Single-cell transcriptomics of multi-site cell therapy in osteoarthritis: Tissue-specific treatment correlations

Abstract

Knee-osteoarthritis (knee OA) is a prevalent joint disorder lacking Food and Drug Administration-approved cell therapies to halt progression. This study uses single-cell RNA sequencing to analyze bone marrow aspirate concentrate (BMAC) and stromal vascular fraction (SVF) samples in a clinical trial of autologous cell therapies. Trial site-specific variability was significant in BMAC, necessitating tailored normalization, whereas SVF was less affected, likely due to uniform subcutaneous fat sampling. Variance partitioning and tensor decomposition identified site effects in BMAC but revealed shared pathways across cell types in both tissues. Differential gene expression (DEG) analysis between responders and non-responders yielded no significant findings, although likelihood ratio test (LRT) revealed enrichment for DEG patterns linked to disease severity, potentially masked by patient heterogeneity. Key BMAC pathways included oxidative phosphorylation, unfolded protein response, and tumor necrosis factor alpha (TNF-α) signaling. Cell-cell communication analysis suggested enhanced human leukocyte antigen (HLA) signaling in non-responder MSCs (mesenchymal stromal cells), consistgenesent with inflammation, while responders showed more coordinated immune interactions. BMAC-MSCs promoted chondrocyte proliferation, whereas SVF-MSCs emphasized immune regulation. This study suggests that variability in therapy outcomes reflects patient heterogeneity beyond genomic factors, complicating the immediate use of genomic profiling to guide treatment. Nonetheless, as molecular pathways become better understood, integrating genomic insights into personalized strategies may become feasible.

Graphical abstract

This multi-site single-cell study reveals how baseline immune-stromal interactions in patient-derived BMAC and SVF products vary by tissue source and donor characteristics. Although clinical response signals were subtle, integrative analyses identified shared transcriptional programs and MSC-centered communication pathways that may inform future biomarker development and regenerative therapy optimization.

Introduction

Osteoarthritis (OA) is a degenerative joint disease that significantly diminishes the quality of life for millions of people worldwide. It is characterized primarily by debilitating joint pain and progressive cartilage loss.¹ Traditional treatment modalities, such as physical therapy, pharmacological interventions, and intra-articular corticosteroid injections (CSIs), provide symptomatic relief but often fail to arrest disease progression or restore damaged tissues effectively. In recent years, the emergence of regenerative medicine has improved the prospective therapeutic landscape for OA by introducing innovative cellular therapies designed to exploit the body’s inherent regenerative capabilities.² Among these therapies, autologous bone marrow aspirate concentrate (BMAC)³ and adipose-derived stromal vascular fraction (SVF)⁴ have gained prominence. BMAC is rich in stromal cells that aim to harness the regenerative and immunomodulatory properties of bone marrow-derived cells to tackle the multifaceted pathology of OA, which includes cartilage degradation, subchondral bone remodeling, and inflammation. Similarly, SVF,⁵ derived from adipose tissue, contains a heterogeneous mix of cells with potent anti-inflammatory and regenerative properties, offering a promising avenue for managing knee OA by mitigating inflammation and promoting repair within the joint.

Despite the promise and growing clinical application of these therapies, a thorough evaluation of their efficacy, particularly in comparison to other cellular therapies and standard of care, remains a critical unmet need.⁶ To address this, a comprehensive phase 2/3 multicenter, single-blind, randomized controlled clinical trial, MILES (NCT03818737), was undertaken. This trial aimed to assess the safety and efficacy of BMAC, SVF, and allogeneic human umbilical cord tissue (UCT)-derived mesenchymal stromal cells (MSCs) against conventional treatments for knee OA in a robust study design involving 480 patients.⁷ The trial’s methodology involved randomized allocation of patients into treatment arms receiving cellular therapies or corticosteroid treatments, with follow-up extending to 3 years. Central to this study was the innovative integration of tissue-specific single-cell RNA sequencing (scRNA-seq) analysis, which was employed to probe the molecular and cellular dynamics within bone marrow and adipose. All BMAC and SVF samples analyzed were collected at baseline, prior to treatment initiation, ensuring that associations reflect pre-treatment cellular states. All UCT-MSC scRNA-seq libraries represent allogeneic donor lots included solely for baseline cross-tissue MSC characterization and were not part of the clinical transplantation or response analyses.

Transcriptomic profiling was performed to determine whether there is a correlation between the cellular dynamics observed in treatments like BMAC and clinical outcomes such as pain relief and improved knee function. Association between molecular profiles and patient-reported outcomes enhances our understanding of how baseline tissue states relate to OA outcomes, potentially leading to more tailored and effective treatment strategies. Our previous study⁸ reported on the use of single-cell transcriptomics to analyze BMAC from OA patients, revealing significant alterations in key cell types, including MSCs, natural killer (NK) cells, monocytes, B cells, and T cells between OA patients and controls. In that analysis of 77 BMAC samples from the MILES cohort, we also observed enrichment of pathways related to PI3K-AKT-mTOR signaling, interferon (IFN) responses, senescence, and autophagy, with MSC-immune cell crosstalk emerging as a prominent feature. While informative, that study relied on a smaller subset of participants and external non-OA controls, limiting generalizability. The present study expands this work to 134 BMAC and 134 SVF samples from the MILES trial, enabling broader assessment of site variability, responder heterogeneity, and multicellular programs across tissues.

