Cohort characteristics and B cell phenotyping profile
We enrolled 150 well-characterized participants into our study and distributed them to the following groups: 1) Control: healthy individuals recruited from volunteers (n = 54); 2) cGVHD: individuals diagnosed using the criteria defined by National Institutes of Health (NIH) Consensus Conference (n = 71); and (3) non-GVHD: individuals who received HSCT but lacked any apparent clinical feature of cGVHD as defined by the NIH criteria (n = 25). These participants were divided into the three cohorts. In cohort1, five cGVHD patients and five healthy donors were enrolled for an exploratory study. In cohort2, 31 cGVHD patients, 15 non-GVHD patients, and 17 healthy donors were enrolled for a retrospective study. In cohort3, 35 cGVHD patients, 10 non-GVHD patients, and 32 healthy donors were enrolled for a further prospective study.
B cell phenotyping profiles of cGVHD patients, non-GVHD patients, and healthy controls were generated by spectral flow cytometry using a 20-marker panel that could accurately distinguish B cell subsets and indicate the developmental stage and activation status of B cells (Table S1). The classical B cell subsets distinguishable by this phenotyping panel included transitional B cells, naïve B cells, memory B cells (including IgM+, marginal zone-like, and switched memory B cells), IgD−CD27− double-negative B cells, and plasmablasts (Figure S1). The B cell phenotyping profiles were further explored to find a cGVHD-related B cell subset, build a cGVHD disease activity evaluation module, and figure out the threshold for cGVHD management through retrospective and prospective studies. Clinical information on the cGVHD and non-GVHD group members are presented in Tables 1, 2, and S1.
Table 1
Clinical characteristics of participants in Cohort 1
Characteristic | cGVHD, n = 5 |
Age, median (range), y | 35 (19–56) |
Gender, n(%) | |
Male | 4 |
Female | 1 |
Source of graft, no(%) | |
BMT | 0 |
PBSCT | 2 |
BMT + PBSCT | 2 |
PBSCT + UCBT | 1 |
Source of donor, no(%) | |
Sibling | 3 |
Unrelated | 0 |
Parents | 0 |
Children | 2 |
Relatives | 0 |
HLA matching, no(%) | |
Identical matched | 1 |
Non-identical matched | 2 |
Haploid matched | 2 |
GVHD prophylaxis, no(%) | |
CSA + MTX | 0 |
CSA + MTX + ATG | 0 |
CSA + MTX + MMF | 1 |
CSA + MTX + ATG + MMF | 4 |
Post-Cy | 0 |
Others | 0 |
Initial disease, no(%) | |
AML/AML from MDS | 3 |
ALL | 1 |
CML | 0 |
MDS | 0 |
Others | 1 |
Table 2
Clinical characteristics of participants in Cohort 2
cohort2 cGVHD patient profile | | cohort2 non-GVHD patient profile |
Characteristic | PR/CR/SD cGVHD, n = 26 | DP cGVHD, n = 5 | P-value | | Characteristic | Stable, n = 11 | Active, n = 4 | P-value |
Age, median (range), y | 30 (17–58) | 27 (17–52) | 0.4976 | | Age, median (range), y | 27 (16–61) | 39.5 (19–57) | 0.5810 |
Gender, No.(%) | | | > 0.99 | | Gender, n0(%) | | | 0.1033 |
Male | 21 (81%) | 4 (80%) | | | Male | 5 (45%) | 4 (100%) | |
Female | 5 (19%) | 1 (20%) | | | Female | 6 (55%) | 0 | |
Source of graft, No.(%) | | | 0.4129 | | Source of graft, No.(%) | | | > 0.99 |
BMT | 1 (4%) | 0 | | | BMT | 1 (4%) | 0 | |
PBSCT | 18 (69%) | 2 (40%) | | | PBSCT | 3 (27%) | 1 (25%) | |
BMT + PBSCT | 7 (27%) | 3 (60%) | | | BMT + PBSCT | 8 (73%) | 3 (75%) | |
PBSCT + UCBT | 0 | 0 | | | PBSCT + UCBT | 0 | 0 | |
Source of donor, No.(%) | | | 0.8704 | | Source of donor, No.(%) | | | 0.7363 |
Sibling | 13 (50%) | 3 (60%) | | | Sibling | 8 (73%) | 2 (50%) | |
Unrelated | 1 (4%) | 0 | | | Unrelated | 0 | 0 | |
Parents | 5 (19%) | 2 (40%) | | | Parents | 2 (18%) | 1 (25%) | |
Children | 2 (7.7%) | 0 | | | Children | 1 (9%) | 1 (25%) | |
Relatives | 2 (7.7%) | 0 | | | Relatives | 0 | 0 | |
others | 3 (11.6%) | 0 | | | others | 3 (11.5%) | 0 | |
HLA matching, No.(%) | | | 0.3892 | | HLA matching, No.(%) | | | > 0.99 |
Identical matched | 15 (58%) | 3 (60%) | | | Identical matched | 4 (36.5%) | 1 (25%) | |
Non-identical matched | 1 (4%) | 1 (20%) | | | Non-identical matched | 3 (27%) | 1 (25%) | |
Haploid matched | 10 (38%) | 1 (20%) | | | Haploid matched | 4 (36.5%) | 2 (50%) | |
GVHD prophylaxis, No.(%) | | | 0.8372 | | GVHD prophylaxis, No.(%) | | | 0.3626 |
CSA + MTX | 3 (11%) | 0 | | | CSA + MTX | 0 | 0 | |
CSA + MTX + ATG | 1 (4%) | 0 | | | CSA + MTX + ATG | 0 | 0 | |
