index.php). The gene expressions of GSE117556 and GSE181063 were mixed into a single matrix file by utilizing the R package “sva”.

To curate a comprehensive list of PCD genes, we sourced genes linked to 19 distinct PCD patterns from esteemed scientific databases and literature. This compilation included Gene Set Enrichment Analysis (GSEA) gene sets, Kyoto Encyclopedia of Genes and Genome (KEGG) pathways, scholarly reviews, and a manual curation process. Post the removal of any redundant genes, our final dataset comprised 2118 unique PCD-associated genes that were subsequently subjected to further analysis.^9-11

2.2 Liquid biopsy data acquisition

The concentration of circulating tumour DNA (ctDNA) in each sample, expressed in human genomic equivalents per millilitre (hGE/mL), was determined by applying the subsequent formula: multiply the mean value of variant allele frequency (VAF) by the cell-free DNA concentration (pg/mL plasma) and then divide by the haploid genomic equivalent weight of 3.3 pg. The results were expressed as base-10 logarithms (log hGE/mL). Post-infusion samples were qualitatively and quantitatively assessed for ctDNA at the end of treatment (EOT) and reported as negative or positive. In cases of a positive result, the concentration of ctDNA was quantified.

We focused on mutations identified in baseline plasma as potential disease biomarkers and investigated their presence in subsequent plasma samples after infusion. MRD status was determined: MRD positive indicated the re-detection of any baseline or new mutation, whereas MRD negative signified the complete clearance of all baseline mutations.

2.3 Signature generated from machine learning-based integrative approaches

To ascertain a reliable consensus on genes linked to pediatric cardiomyopathy (PCD) with precision and consistency, a comprehensive integration of 10 machine-learning algorithms and 70 algorithmic combinations was employed. Within the training dataset, we developed 70 distinct machine learning-based models, each trained using leave-one-out cross-validation (LOOCV). Among these models, the combination of coxBoost and GBM was identified as the most suitable model, as it demonstrated the highest area under the curve (AUC) score in the validation dataset.

2.4 Tumour microenvironment analysis and drug sensitivity prediction

Expression profiles of key genes and the levels of various immune cell infiltrations were acquired and processed using a suite of computational algorithms. Furthermore, we employed the ‘oncoPredict’ R package to calculate the half-maximal inhibitory concentration (IC50) for a range of chemotherapeutic agents, thereby identifying which drugs might be most effective for the high-risk cohort.

2.5 Statistical analysis

Discrepancies among various risk groups were assessed using Unpaired Students' t-tests, one-way analysis of variance, Chi-square analyses, or Fisher's exact tests, depending on the context. A p-value of less than .05 was considered to indicate statistical significance. The data analyses were conducted utilizing R software (version 4.3.1), SPSS software (version 20.0) and GraphPad Prism (version 10.0).

The detailed methods of clinical data, genetic data, liquid biopsy data acquisition, a signature generated from machine learning-based integrative approaches, tumour microenvironment analysis and drug sensitivity prediction, functional enrichment analysis, nomogram building and assessment, unsupervised clustering of model genes and statistical analysis were shown in Supporting Information methods.

3.1 Preliminary screening of PCD-related regulators in less-defined subtype DLBCL

The reanalysis involved integrating data from two bulk RNA sequencing cohorts, identified as GSE117556 (N = 116) and GSE181063 (N = 154), along with the JSPH cohort (N = 69). We curated a set of 2118 regulatory genes encompassing 19 PCD patterns sourced from literature (Figure 1A and Table S2). Patients from two public databases were divided in an 8:2 ratio based on PFS endpoint status, resulting in a training dataset of 216 patients and an internal test dataset of 54 patients. The JSPH cohort served as the external test dataset. Subsequently, we conducted univariate Cox regression analysis, identifying 212 regulatory genes associated with PFS in the training set (Figure 1B).