The current study expands on this work by incorporating additional BMAC profiles along with data from SVF from OA patients and exploring their relationship to therapeutic response. This study thus provides a more detailed characterization of the cellular environments and interactions within these tissues. Overall, the inter-individual variability of profiles obscures any major signatures of remission or disease progression, although we do identify gene expression features that may be enriched in some recipients with poor outcomes. Efforts to incorporate personalized genomic profiling into therapeutic decision-making will need to account for the patient variability. An overview of mILES trial design, patient cohorts, and sample inclusion for single-cell RNAseq analysis is shown in Figure 1.

Figure 1.

Experimental and analytical design and for scRNA-seq data

(A) MILES clinical trial data collection design: the trial consisted of 480 total participants, randomized into one of 3 arms. In each arm, patients received either cell therapy injection or CSI (control) to the knee. BMAC and SVF cell product consisted of autologous therapeutic cells collected from the patient’s bone marrow or adipose tissue, respectively. The UCT-MSC allogeneic cell therapy was manufactured from UCT, with 8 lots included. (B) Of the total enrolled patients in the study, samples from 134 patients from each of the BMAC and SVF arms and 8 samples from UCT-MSC lots were analyzed using 10x Genomics 3′ scRNAseq kit V3.1. All the autologous data from BMAC and SVF are collected pre-treatment. For BMAC, among 134, there are 32 patients treated with CSI, 11 patients dropped out, 29 patients were grouped as non-responders as of VAS responsiveness, and 62 patients were responders as of the VAS responsiveness after 12 months’ follow-up. For SVF, among 134, there are 24 patients treated with CSI, 30 patients dropped out, 33 patients were grouped as non-responders as of VAS responsiveness, and 47 patients were responders as of the VAS responsiveness after 12 months’ follow-up. For UCT-MSCs being allogeneic, there were data generated from 8 donor samples who are not OA patients. (C) scRNA-seq analysis workflow. Major steps included quality control and normalization, integration, exploratory analysis (variance partitioning, cell proportions, differential expression, cell-cell communication), and baseline correlates of treatment response. (D) Timeline of trial events. Baseline sample collection occurred prior to treatment, followed by randomization and injection (BMAC, SVF, UCT-MSC, or CSI). MRI follow-up occurred at 6 and 12 months in the parent trial. For this analysis, we had access only to treatment response classification (responder vs. non-responder based on pain score), not MRI or other secondary clinical endpoints.

Results

Cluster analysis reveals distinct cell type compositions in BMAC and SVF datasets

Cluster analysis of the BMAC and SVF datasets revealed 18 and 19 distinct cell-type clusters, respectively (Figures 2A and 2B). In the BMAC dataset, major cell types identified included immune cells including monocytes, erythrocytes, megakaryocytes (MEGA), MSCs, and hematopoietic stem cells, with MSCs comprising a lower fraction of the BMAC dataset (543 cells of 224,976; ∼0.24%) after filtering and quality control, compared with 14,079 MSCs in SVF (5.1% of cells). In the SVF dataset, predominant cell types included adipose-derived stromal cells (ASCs), epithelial cells, fibroblasts, macrophages, and MSCs, which constituted approximately 5% of the total cell population. The distinction between ASCs and MSCs was based on marker expression: ASCs exhibited high CD34 and human leukocyte antigen (HLA)-DR expression but low CD90, while MSCs were characterized by high CD90 and low HLA-DR expression. Following extensive comparison of different procedures for integration of datasets obtained from the different collection sites, we settled on reciprocal principal-component analysis (rPCA) for all subsequent analyses. As with all scRNA-seq analyses, the final cluster assignment represents a balance between consistency across datasets, noise reduction, and matching cell types to biological expectation. Erythrocytes were removed from the final clustering as their proportion is highly variable and obscures trends in proportions of the other cell types.

Figure 2.

UMAPs and cellular compositions for scRNA-seq cell type clusters for BMAC and SVF datasets

(A) Cell types identified for BMAC datasets showing cellular composition; 18 large cell type clusters are identified. The erythrocyte cluster was excluded to reduce the noise for BMAC. Large immune cell types and monocytes are identified. (B) Cell types identified for SVF datasets showing the cellular composition of SVF as a product. Nineteen large cell type clusters are identified; among these, immune cell, epithelial cell, macrophages, adipose-derived stromal cells, and fibroblasts are identified. (C) BMAC cell composition barplot aligned with clinical collection site. (D) SVF cell composition barplot aligned with clinical collection site. dittoSeq package is used for the barplots, and Propeller (proportions per donor) tool is used for the statistical analysis on the cell proportions to investigate any significant correlation between cell types and the metadata like sex, site, age, individuals, and treatment responses.

