2.1 Patients and sample characteristics

This study included 119 DLBCL patient samples and 117 healthy control samples from Peking University Third Hospital (Table S1). In all cases, DLBCL was diagnosed using appropriate diagnostic criteria from the 2016 WHO classification of lymphoid neoplasms, and 81 patients received standard R-CHOP therapy. Before any treatment, lactate dehydrogenase (LDH), beta2 microglobulin (β2MG) and cfDNA were collected. Disease stage was defined by the Ann Arbor staging system, this study was conducted in accordance with the Declaration of Helsinki, and all samples had signed patient consent. In the actual data analysis, due to the small amount of DMRs in some patient samples, 97 patients among 119 patients were retained for early diagnosis cohort study, and 80 patients among 81 follow-up samples treated with R-CHOP were retained for further treatment response cohort study. Some of the discarded samples participated in the DMR identification process.

2.2 Isolation of plasma cfDNA

In this study, 2–8 mL of whole blood was collected into 8.5 mL Cell-Free DNA Blood Collection Tubes (Cat#07785674001, Roche), and plasma separation was performed within 72 h. First, we centrifuged at 1350 × g for 12 min at 4°C, carefully removed the light yellow supernatant liquid, to avoid contaminating the white blood cell layer, and immediately transferred the supernatant liquid to a 2 mL DNase-free sterile centrifuge tube. Second, we centrifuged at 13 500 × g for 5 min at 4°C, transferred the supernatant to a 2 mL DNase-free sterile centrifuge tube, and then used QIAamp MinElute ccfDNA Mini Kit (Cat#55204, Qiagen) to extract cfDNA, followed by using Qubit dsDNA HS Assay Kit (Cat#Q32854, Thermo Fisher Scientific) to determine DNA concentration and using Agilent 4200 TapeStation (Agilent) for fragment analysis.

2.3 Library preparation

We constructed methylation libraries by combining multiple kits, and used NadPrep Methyl Library Preparation Module (Cat#1002502, Nanodigmbio) to end repair cfDNA, add ‘A’, and ligase adapters containing fully cytosine-methylated unique molecular identifier (UMI). For methylation conversion, NEBNEXT Enzymatic Methyl-seq Conversion Module (Cat#E7125L, New England Biolabs) was used. The methylation library used the Qubit dsDNA HS Assay Kit (Cat#Q32854, Thermo Fisher Scientific) to determine the concentration and Agilent 4200 TapeStation for library fragment analysis.

2.4 Target region capture sequencing

Target region hybridization was performed by Twist Fast Hybridization and Wash Kit (Cat#101174, Twist Bioscience). We mixed eight methylation libraries with different indexes; each library was 187.5 ng. The resulting mixture was combined with 4 µL Twist Custom Methylation Panel (Cat#105520, Twist Bioscience), 8 µL Universal Blockers and 5 µL Blocker Solution (Cat#103557, Twist Bioscience) and concentrated to a dry powder state. The Custom Methylation Panel covered 123 M methylation region, which came from the latest database versions UCSC, Ensemble, ENCODE and so on.Thereafter, the powder was dissolved by 20 µL Fast Hybridization mix, then 30 µL Enhancer was added and mixed thoroughly. The following condition was carried out: incubation at 95°C for 5 min, then 60°C for 16–24 h. After hybridization was completed, streptavidin beads were used for binding and cleaning. The product was amplified and enriched by 25 µL KAPA HiFi Hotstart Readymix (Cat#07958927001, Roche) and 2.5 µL Amplification Primer under the following conditions: pre-denaturation at 98°C for 45 s, denaturation at 98°C for 15 s, annealing at 60°C for 30 s, extension at 72°C for 30 s, including denaturation to extension steps for eight cycles, and final extension at 72°C for 1 min. The sequencing platform is Illumina NovaSeq 6000 (Illumina), and the sequencing raw data for each sample was approximately 20–40G.