3.2 Integrative construction of a consensus signature

We utilized a suite of 10 distinct machine-learning algorithms to develop the PCDI. In the training set, we constructed 70 machine learning models, employing LOOCV for training each one. The concordance index (C-index) was utilized to assess the performance of the models. It was employed as a criterion for selecting the most effective model, focusing on the one that exhibited the highest mean C-index across all validation datasets (Figure 1C). The coxBoost + GBM was selected among the top five performing model combinations for gene model selection (C-index = 0.75, Figure 1D). Ultimately, our algorithm identified eight key PCDI signature genes: FLT3 (FMS-like tyrosine kinase 3), SORL1 (sortilin-related receptor 1), CD8A (cluster of differentiation 8A), COL13A1 (collagen type XIII alpha 1), BCL2L1 (BCL2-like protein 1), MPG (methylpurine glycosylase), DYRK2 (dual-specificity tyrosine-(Y)-phosphorylation regulated kinase 2), and CAMK2B (calcium/calmodulin-dependent protein kinase II beta). Importantly, DLBCL patients from public databases with elevated expression of these signature genes exhibited prolonged survival (Figure S1A–H). Similar results were demonstrated in the JSPH cohort (Figure S1I–P).

Our analysis of patients diagnosed with less-defined subtype DLBCL in the GSE101063, GSE117556, and JSPH cohort revealed that high PCDI scores were significantly associated with poorer prognosis, as depicted in Figure 1E (cutoff = 0.916, training dataset: p < .001, test dataset1: p = .003 and test dataset2: p < .001). Patients with high PCDI scores were classified as the high-risk group, while the rest of the patients were designated as the low-risk group. In the training cohort, the AUC scores were 0.84, 0.90, and 0.89 for each of the three consecutive years, respectively. The AUC scores, as indicated by the internal test dataset, were 0.71, 0.80 and 0.85, respectively. Furthermore, the external test dataset, known as the JSPH cohort, presented AUC scores of 0.77, 0.86 and 0.81 (Figure 1F). Furthermore, we performed univariate Cox analysis on various clinical features in public databases and the JSPH cohort. The results revealed survival differences among patients with different risk groups within subgroups defined by clinical features (Figure S2A,B). These findings underscore the prognostic significance of PCDI.

3.3 Characteristics annotation of PCDI signature genes

Heatmaps of the PCDI-related prognostic model and clinical characteristics of public databases and the JSPH cohort are depicted in Figures 2A and 2B. In public databases, discernible disparities were identified between low-risk and high-risk groups with respect to PCDI scores, progression of disease within 2 years (POD24), cell of origin (COO), Ann Arbor stage, and IPI risk group (p < .001 for all above), as determined by Chi-square tests. The JSPH cohort showed similar results. Additionally, liquid biopsy analysis indicated higher pre-treatment ctDNA concentrations in high-risk patients (p < .05). At the end of therapy response evaluation, ctDNA was more frequently detected in the high-risk group (p < .05). We also compared PCDI scores across different subgroups based on multiple clinical variables in both the combined public database and the JSPH cohort. In the combined public database, the advanced Ann Arbor stage and POD24 were associated with higher PCDI scores (Figure S3A–I). Similarly, in the JSPH cohort, non-GCB subtype, advanced Ann Arbor stage, extranodal involvement > 1, IPI high-risk group, PD/stable disease (SD) at the end of the first-line therapy, MRD positivity, and POD24 correlated with higher PCDI scores (Figure S3J–T). We also analyzed the association between distinct risk groups and established markers of DLBCL. We observed a higher prevalence of whole blood EBV-DNA positivity in the high-risk group, which correlates with a poorer prognosis. However, we regret to report that no significant statistical differences were found in the expression of MYC, BCL2 and BCL6 between high-risk and low-risk groups in either the integrated cohort or the JSPH cohort (Table S3). Figure 2C provides an overview of the relationship between the expression of eight PCDI signature genes and the 24 distinct types of immune cells, encompassing 18 kinds of T cells along with other types of immune cells infiltrated in less-defined subtype DLBCL. Additionally, we found that patients with high risk are more prone to exhibit higher central nervous system (CNS)-IPI scores (p < .001), CNS involvement, and involvement of the adrenal glands, kidneys, and breasts, which are also associated with CNS involvement (Figure S4A,B). As illustrated in Figures 2D and 2E, several pathways significantly correlated with these 8 genes, encompassing position and regulation of PCD patterns, lipid metabolism, and various signalling pathways related to immunological regulation and LME.

3.4 The dynamic analysis of liquid biopsy between different PCDI groups

ctDNA is considered a novel prognostic and predictive biomarker with significant prognostic value for PFS and OS.¹² Our results revealed that high-risk patients typically exhibited higher ctDNA concentrations at pre-treatment, and post-treatment (Figures 3A and 3B). However, no significant difference was observed in the log-fold change of ctDNA burden (Figure 3C). Through dynamic analysis, we detected that almost high-risk patients tend to exhibit not only MRD positivity but also an elevated change of ctDNA burden (Figures 3D and 3E).