Cell abundance analysis

Inferred cell type proportions per donor group by collection site are shown in Figures 2C and 2D for BMAC and SVF, respectively. Using the propeller tool (speckle R package) to evaluate the variability of cell type abundance, we calculated proportions on a per-donor basis; BaselineProp represents the mean donor-level proportion across all samples, while PropMean.1–PropMean.5 represent site-specific averages. In BMAC, significant site-to-site variability was detected for several immune and stromal lineages (CD14 monocytes, macrophages, pre-B, plasma B, MEGA, and MSC; one-way ANOVA false discovery rate (FDR) <0.05; Table S1). These effects are illustrated in Figure S9, which shows donor-level proportions and site means, with significant cell types indicated by an asterisk in the panel titles. This pattern is most likely driven by subtle differences in bone marrow sampling by different physicians. Since pro-B and mature B cells, dendritic cells, MSCs, as well as naive CD4 and CD8 T cells were additionally significantly variable among sites when using Harmony, we regarded this method as less robust but show the proportions in Figure S1. In contrast, variability across normalization methods was not observed for the SVF datasets shown in Figure 2D. After controlling for collection site, age, sex, body mass index (BMI), and Kellgren-Lawrence (KL) disease severity index, no cell type proportions varied significantly between responders and non-responders (Tables S1–S4). We also examined donor covariates including age, sex, and BMI in relation to baseline BMAC cell type proportions. Across 134 donors, associations were modest and lineage specific, with no consistent directional trends (Figure S11). This indicates that demographic covariates did not drive the major inter-site differences observed in cell composition.

Variance partition analysis for BMAC and SVF

Turning our attention to gene expression within cell types, dreamlet was used to perform variance partition analysis of the influence of collection site (Figure 3A) and therapeutic response status (Figure 3B). The violin plots show the range of variance explained by these factors for each expressed gene with the main effect shown in brown and residual unexplained variance in gray. These results imply a substantial influence of collection site on BMAC gene expression variability, with a median between 10% and 25% in most BMAC cell types, for some genes over 50% of the variability due to site, most notably in MSCs and plasma B cells. Only CD16-positive monocytes showed an elevated collection site influence in the SVF dataset. We, therefore, interpret site-level donor differences with caution. To mitigate this effect, we applied rPCA (integration) for clustering and ComBat-Seq batch correction for pseudobulk analyses, which consistently revealed enrichment for oxidative phosphorylation, unfolded protein response (UPR), and tumor necrosis factor alpha (TNF-α) pathways across BMAC donors. By contrast, SVF was less affected by site variability and supported immune-regulatory signals with greater robustness. This variation itself is informative, underscoring the need to account for collection-site effects when designing multi-site trials and interpreting single-cell data in a clinical context.

Figure 3.

Assessment of site variance impact on gene expression in two separate datasets

Assessment of site- and response-related variance impacts on gene expression across two datasets. (A) This panel depicts the impact of the site variable on gene expression across multiple cell types within two independent datasets, visualized side by side for comparison (left: BMAC; right: SVF). Violin plots represent the distribution of variance percentages explained by site within each cell type. Black dots indicate the median variance explained by site, and vertical lines denote the interquartile range. The distribution and spread of the violin plots convey the degree to which site-level effects contribute to gene expression variability across cell types. (B) This panel shows the variance in gene expression explained by responder versus non-responder status across the same cell types in BMAC (left) and SVF (right). Violin plots summarize the distribution of variance explained by response status, highlighting comparatively smaller contributions relative to site-level effects. Variance partitioning was performed using the variancePartition framework from the dreamlet package, which applies linear mixed-effects models to quantify the fractional contribution of covariates to overall gene expression variability.

In strong contrast, treatment response contributed minimally to the variance in both BMAC and SVF datasets, explaining less than 10%–15% of the variability of all genes across most cell types. Interestingly, the only exception was MSCs in the BMAC with approximately 20% of the variance in gene expression associated with treatment response, although this result is tempered by the relatively small counts of MSC in that tissue compartment. There was also a slightly elevated response of ASC/MSC and CD16 monocyte gene expression in the SVF.

Identification of multicellular gene expression programs across diverse cell types using scITD in OA SVF: Insights into donor-specific variabilities and immune-stromal interactions