CSA + MTX + MMF | 10 (38%) | 3 (60) | | | CSA + MTX + MMF | 3 (27%) | 0 | |
CSA + MTX + ATG + MMF | 6 (24%) | 1 (20%) | | | CSA + MTX + ATG + MMF | 6 (55%) | 4 (100%) | |
Post-Cy | 2 (8%) | 1 (20%) | | | Post-Cy | 2 (18%) | 0 | |
Others | 4 (15%) | 1 (7) | | | Others | 0 | 0 | |
Initial disease, No.(%) | | | 0.3699 | | Initial disease, No.(%) | | | > 0.99 |
AML/AML from MDS | 14 (54%) | 1 (20%) | | | AML/AML from MDS | 8 (73%) | 3 (75%) | |
ALL | 6 (23%) | 3 (60%) | | | ALL | 3 (27%) | 1 (25%) | |
CML | 0 | 0 | | | CML | 0 | 0 | |
MDS | 3 (11.5%) | 0 | | | MDS | 0 | 0 | |
Others | 3 (11.5%) | 1 (20%) | | | Others | 0 | 0 | |
Identification of a cGVHD-specific B cell subpopulation
Spectral flow cytometry data obtained from cohort1 (10 participants; 5 cGVHD patients and 5 healthy donors) were used to identify a cGVHD-specific B cell population. Uniform Manifold Approximation and Projection (UMAP) and Louvain clustering of cells 33,34 yielded three cell clusters (Clusters 0–2; Fig. 1A, 1B). From the cell composition of each cluster, Cluster 2 was found to be a cGVHD-specific cell cluster (Fig. 1C): cGVHD samples accounted for 82.0% of the Cluster 2 cells (1748 cells), while healthy donor samples accounted for only 18% (383 cells). To phenotypically characterize the cluster 2, we examined the cell surface phenotypes of each cluster. From among the 20 studied markers, CD20, CD268, CD5, CD27, CD38, CD86, CD269, and CD319 were identified as specific markers for Cluster 2 based on their expression levels (Fig. 1D).
We used these eight markers to generate an unbiased random forestry machine learning model and found that these markers could accurately label a disease-specific cell cluster. We then explored whether these eight markers could be further streamlined, with the goal of improving the clinical applicability of our method. We iterated all possible marker combinations and evaluated their accuracy. As presented in Fig. 1E, the accuracy scores indicated that certain combinations of only two or three markers could substitute for the full eight-marker panel with little loss of accuracy.
Optimal marker combination for identifying a cGVHD-specific B cell subpopulation
To identify a simple marker combination that could identify a cGVHD-specific B cell subset, we used an independent cohort (48 participants; 31cGVHD patients and 17 healthy donors) and marker combinations with accuracy scores > 0.9 (20 combinations; Fig. 2A). The three-marker combination of CD20/CD27/CD86 yielded the best AUC score (AUC score = 0.958, Fig. 2B) and thus appeared to be the optimal option.
The results of further cross-validation using flow cytometry indicated that this CD20/CD27/CD86 marker set could precisely identify a cGVHD-specific B cell subpopulation, as the CD27+CD86+CD20− B cells were extremely frequent in cGVHD patients and almost entirely absent from healthy donors (Fig. 2C). Thus, this three-marker combination appeared to successfully substitute for the original 20-marker panel. To further increase the clinical applicability of this marker set, we utilized its distribution to develop a module known as the cGVHD Progress Score (cGPS). This module aims to provide a single, easily interpretable score, serving as a direct reflection of the disparity between cGVHD and health. Using the formula presented in Fig. 2D, we calculated the cGPS for each participant in the cohort. The cGPS was significantly different between cGVHD patients and healthy donors (P = 0.00003, Fig. 2E).
Using a cGPS threshold to predict progression risk among non-GVHD patients
Recognizing or predicting the imminent onset of cGVHD at an early stage remains a significant challenge because non-GVHD patients do not present obvious signs (per the NIH consensus criteria) at the early stage of developing cGVHD 35–37. Hence, we tested whether cGPS could potentially predict progression of non-GVHD patients to cGVHD. Details on the non-GVHD patients enrolled into Cohort-2 are presented in Table 2. According to their progression (or lack thereof) to cGVHD within a 3-month period observation, the subjects were categorized as stable state (those remaining in non-GVHD status) and active state (patients that developed cGVHD). The patients of the two groups had similar baseline clinical characteristics (Table 2).