2.5 Sequencing data processing and comparison

We used umitools⁴⁰ (v1.1.2) to extract the UMI of each read and merged it into the fastq sequence identifier, then used trim_galore (v0.6.6) to filter the data, and used bismark⁴¹ (v0.23.1) to map the filtered fastq to the hg19 reference genome prepared through genome preparation, and the ‘deduplicate_bismark –barcode’ mode was used to deduplicate.

2.6 Multimodal information extraction

2.7 Analysis of methylation levels in standard samples

We extracted CpG site methylation levels and performed depth statistics on gradient methylation level standard samples with different DNA contents. The methylation levels of CpG sites on different gradient methylation level standard samples with the same DNA content were fit by using the linear regression function of R and p-values were calculated.

2.8 Model construction

We defined three classifiers, respectively: logistic regression, random forest and AdaBoost, and defined random seeds as 42. Standardization was performed using the sklearn StandardScaler function. Specifically, we used StandardScaler to fit the training set and normalized the entire dataset. For each omics data, the model construction methods of different classifiers were as follows:

In the training set, recursive feature elimination with cross-validation (RFECV) was used for feature screening, where cv was set to 10 and the evaluation metric was set to ‘roc_auc’.

We performed 10-fold cross-validation on each classifier after the hyperparameters, and calculated the average area under the curve (AUC).

For each classifier after the hyperparameters, we fit it on the training set, and performed prediction evaluation on the verification set, and computed the performance metrics related to the model.

2.9 Model integration

By computing the average of 10-fold AUC of three classifiers after the hyperparameters under each omics, the model with the largest average of AUC was the optimal model for the omics. We set the initial classification threshold to 0.5, increased it in steps of 0.01, counted the sensitivity of the optimal model under each omics when the specificity reaches 99% on the training set and saved the current classification threshold for subsequent analysis. We expected that the classification threshold obtained through this step could make the specificity of the model in the validation set high enough, in order to meet the high requirements of specificity in different tasks (early diagnosis or treatment response prediction). The positive prediction probability value generated by seven individual omics optimal classification model was used as the feature of integrated model training. Specifically, we concatenated the positive prediction probabilities of the seven individual omics as columns, with each row representing the positive prediction probabilities for a sample across different omics. The sample cohorts for the training and test sets were the same as those for the single-omics datasets. We used the random forest to build the COMOS model on the dataset of all individual omics positive predictive probability, and the random seed was set to 42. The default parameters were used for early diagnosis, and since the sample size of the treatment response prediction is smaller than that of the early diagnosis, the n_estimators was set to 5 to prevent the model from over-fitting and reduce the model complexity. Similarly, the classification threshold when the specificity of the integrated model reaches 99% on the training set was saved for subsequent analysis.

2.10 Model evaluation

The 1000 bootstrap resamples were used to obtain the 95% confidence intervals of AUC, sensitivity, specificity and other indicators. The DelongTest method was used to calculate the difference between the two models.

In the early diagnosis of cancer at different stages, due to the limited number of patients, cancer stages I and II were combined into the early stage, and cancer stages III and IV were combined into the late stage. We utilized the Wilson method to separately assess the diagnostic ability of the model in early- or late-stage patients and normal control sample sets, and calculated the 95% confidence intervals.

2.11 Statistical analysis

The heatmap for early diagnosis was drawn by calculating the Z-score for each omics feature in tumour patients relative to normal controls. The calculation method for the Z-score of a cancer group/normal control group was: $$Z_{T} = {\frac{X_{T} - \mu_{\textit{NC}}}{\sigma_{\textit{NC}}}}/ Z_{\textit{NC}} = {\frac{X_{\textit{NC}} - \mu_{\textit{NC}}}{\sigma_{\textit{NC}}}} $$, and mapped the Z-score value to the corresponding colour. t-SNE dimensionality reduction was accomplished using the t-Distributed Stochastic Neighbour Embedding (tSNE) function from the sklearn (0.24.2) package,⁴⁵ and ‘n_components’ and the random seed was set to 2 and 42. We used ChIPseeker⁴⁶ (v1.34.1) to annotate DMRs to related genes and genomic elements, set the TSS region range to −2 Kbp to 2 Kbp and used clusterProfiler⁴⁷ (v4.6.2) to annotate related KEGG signalling pathways. In order to compare the advantages and disadvantages of the treatment response prediction model and the clinical markers LDH and β2MG, the AUC of LDH and β2MG under a certain number of samples were calculated by numerical integration method. This study used the sklearn package of Python (v3.6.9) for machine learning modelling, and matplotlib and R for plotting.