We next performed tensor decomposition to explore how gene expression programs shared across multiple cell types associate with biological or technical aspects of the study. For this, we used single-cell tensor decomposition (scITD)⁹ software, which provided the most informative results for gene expression programs associated with immune and stromal cell interactions in the SVF dataset. Applying scITD to SVF yielded five latent factors, defined as components of donor-level variation that capture coordinated gene expression programs across multiple cell types. Latent factors identified by this approach can be considered as analogous to principal components of variation shared across multiple cell types. None of these factors were associated with Condition after FDR correction, consistent with the modest responder signal observed in variance partitioning. Here, “Condition” refers to clinical responder vs. non-responder status, which is determined by 12-month WOMAC (Western Ontario and McMaster Universities Osteoarthritis Index) improvement following BMAC or SVF treatment. These clinical outcome labels are independent of the sequencing data and allow us to test whether baseline gene expression programs relate to later treatment response. Factor loadings, nevertheless, highlighted coordinated immune-stromal programs (Figure S2). The software first identifies components of variation shared across cell types and shows how these covary with collection site or treatment response (Figure 4A) and then allows visualization of the expression of key transcripts associated with the components in the contributing cell types. Factors 2 and 4 weakly associated with site, but none of the first five factors associated with response. Nevertheless, Factor 1 exhibited a strong association with specific immune cell types (Figure 4B), notably antibody-secreting cells, B cells, CD4 T cells, NK cells, macrophages, and monocytes. Upregulated genes associated with Factor 1 included FBN1, IGFBP5, and MFAP5, all potentially engaged in immune-stromal interactions, while TGFBR3 and PRG4 were downregulated in these cell types. Gene set enrichment analysis (GSEA) of the Factor 1 loadings further supported stromal cell enrichment for pathways related to innate immune response, B cell-mediated immunity, and T cell activation. In parallel, pathways linked to cytokine signaling, antigen presentation, and immune effector functions were highlighted in the immune cells in the SVF. Loadings of other gene sets and cell types for Factors 2 through 5 are shown in Figure S2.

Figure 4.

Tensor factorization of gene expression profiling in SVF

(A) The heatmap depicts Z score-normalized expression data across multiple donors, stratified by site and condition, with sidebars indicating sample origin (Site) and clinical outcome (Condition). Red shades indicate higher factor levels, and blue shades represent lower factor levels relative to the mean. (B) The heatmap illustrates the association of various genes with “Factor 1” across different cell types. Red intensity signifies a stronger association. Listed genes demonstrate a pronounced association with Factor 1, suggesting a potential biological significance in these specific cell types.

Pseudobulk analysis of BMAC shows enrichment of DEGs correlated with clinical response

Donor-level single cell interpretable tensor decomposition (scITD) confirmed that site is the dominant axis of donor variation in BMAC. Factor 1 explained 41.5% of variance and was significantly associated with Site (p = 5.62 × 10⁻⁵; q = 2.81 × 10⁻⁴, Benjamini-Hochberg [BH] ), whereas no factor associated with Condition after FDR correction (all q ≥ 0.16). These results are shown in Figures S6A–S6C and Table S5. Reflecting the above-documented strong site effect, scITD Factor 1 for BMAC was strongly associated with collection site. Since the remaining BMAC factors did not clearly associate with a biological covariate, an alternative analytical approach was taken to better address whether response to therapy correlates with differentially expressed genes. Single-cell profiles were converted to pseudobulk profiles for each cell type in each donor. Responder and non-responder categories reflect patients’ clinical pain response after treatment, not differences in sequencing or cell product. We used these clinical labels to evaluate whether baseline transcriptomic signatures in BMAC relate to therapeutic outcome. Since there was no significant differential gene expression between responders and non-responders below the 5% FDR threshold in any of the individual cell types after standard hypothesis testing, we utilized the likelihood ratio test (LRT) option in DESeq2 (model comparison to detect patterns) to contrast a null model of no differences among response groups with alternate models. We thus searched for genes whose profile in pseudobulk cells representing each cell type in all individuals was more likely to fit a specific pattern of differential expression with respect to treatment response. Figure 5A shows that in all 18 BMAC cell types, by far the most supported model showed a gradient from lowest expression in samples from individuals who did not complete the study to highest in responders, who generally resemble control. The number of genes following this pattern ranged from 2,923 in mature B cells (and similarly in the major T cell subtypes) to fewer than 100 in MSC and macrophages (reflecting the small number of such cells).

Figure 5.

Differential gene expression analysis across treatment groups using DESeq2 LRT

(A) The plots visualize the Z score of gene abundance for groups of DEGs clustered by similarity in their expression trajectories. Each subplot represents one such cluster, labeled by “Group” number and the number of genes it contains. The categories on the x axis denote the treatment response groups: control, incomplete responder, non-responder, and responder. Points represent individual genes, and the lines connect the median Z scores across treatment conditions, summarizing expression trajectories. Boxplots overlay the data points to show within-condition distributions. Note that “Group 1, 2, 3” refer to DEG clusters with distinct expression patterns, not treatment groups. The red color highlights one DEG cluster category, as indicated in the legend. (B) The same clustering pattern after ComBat-seq adjustment for site variability. (C) Heatmap of pathway enrichment comparing non-responders versus responders (hallmark GSEA) of the BMAC samples. Rows correspond to pathways and columns to cell types. Colors indicate t statistics for the non-responder minus responder contrast (red, higher in non-responders; blue, higher in responders). Asterisks denote pathways reaching FDR <0.10. Non-responders show relative enrichment of oxidative phosphorylation, UPR, and TNF-α signaling via nuclear factor (NF)-κB, whereas responders show enrichment of KRAS signaling and hypoxia.