Tumour mutation analysis was performed on tissue and plasma samples diagnosed with less-defined subtype DLBCL. Using the Fisher test, CCND3 mutations, associated with worse prognosis, were more prevalent in the plasma of the high-risk group¹³ (Figure 3F and Table S4). Conversely, a higher frequency of BCL2 mutations was observed in the tissue samples of the low-risk group, while NSD2 mutations which are associated with resistance of Imbruvica, were more frequently observed in the high-risk group^{14, 15} (Figure S5 and Table S5). Additionally, in the combined public database, SGK1 mutations were more frequently observed in patients with low PCDI scores (Table S6).

We then compared TMB between individuals with high and low PCDI scores. To explore the link between TMB and PCDI, our findings revealed that the TMB was considerably elevated in the high-risk subgroup, as depicted in Figure 3G. Notably, patients exhibiting reduced TMB demonstrated improved survival rates in contrast to those with elevated TMB levels, with a marked statistical disparity (p = .008, Figure 3H). To unravel the combined predictive value of TMB and PCDI for survival, we stratified patients according to these criteria and performed a survival analysis. Each factor's presence was allocated one point: a score of 2 was indicative of high risk, a score of 1 was of intermediate risk, and a score of 0 of low risk. Patients with low risk exhibited the most favourable prognosis, whereas those classified as high risk had the least favorable (low-risk vs. intermediate-risk, p = .668; intermediate-risk vs. high-risk, p < .001, Figure 3I).

3.5 The underlying biological mechanisms of PCDI groups

To deepen our comprehension of the biological mechanisms associated with PCDI groups, we conducted a pathway enrichment analysis. The results of KEGG are shown in Figure 4A. In terms of cellular processes, PCDI were predominantly enriched in focal adhesion, lysosome, efferocytosis and adherens junction. For environmental information processing, PCDI were most highly concentrated in the PI3K-Akt and NF-kappa B signalling pathway. Additionally, PD-L1 expression and PD-1 checkpoint pathway were also enriched in the high-risk group. The gene ontology (GO) results of patients with high risk are presented in Figure 4B.

Furthermore, we utilized GSEA to uncover potential pathways correlated with PCDI. As illustrated in Figure 4C and Figure 4E, the high-risk patients showed significant enrichment in pathways similar to those identified by KEGG. To validate these results, we applied similar methods to patients in the JSPH cohort. Six of the eight regulatory genes have lower expression in individuals with high PCDI scores (Figure S6A). The results of KEGG and GSEA also exhibited disorder of several significant pathways, including Th1, Th2 and Th17 cell differentiation, JAK-STAT, NF-kappa B and PD-L1 expression and PD-1 checkpoint pathway in cancer (Figure S6C–E).

Based on the GDSC database, patients classified as high risk were found to be more responsive to a variety of specific anti-neoplastic medications, such as 5-Fluorouracil, sorafenib, MK-1775, BMS-345541, JAK1 inhibitor and AZD6738 (Figure 4E). Among the anti-cancer drugs mentioned above, JAK1-inhibitor showed a comparatively lower IC50 in patients in the high-risk group compared to those in the low-risk group.

3.6 Potential biological mechanisms in tumour immune process related to the PCDI signature

Cancer stem cells (CSCs) constitute a minor subset of cells within a lymphoma that share characteristics with normal stem cells, notably their capacity to differentiate into the various cell types present within a specific tumour sample. In our quest to explore the gene expression profiles and epigenetic signatures of CSCs, we determined mRNAsi scores in samples retrieved from publicly accessible databases. mDNAsi scores were not calculated due to the lack of DNA methylation data. A heatmap of mRNAsi scores and clinical characters is displayed in Figure 5A. No significant correlation between risk scores and mRNAsi scores was observed (Figure 5B). However, higher mRNAsi scores can be observed in patients with high risk. (p = .016, Figure 5C).