3.1 COMOS overview for methylation, CNA and fragmentation detection

The COMOS approach mainly included three steps, a deep targeted whole methylome sequencing (TWMS), multi-omics features extraction and model integration (Figure 1). To preserve the integrity of the cfDNA,⁴⁸ an enzyme-mediated methylation conversion was utilized to construct a sequencing library containing a fully cytosine-methylated UMI. The whole target methylome was then captured by hybridization to a custom probe panel (Figure S1A,B). CNA, FSR, BSN, BSC, BSD and BSE were consequently extracted and utilized to construct an integrating model for early diagnosis and treatment response prediction of DLBCL.

To verify methylation detection of TWMS, the fully methylated human genomic DNA and unmethylated DNA were mixed to prepare reference standards with 0%, 5%, 25%, 50% and 100% methylation frequencies. Standards of 1, 5, 10, 20 and 50 ng were input for each sequencing experiment with two technical replicates, and the methylation sites with coverage of ≥300 were selected for statistical analysis. The results showed that the methylation frequencies of 0−100% were detected, consistent with theoretical frequencies, and displayed a linear correlation (R = .99, p <0.01, Figure 2A,B). Furthermore, as little as 1 ng of sample could be detected and the detection capability was reproducible at all methylation sites evaluated (R = 0.99, p < 0.01, Figure 2C). These results confirmed the reliability of TWMS in detecting methylation variations.

Given that TWMS ensured the integrity of cfDNA and covered the entire methylation region, it was theoretically effective to obtain a large amount of CNA and fragmentation information. Therefore, to evaluate the feasibility of using the TWMS methylation data for fragmentation and CNA analysis, we randomly included four cancer patient samples to perform a consistency analysis between TWMS and non-methylated converted TWMS (non-TWMS) data (see Methods − Multimodal information extraction for details). TWMS and non-TWMS data from sample 4–14 and other samples 4–06, 4–65 and 4–36 (Figure S2A−C) showed a highly consistent copy number pattern with a linear correlation (R = .86, p < .01, Figure 2D). Similarly, cfDNA fragmentation analysis of sample 4–14 and other cancer samples was performed simultaneously (Figure S2D−F). The ratio of the number of fragments in each 5-Mb interval to the total number of fragments was calculated (see Methods − Multimodal information extraction for details), which showed a highly consistent fragmentation pattern and a linear correlation (R = .94, p < .01, Figure 2E). These results demonstrated the accuracy of TWMS in detecting CNA and fragmentation. Besides, a fully cytosine-methylated UMI adapter (Figure S1B) was used to perform deduplication analysis, which increased the number of unique reads and avoided excessive deduplication (Figure S1C). In short, the COMOS approach, on the basis of an elaborate experimental design and data analysis, could improve the detection accuracy of methylation, CNA and fragmentation.

3.2 Multi-omics profiling between DLBCL and healthy controls

In order to comprehensively prove the performance of COMOS in cancer detection, a clinical cohort sample study was conducted and validated, including 117 healthy controls and 97 pathologically confirmed DLBCL patients (Tables S2). In the early diagnosis study, stages I and II were used as early stage, stages III and IV as late stage, and the cohort samples were divided proportionally into the training set (healthy control = 70, DLBCL = 58) and the verification set (healthy control = 47, DLBCL = 39). The extracted DMR, CNA, FSR, BSN, BSC, BSD and BSE features from the training set were used to build an integrating model, and then evaluated on the validation set (Figure 3A). Overall statistical results of DMR, CNA and FSR were analysed and the results showed differences between the groups (Figure 3B). A total of 264 DMR regions were identified. The DLBCL cohort showed a decrease in the methylation rate of DMR regions compared to the healthy cohort (Figure 3C,D). Except for intron regions, DMRs were mainly clustered in the promoter region, accounting for approximately 20% (Figure S3A,B). There was a total of 567 5-Mb bin regions in the FSR, and averaging the FSRs of all patients in each bin region of the two different cohorts revealed that the mean FSR of the DLBCL cohort was higher than that of the healthy cohort in each bin interval (Figure 3E). Meanwhile, the CNA of the DLBCL cohort relative to the healthy cohort was obtained for 1942 1-Mb intervals, revealing 1283 regions of copy number changes and a total of 686 regions of copy number gain (Figure 3F).