Further adjustment was performed using ComBat-seq to remove the batch effect of collection site on the pseudobulk expression profiles. This considerably reduced the numbers of significant genes as much as two orders of magnitude; yet around 100 genes with this pattern were retained, with examples shown in Figure 5B. GSEA of these DEGs identified key biological pathways impacted by the treatment. GSEA of these DEGs showed pathways higher in non-responders, oxidative phosphorylation, UPR, and TNF-α signaling via NF-κB, while responders showed higher KRAS signaling,^10,11 a RAS-MAPK pathway that regulates cellular proliferation, stress responses, and inflammatory gene programs and hypoxia (Figure 5C).

Restricting the analysis to the responder and non-responder categories shows that the lack of significant differential gene expression between groups may be due to the high variability within groups. This is clearly illustrated by the dot plots for the ten most differentially expressed genes in four cell types in Figure 6. Across cell types, no responder vs. non-responder differences reached the 5% FDR threshold. However, LRTs suggested enrichment of pathways including oxidative phosphorylation, UPR, and TNF-α signaling. These findings are exploratory given subgroup variability, but they point to potentially relevant biological processes warranting further investigation. The size of each dot represents the proportion of cells for the donor who expresses the transcript, and the shading represents the average gene expression. We chose 10 individuals of each class at random from across the collection sites and grouped them by response group, but visually it is clear that there is little correspondence between response status and expression. The four selected cell types—CD16 monocytes, pro-B cells, plasma B cells, and MSCs, display variation in gene expression without clear separation between responders and non-responders. In CD16 monocytes, for example, while most of the 10 genes are co-expressed at a higher level in non-responders, one responder (labeled 04bmac039) is among the high expressors, while others have elevated expression of one or a few of the genes. A similar pattern of high inter-individual variability without clear separation between responders and non-responders is observed across the other cell types, with MSCs showing particularly variable expression of multiple ribosomal protein–encoding genes. These results indicate that gene expression variation does not distinctly separate responders from non-responders. To assess whether limited power could mask group differences, we performed repeated random downsampling (10 responders vs. 10 non-responders; 30 iterations) and re-fit the pseudobulk models for four representative cell types (CD16 Mono, MSC, plasma B, pro-B). Across iterations, the number of FDR <0.10 DEGs was zero in the majority of runs, with only sporadic small counts that were not reproducible; nominal hits (p < 0.05) were more common but variable (Figures S10A and S10B; Table S7). These resampling results corroborate that baseline responder/non-responder differences are subtle and unstable to resampling, consistent with the variance partitioning and scITD findings. Equivalent contrasts in SVF did not identify statistically significant DEGs or consistent patterns across downsamplings, so Figure 6 focuses on BMAC where sufficient donor numbers enabled formal resampling validation.

Figure 6.

Expression variability of selected genes in BMAC cell types across random batches

This figure illustrates the expression variability of the top 10 most differentially expressed genes (DEGs) in selected BMAC cell types across randomly assigned batches of 20 samples. We randomly selected 10 responders and 10 non-responders across sites; dot size = % of cells expressing the gene; dot color = average expression. Each dot represents a gene (y axis) within a specific cell type (x axis). Dot size reflects the percentage of cells expressing the gene, and color intensity denotes the average expression level, with red indicating higher expression and blue indicating lower expression. (A–D) The selected BMAC cell types display expression diversity for the most differentially expressed 10 genes, highlighting variability in gene expression patterns across sample batches.

Cell-to-cell communication analysis in BMAC and SVF

Cell-to-cell communication analysis was conducted to ask whether there is potential for interactions between MSCs and other cell types to differ between responders and non-responders. Using the CellChat tool, which infers likely receptor-ligand interaction networks from scRNA-seq data, a rich network of interactions was observed, with MSCs potentially capable of establishing communicative links with every cell type analyzed and significant autocrine interactions noted in every cell type except macrophages (Figure S3A). Focusing on the top 10 differentially expressed ligand-receptor interactions (Figure S3B), it was found that MSCs were pivotal predicted communicators in responder-derived BMAC cells. The relative information flow in Figure S3C suggests that MSCs in responder samples exhibited interactions enriched for ligands such as THY1, CD40, and IGF1, which are associated with immune modulation and tissue repair. By contrast, non-responder samples seem to be enriched for ligands such as IGFBP and CSF that are more typically linked to inflammatory and cellular stress responses.

A similar analysis was performed on the SVF single-cell RNA-seq dataset to compare predicted cell communication differences between responders and non-responders at baseline. In this analysis, ASCs, MSCs, and M2 macrophages were defined as sender populations, while the remaining immune and stromal cell types were treated as receivers. This revealed ASCs, MSCs, and M2 macrophages as the most prominent signal-sender populations in SVF (Figure 7A). The identities of the ASC and MSC populations, shown by expression of key markers in Figure 7B, were used to separate ASCs and MSCs based on distinct expression profiles. MSCs exhibited higher expression of THY1 (CD90) and ENG, while ASCs were characterized by reduced CD34, HLA-DRA, and HLA-DRB1 expression.

Figure 7.