The development of an antitumor immune response involves a series of steps, succinctly encapsulated within the framework of the “cancer immunity cycle¹⁶”. In our quest to explore the biological mechanisms underlying the PCDI signature, we assessed the trajectory of cancer immunity. Notably, within the cohort identified as high-risk, disruptions were observed in three distinct stages of the tumour immune cycle: phase 2 (cancer antigen presentation), phase 4 (tumour immunized infiltrating cells recruitment) and phase 5 (immune tissues influx) (Figure 5D). Furthermore, the levels of expression for 178 genes linked to protective mechanisms against cancer, spanning the seven phases of the cancer-immunity process, are illustrated in Figure 5E.

3.7 Immune characteristics between distinct PCDI groups

To assess the characteristics of the tumour microenvironment between different PCDI groups, we examined the presence of immune cell infiltration utilizing seven distinct algorithms (Figure 6A). Low PCDI scores were associated with higher infiltration of CD4+ T cells, CD8+ T cells and M1 macrophage. Patients exhibiting high PCDI scores had notably reduced Stromal and ESTIMATE scores. Similar results were observed in patients from the JSPH cohort grouped by PCDI scores (Table S7). Furthermore, we employed the ssGSEA enrichment score to investigate the correlation between the PCDI signature and various immune cell subsets as well as their functional activities. Our results showed that among CD56 bright natural killer cell, eosinophil, natural killer T cell and immature dendritic cell, ssGSEA scores of patients with high PCDI scores were significantly different from those with low PCDI scores for these immune-related cell functions (Figure 6C). In the JSPH cohort, central memory CD4 T cell was more commonly observed in the low-risk group (Figure S7). Additionally, PCDI characteristic scores were negatively correlated with immune checkpoints, including HLA-DOA, HLA-DOB, ICOS, CD27, CTLA4, PD-L2 and so on (Figure 6D).

3.8 Development and evaluation of the nomogram prognostic model

To evaluate the prognostic impact of PCDI independently, separate univariate and multivariate Cox regression analyses were conducted. The results indicated that PCDI emerged as a substantial risk factor in the univariate Cox regression analysis (hazard ratio [HR] = 4.4, 95% confidence interval [95%CI]: 3.5–5.6, p < .001, Figure 7A). When considering multivariate analysis, the PCDI continued to show its independent prognostic significance in patients with less-defined subtype DLBCL, even after considering other potential influencing factors (HR = 1.7, 95%CI: 1.1–2.7, p < .001, Figure 7B). Utilizing multivariable Cox and stepwise regression analyses, a prognostic nomogram was established within public databases to estimate the 1-, 3- and 5-year PFS for patients with less-defined subtype DLBCL (Figure 7C). Given the results of our study, we have developed an online tool for risk assessment (https://xulymphoma.shinyapps.io/PCDI_pred/). This user-friendly platform is designed to efficiently evaluate risks and serves as a basis for the creation of a practical algorithm suitable for clinical use. The accuracy of the nomogram model's predictive capabilities for 1-, 3- and 5-year survival rates was confirmed through calibration and ROC curve analysis (Figure 7D,E). Furthermore, decision curve analysis (DCA) indicated that our nomogram model's predictive performance surpassed that of other models in the study (Figure 7F–H). Similar results were observed in the JSPH cohort (Figure 8A–E).

3.9 Consensus clustering to identify two clusters of patients

To investigate the role of PCDI signature genes in the progression of less-defined subtype DLBCL, we undertook an exhaustive clustering analysis focusing on the activity values associated with the eight signature indicators. K = 2 was determined to be optimal stable clustering (Figure 8A,B). This analysis divided the samples from public databases into two subgroups: C1 (N = 160) and C2 (N = 110) (Figure 8C). Notably, there were plenty of variations in PFS between the two subgroups (p = .044, Figure 8D). C1 was predominantly composed of patients with high PCDI scores, while C2 was mainly composed of those with low PCDI scores (p = .006). Furthermore, the heatmap depicting PCDI consensus clusters highlighted significant disparities between clusters C1 and C2 sections in terms of risk score, COO and age (Figure 8E). Moreover, the pathway-based Gene Set Variation Analysis (GSVA) findings revealed that the C1 subtype exhibited increased activation of pathways related to tumours and immunity, including the T cell receptor, B cell receptor, and JAK-STAT signalling pathway, along with endocytosis and apoptosis. This suggests a strong correlation between the PCDI signature and these canonical tumour-associated pathways (Figure 8F). Additionally, the difference in immune cell infiltration is depicted in Figure 8G.