Most importantly, the cfDNA breakpoint patterns in regions surrounding chromatin-associated features, including nucleosomes, CpG islands, DNase clusters and enhancers, were analysed in depth to compare differences between the healthy and cancer groups. To quantify this difference, a breakpoint score was developed. Specifically, using chromatin-associated features as a centre point, all cfDNA fragments within 200, 300, 150 and 500 bp upstream and downstream of the centre point were counted, as well as the number of cfDNA fragments at the breakpoint. The breakpoint score was calculated as the ratio of the latter to the former (Figure 4A). In certain regions, the differences tended to be present and varied (Figure 4B−E). Taking BSN as an example, the breakpoint score of the DLBCL cohort fluctuated greatly in the area upstream and downstream of the nucleosome-occupied region compared to the healthy control cohort, showing extremely significant differences between the two groups in one local area. The breakpoint score was lower in nucleosome occupancy regions than that in nucleosome binding regions, consistent with previous studies.³⁹ Similarly, there were significant differences in some local areas of BSC, BSD and BSE. In a word, the multi-omics results produced by COMOS provided comprehensive and robust data underpinning for the construction of early diagnosis models.

3.3 COMOS performance in DLBCL early diagnosis

In the training set, through comparative analysis of healthy controls and DLBCL patients, the candidate feature of each group was screened out using RFECV, and a differential heat map was obtained (Figure 5A). The results showed that in addition to FSR and CNA, other omics could distinguish healthy controls from cancer patients to a certain extent. At the same time, the characteristics of all omics through tSNE (Figure S4) were visually analysed, suggesting that DMR could significantly discriminate cancer patients from healthy controls (Figure S4A). Based on the above-selected omics features, a set of classification models, as well as integration models using the selected DMR, FSR, CNA, BSN, BSC, BSD and BSE were constructed (Table S3) and evaluated on the training and validation sets (Figure 5B, Figure S5A and Table S2). The results showed that COMOS had the highest AUC of .993 (95% CI: .978–1). Further analysis of the model AUC revealed that COMOS was significantly better than DMR (p < .05), BSN (p < .05), BSC (p < .01), BSD (p < .01) and BSE (p < .01) (Figure 5C). By adjusting the classification threshold on the training set, the sensitivity at 99% specificity of the training set was obtained (Figure S5B). In the validation set, COMOS demonstrated a detection sensitivity of 95% (95% CI: 88−100%) and a specificity of 98% (95% CI: 93−100%) using the same classification threshold (Figure 5D and Figure S5C). COMOS calculated a score corresponding to the probability of being a tumour patient, and the results showed that the healthy control group and the DLBCL group were effectively identified (Figure S5D). As the classification threshold was changed, the sensitivity, specificity and accuracy of COMOS fluctuated in a gradual manner (Figure S5E). In addition, the Youden Index of COMOS on the validation set was .927 (95% CI: .838−1) (Figure 5E), which was superior to that of the individual omics model, suggesting that COMOS had notable advantages in sensitivity and specificity. Furthermore, we randomly combined DMR, FSR and CNA to construct integrating models. The integrating model containing FSR was improved to a certain extent. The AUC value and specificity of COMOS, integrating BSN, BSC, BSD and BSE were superior to the multi-omics model mentioned above. These results indicated that the fragmentation modalities further exploited the potential of the fragmentomics and improved the performance of COMOS (Figure 5F and Figure S5F).