Interaction of ligand activity with gene expression and predicted regulatory potential in SVF

(A) Chord diagram of predicted cell-cell communication in SVF. Each node represents a cell type, and connecting edges visualize signaling pathways, with edge thickness reflecting interaction strength. (B) Violin plots showing ASC- and MSC-specific marker genes used to distinguish ASCs from MSCs. (C) Prioritized ligand activity across ASCs, MSCs, and M2 macrophages. Panels are arranged side by side for ASCs (left triptych), MSCs (middle), and M2 macrophages (right). Dot size indicates the percentage of sender cells expressing each ligand, and dot color reflects log2 fold-change (responders−non-responders), with red indicating higher expression in responders and blue indicating lower expression. Ligand activity scores were computed with NicheNet as the correlation between each ligand’s target predictions and observed differential expression; AUPR (area under the precision-recall curve) summarizes target-prediction ability. (Right) Heatmap of normalized expression levels for the same ligands within sender populations. Red, higher expression; blue, lower expression (log2 fold-change scale). Columns are predicted target genes, with purple shading indicating regulatory potential. Darker purple, stronger ligand-target association, as measured by AUPR. The full, unfiltered ligand panel is provided in Figure S4.

Focusing on specific ligand-receptor interactions and target gene predictions, several ligands enriched in the responder group were identified, both in terms of the number of interactions and their relative strengths. The dot plot in Figure 7C illustrates differences in prioritized ligand expression across ASCs, MSCs, and M2 macrophages. Dot size represents the proportion of sender cells expressing each ligand, and color reflects the log2 fold-change between responders and non-responders. Panels are arranged side by side for ASCs (left triptych), MSCs (middle), and M2 macrophages (right). Notably, FGF7 was significantly upregulated in MSCs but downregulated in ASCs, while interleukin (IL)-15 showed the opposite pattern (downregulated in MSCs and upregulated in ASCs). Other significant ligands enriched in MSCs included COL5A2, ANGPT1, POSTN, LAMA4, and S100A4, while ligands such as F8, NRG2, and VWF were enriched in ASCs.

The heatmap in Figure 7C (right) highlights normalized expression patterns of the same ligands within each sender population and their predicted regulatory potential. Here, M2 macrophages displayed distinct ligand expression profiles compared to ASCs and MSCs. IFITM1 emerged as a significantly upregulated ligand in M2 macrophages, whereas its expression remained relatively lower in the other two cell types. Similarly, HBEGF, TGFB2, and FGF7 showed increased expression in M2 macrophages, reinforcing their potential role in modulating the responder group’s microenvironment. In contrast, ligands such as CD86, BACE2, and VWF were more prominently expressed in ASCs and MSCs, suggesting differential contributions to the signaling network. This extended analysis underscores the nuanced cellular interactions within SVF and refines our understanding of how distinct cell populations contribute to the observed communication dynamics.

To provide context across MSC tissue sources, we also profiled UCT-MSCs. Because these represent allogeneic donor-derived lots rather than autologous patient samples, they were analyzed descriptively and are presented in Figure S7.

Discussion

Our study provides a comprehensive transcriptomic analysis of BMAC and SVF samples in the context of OA therapy. By analyzing cellular diversity and molecular expression patterns, our purpose was to identify key factors influencing treatment responses. We explored cell composition, gene expression variability, and cell communication networks, building on our previous work with a focus on treatment response variability.

Our findings reveal marked study site-influenced variability in BMAC, with significant differences across five study sites in the United States affecting both cell composition and gene expression. This variability was more pronounced in BMAC than SVF, likely due to the heterogeneity incurred by the technique of needle aspiration from the bone marrow compared with more uniform sampling from subcutaneous fat. Operational variability may be unavoidable in real-world settings, but this finding nevertheless emphasizes the importance of testing and optimizing normalization techniques to enhance the likelihood that observed patterns reflect biological variability rather than procedural artifacts. We applied a range of normalization methods, including Harmony and rPCA, to adjust for site-related effects while preserving biological variability, showing how application of normalization techniques can potentially influence data interpretation. Importantly, the marked site-specific variation observed in BMAC is itself an informative finding. It highlights the critical role of collection practices in shaping single-cell readouts and underscores the need for harmonization in multi-site clinical trials. The extent of donor-to-donor variability across clinical sites is further evident in Figure S9, where several BMAC cell types show pronounced site-to-site shifts in mean proportions. Recognizing this variability is essential not only for data interpretation but also for designing standardized protocols and guiding clinical decision-making in future cell therapy studies. Consistent with these observations, single cell interpretable tensor decomposition (scITD) confirmed that inter-site heterogeneity was the dominant axis of donor-level variation in BMAC. These latent factors, defined as multicellular gene expression programs spanning multiple cell types, capture coordinated donor variability. Factor 1 explained 41.5% of variance and was significantly associated with Site (q = 2.8 × 10⁻⁴), while no factors were associated with Condition after FDR correction (Figure S6; Table S5). This underscores that technical heterogeneity across sites can outweigh biological response effects in BMAC single-cell data.