Our statistical analysis indicated that the cGPS values were significantly different between the stable and active groups (P = 0.0029). As shown in Fig. 3A, the cGPS of the patients in the stable group were largely < 1 whereas those of the patients in the active group were > 1. This suggested that a threshold cGPS approximate to 1 might potentially be used to distinguish between patients in the stable and active states. Our receiver operating characteristics (ROC) analysis revealed that the precise cGPS threshold for this cohort was 1.150; indeed, 100% of non-GVHD patients scoring below this threshold were in stable state (Fig. 3B). In addition to being able to distinguish the state of non-GVHD patients (Fig. 3C), this cGPS threshold could also be used to predict whether non-GVHD patients would become sick: As shown in Fig. 3D, four out of five (80%) non-GVHD patients with cGPS > 1.150 developed into cGVHD during the evaluation period, whereas all 10 non-GVHD patients with cGPS < 1.150 remained in stable state.
Using a cGPS threshold to estimate the treatment efficiency of cGVHD
It is difficult to quickly and efficiently estimate whether a cGVHD patient is likely to become disease-progressed38,39. As NIH consensus criteria for organ scoring was a follow-up evaluation system, cGVHD real time state monitor was hard to captured. Since the cGPS module was generated from the disease-related B cell subset, we questioned whether it might be able to distinguish disease-progressed cGVHD patients. To address this question, we classified the cGVHD patients of Cohort-2 as having stable disease (SD), partial response (PR), complete response (CR), or disease progression (DP), and collected them into two groups: the DP group and the non-DP group (PR/CR/SD groups). The baseline clinical characteristics of the two groups are shown in Table 2
As shown in Fig. 4A, the cGPS were significantly different between the DP and non-DP groups (P = 0.00029), with an apparent threshold around 1.5. Indeed, our ROC analysis revealed that a cGPS of 1.505 could be used to successfully separate the DP and non-DP groups (Fig. 4B). We further validated this finding by exploring the cGPS distribution between two groups, and found that a cGPS threshold of 1.505 could be used to clearly separate the two groups (Fig. 4C). No patient with cGPS < 1.505 entered the DP state, whereas five out of 11 (45.5%) patients with cGPS > 1.505 entered the DP state (Fig. 4D).
Using the above-identified cGPS thresholds in a prospective study
To validate the potential utility of the cGPS thresholds for cGVHD management in the clinic, we performed a prospective study in a new cohort of 77 participants (Cohort-3: 35 cGVHD patients, 10 non-GVHD patients, and 32 healthy donors). The cGVHD-specific B cell subpopulation of each participant was analyzed and cGPS values were generated (Figure S2). The cGPS thresholds were then used to divide the cGVHD and non-GVHD patients into high-risk and low-risk groups. Their clinical information is summarized in Table S1. At 3 months after the initial evaluation, the 10 non-GVHD patients were re-evaluated based on the NIH consensus criteria, and the results were compared to those generated by our patient classification strategy. As shown in Fig. 5B and 5C, four out of the four non-GVHD patients in high-risk group (with an initial cGPS > 1.150) were diagnosed with cGVHD at the second evaluation, while six of the six cases with cGPS < 1.150 (low-risk group) were still in stable state at the second evaluation. The evaluation accuracy of the cGPS module therefore reached 100% in predicting non-GVHD patients who would become disease-progressed.
We then assessed the ability of cGPS to predict the disease progression of cGVHD. As shown in Fig. 5D and 5E, all 24 patients in the low-risk group (with an initial cGPS < 1.505) were diagnosed as PR/CR/SD (i.e., non-DP) at the second evaluation. The accuracy of the cGPS module was thus 100% for predicting no disease progression. Of the eight patients in the high-risk group (with cGPS > 1.505), four (50%) were classified into the DP group at the second assessment, one was lost to follow-up, and one died of infection. The results of our prospective study therefore collectively indicated that the cGPS module may be a powerful tool for predicting the future course of disease for both cGVHD and non-GVHD patients in the clinic. This strategy should help doctors obtain timely and effective information that can be leveraged to improve patient care.
Overall summary
We herein leveraged previous reports that aberrant B-cell homeostasis could influence the progression of cGVHD to develop a new strategy for recognizing and predicting the disease. Using spectral flow cytometry in a small cohort, we successfully identified a cGVHD-specific subpopulation of B cells and delineated the defining markers of this subgroup. By applying machine learning in an independent cohort, we derived a simple marker combination and used it to develop an all-in-one scoring module named cGPS. ROC analysis was used to identify cGPS thresholds that can be used to evaluate disease progression in cGVHD and non-GVHD patients of this cohort. Finally, we performed a prospective study in a new cohort using these thresholds, and verified that our cGPS module could be a powerful tool for predicting the future course of disease among cGVHD and non-GVHD patients in the clinic.