At the same time, to further evaluate the diagnostic performance of COMOS in different stages of DLBCL, a total of 89 patients with clinical stages were obtained, and the early stage (stages I and II, 11 validation set samples out of 26 samples) and the late stage (III and IV, 25 validation set samples out of 63 samples) were evaluated. In the validation set, the sensitivity for early-stage and late-stage DLBCL reached 91% (95% CI: 62−98%) and 96% (95% CI: 80−99%), respectively (Figure 5G). Our findings indicated that COMOS performed much better in the early diagnosis of DLBCL.

3.4 Multi-omics profiling between partial response/complete response and progressive disease/stable disease patients

According to treatment response, patients were defined as progressive disease (PD), stable disease (SD), partial response (PR) and complete response (CR).⁴⁹ PD and SD were assigned to the R-CHOP ineffective group, while PR and CR to the R-CHOP effective group. Besides, 80 patients were divided into a training set and a validation set (Table S4). To explore the potential of COMOS in predicting R-CHOP treatment response, the statistical results under different omics were analysed. DMR region exhibited lower methylation levels in the PR/CR patient group compared to the PD/SD patient group (Figures 6A and S6A,B). KEGG analysis suggested that DMR-related genes were enriched in the MAPK signalling pathway, calcium signalling pathway, Ras signalling pathway and other pathways closely related to tumour occurrence and development (Figure 6B). In each bin interval, the average of FSR in PR/CR patient cohort was higher than in the PD/SD patient cohort (Figure 6C). In CNA, the PR/CR patient cohort had 260 gain regions and 231 loss regions compared to the PD/SD patient cohort (Figure 6D). Similarly, the statistical results of the BSN, BSC, BSD and BSE also showed varying degrees of difference (Figure 6E−H). These data demonstrated that COMOS could extract faithful multi-omics information from the PR/CR and the PD/SD patient cohort.

3.5 COMOS performance in DLBCL response prediction

In the training set, through the comparative analysis of PD/SD control and PR/CR groups, the candidate features of each omics were screened by RFECV, and the characteristics of all omics were visually analysed by tSNE (Figure S7). The results showed that the DMR characteristic sites (Figure S7A) could clearly separate PR/CR and PD/SD cohorts, supporting their role in treatment response detection. An integrating multiple classification models on the basis of the above-selected omics features were constructed (Table S5), and the performance was evaluated on the training and validation sets (Figure S8A,B). The results showed that the AUC of COMOS reached .903 (95% CI: .764–1), which was better than any individual omics model, and the AUC difference analysis showed that the COMOS performance was significantly better than CNA (p < .05), BSC (p < .01) and BSD (p < .05) (Figure S8C). In addition, the performance of COMOS was compared with a randomly combined model of DMR, FSR and CNA. Interestingly, COMOS also achieved the best AUC value (Figure 7A), and the sensitivity at 99% specificity of the training set was obtained by adjusting the classification threshold (Figure S8D). In the validation set, COMOS sensitivity reached 88% (95% CI: 72−100%) and the specificity was 86% (95% CI: 50−100%) (Figure 7B,C and Figure S8E). Although the sensitivity and specificity of COMOS were the same as those of the FSR model and the FSR&CNA model, the AUC value was the highest, indicating that BSN, BSC, BSD and BSE improved the performance of COMOS in predicting the treatment response of DLBCL. The sensitivity, specificity and accuracy of COMOS showed a certain stability when the classification threshold was changed (Figure S8F).

In addition, the predictive performance of the integrating model was also compared with DLBCL prognostic indicators, such as LDH and β2MG (Figure 7D). In the validation set, 21 patients had LDH and β2MG data available (16 PR/CR, 5 PD/SD). The performance of COMOS and the two biomarkers was calculated and compared, showing an AUC of .95 for the integrating model, better than .85 and .76 for LDH and β2MG, respectively. Thus, COMOS achieved a much better predictive performance in comparison to the classical approach by integrating multiple fragmentation features.