Overall, there was no significant enrichment of differential gene expression between responders and non-responders after adjusting for the multiple cell types and number of genes considered. In BMAC, scITD further supported this, indicating that no latent factor aligned with Condition after FDR correction and reinforcing that treatment-response signatures are modest relative to site-level effects. Our current findings also align with our earlier study of 77 BMAC samples,⁸ where we reported that OA patients exhibited enrichment of pathways such as PI3K-AKT-mTOR signaling, IFN responses, senescence, and autophagy compared with non-OA controls. The recurrence of related stress- and immune-regulatory programs in the present responder/non-responder analysis underscores that these pathways are reproducible features of the osteoarthritic bone marrow microenvironment, even though they remain modest in magnitude relative to site-level effects. This reinforces that treatment-response signatures are modest relative to site-level effects. Nevertheless, exploratory analyses identified shared pathways of co-regulated genes across cell types in both BMAC and SVF. While responder versus non-responder contrasts did not yield statistically significant DEGs at a stringent FDR cutoff, these exploratory analyses revealed consistent enrichment of pathways such as oxidative phosphorylation, UPR, and TNF-α signaling. We present these results as hypothesis-generating rather than definitive, highlighting processes that may contribute to differential response and that merit confirmation in larger cohorts. In BMAC, application of a likelihood ratio test (LRT) provided preliminary evidence that several pathways may be associated with treatment response, associations that are likely to be clarified with larger sample sizes. These included oxidative phosphorylation,¹² UPR,¹³ and TNF-α signaling,¹⁴ suggestive of the initiation of a senescent,¹⁵ pathophysiological environment,¹⁶ related to cellular stress¹⁷ and immune regulation.¹⁸ These stress-associated pathways, particularly TNF-α signaling, KRAS signaling, and oxidative phosphorylation have been independently implicated in OA pathogenesis and immune-metabolic imbalance. TNF-α signaling contributes to chronic inflammatory remodeling of the synovium and cartilage through immune-stromal crosstalk.¹⁰ KRAS pathway activity has been linked to chondrocyte stress adaptation and inflammatory signaling loops that promote joint degeneration.¹¹ Oxidative phosphorylation, which sustains chondrocyte energy metabolism, has been shown to mitigate cartilage degradation when experimentally enhanced in OA mouse models.¹² Collectively, these findings reinforce that the multicellular programs captured by scITD in our study are not merely technical artifacts but reflect conserved OA-associated immune-metabolic mechanisms that may influence therapeutic response. Notably, the UPR, which is linked to endoplasmic reticulum stress, has been implicated in OA,¹⁹ providing a potential mechanistic connection to disease progression. These findings suggest a heightened cellular stress response and metabolic dysregulation in non-responders, possibly contributing to the diminished therapeutic efficacy observed in these patients.

In contrast, SVF samples showed distinct immune regulatory profiles,^20,21 with stromal cells exhibiting the potential to activate pathways related to innate immune responses and T cell activation,²⁰ while immune cells being enriched in cytokine signaling and antigen presentation pathways. These differences suggest that while BMAC cells may primarily modulate metabolic and inflammatory responses, SVF cells may have more potential to contribute to immune modulation. This dual role of metabolic dysregulation in BMAC and immune regulation^22,23,24 in SVF underscores the complexity of OA pathophysiology and highlights the need to tailor cell-based therapies based on these distinct molecular profiles. The pathway distinctions observed between responders and non-responders in BMAC and the immune-modulatory roles of SVF further emphasize the importance of understanding cellular contributions to therapeutic responses in OA.²⁵

The reason why the standard differential expression analysis did not reveal any statistically significant genes (after FDR correction) between responders and non-responders in either BMAC or SVF, whereas the LRT showed a strong pattern of enrichment, has to do with the combination of individual variability and study site confounding. The subset of individuals who failed to complete the study included many who dropped out due to non-response, explaining why their baseline profiles appear to be even more deviant from the healthy controls than the included non-responders. We expect that larger sample sizes will confirm the potential markers of non-response, and better address how to avoid study-site differences.

Cell-cell communication analysis further underscored the potential therapeutic roles of MSCs within BMAC and SVF samples. In BMAC, MSCs were identified as major communicators, interacting with various immune cells. Notably, MSCs from responders showed enriched ligand interactions with THY1,²⁶ CD40, and IGF1,²⁷ which are linked to immune modulation and tissue repair, whereas MSCs from non-responders exhibited elevated HLA signaling,²⁸ indicative of an inflammatory and possibly an autoimmune profile. This suggests that MSC-mediated signaling may be more coordinated in some responders, potentially contributing to treatment efficacy through enhanced immune cell interactions. In SVF, both MSCs and ASCs demonstrated active communication networks, particularly in responders. Ligands such as IL15, FGF7, and TNFSF12 showed varying expression patterns,^29,30,31 with FGF7 upregulated in MSCs and downregulated in ASCs,³² indicating nuanced differences in cellular interactions that could affect immune responses and tissue repair in OA.^33,34 All of these cell communication inferences are subject to inter-individual variability.

These findings, together with scITD, underscore that cell therapy outcomes in OA are shaped by both patient-specific factors and site-dependent variability. In BMAC, scITD reinforced the dominance of inter-site heterogeneity, with no latent factors associated with Condition after FDR correction. MSCs were less abundant in BMAC (∼0.24% of all cells) compared with SVF (∼5.1%), which inevitably leaves pathway analysis for this key cell type underpowered; yet the cell-type-specific analyses, nevertheless, show enrichment of oxidative phosphorylation, UPR, and TNF-α pathways despite the minimal contribution to overall BMAC expression. MSCs likely influence the osteoarthritic microenvironment through their influence on other cell types as indicated by the inferred ligand-receptor interactions, despite their low frequencies. Ex vivo manipulations and organoid studies may illuminate the regulatory functions of MSC further.

Thus, while site variability constrained some donor-level comparisons in BMAC, it provides a valuable translational insight: without proper normalization, clinical studies risk misattributing procedural differences as biological effects. By documenting this variability and showing that pathways such as oxidative phosphorylation, UPR, and TNF-α signaling reproducibly emerge after correction, our work offers a framework for balancing technical artifacts with physiological interpretation.

Overall, this study highlights the importance of accounting for site variability in BMAC samples, the distinct roles of MSCs in BMAC and SVF, and the potential of integrating scRNA-seq and cell communication analyses to refine therapeutic strategies. Our findings contribute to a growing understanding of OA’s cellular dynamics and suggest that personalized cell therapy may require a more nuanced approach, considering both molecular signatures and sample-specific factors to improve treatment outcomes.

Materials and methods

Study design and participants

The MILES³⁵ study was a multicenter single-blinded clinical trial performed in accordance with guidelines and oversight from the Food and Drug Administration (FDA), and under the management of a contracted research organization. Study approval was obtained from the Western Institutional Review Board and by the institutional review boards of Duke and Emory Universities. In accordance with the FDA requirements, an investigational new drug application was filed (#18414), referencing investigative device exemption #17894. The study was also registered with⁷ ClinicalTrials.gov (NCT03818737). A total of 570 patients were screened to identify 480 eligible patients who were randomized at five clinical sites across five states within the United States. The five clinical sites included clinics from Emory University (GA), Sanford Health (SD and ND), the Andrews Institute (FL), and Duke University (NC). Participants were randomized to a four-cohort parallel-design study. Participants were further subdivided into three arms, each of which was matched with one-third (n = 40) of the CSI controls (Figure 1A): arm 1, autologous bone marrow concentrate (BMAC) (n = 120); arm 2, autologous SVF (n = 120); arm 3, allogeneic UCT (n = 120). The control cohort of CSI was subsequently aggregated to allow a 1:1:1:1 comparison for analysis. Subjects were enrolled if they were between 40 and 70 years of age and carried a diagnosis of knee OA as determined by radiographs within 3 months of their clinical visit, also having a KL OA severity score between 2 and 4. A full list of eligibility criteria is provided on ClinicalTrials.gov (NCT03818737). In the BMAC arm, a total of 12 subjects did not complete the study: 1 related to release criteria, 10 related to subject withdrawal, and 1 lost to follow-up. In the SVF cohort, 12 subjects did not complete the study: 1 due to investigator withdrawal, 4 related to release criteria, 6 related to subject withdrawal, and 1 due to screen failure. In the UCT cohort, 13 subjects did not complete the study: 2 due to investigator withdrawal, 10 related to subject withdrawal, and 1 lost to follow-up. In the CSI cohort, 11 subjects did not complete the study: 1 due to investigator decision, 6 related to subject withdrawal, and 4 lost to follow-up. Of the remaining patients, samples from 134 in each of the BMAC and SVF arms met minimum criteria for additional processing and were analyzed for scRNA-seq (Figure 1B). In addition, scRNA-seq libraries were generated from eight allogeneic UCT-MSC donor lots prepared under Uniform Manifold Approximation and Projection (cGMP) conditions. These UCT-MSC samples were included for cross-tissue MSC comparisons; however, because they were not autologous or patient-matched, they were not incorporated into responder versus non-responder analyses. Although UCT-MSCs were administered clinically in the parent MILES trial, the scRNA-seq data generated here are from allogeneic donor lots rather than patient-derived cells and therefore cannot be linked to individual clinical outcomes. A total of 40 participants discontinued participation prior to the 12-month endpoint, as reported in the corresponding Nature Medicine publication. Responder versus non-responder classification was determined in the parent MILES trial based on the co-primary endpoints of pain improvement at 12 months, measured by the visual analog scale (VAS) and Knee injury and Osteoarthritis Outcome Score (KOOS)-Pain subscale, as previously reported (Centeno et al., Nat. Med. 2023). Patients returned to the clinic for magnetic resonance imaging (MRI) to assess cartilage and joint health³⁶ at baseline, 6 months, and 1 year.

Randomization and masking

Subjects were stratified by clinical center, and after informed consent was obtained and eligibility criteria verified, they were randomized to treatment groups.³⁷ Treatment assignments were also stratified by clinical center and generated using a pseudo-random number generator with randomly permuted blocks. These assignments were stored in the Medidata Rave cloud-based data management system developed by the contracted research organization. All subjects underwent the harvesting procedure per their assigned study arm, and then, according to the randomization scheme, they were blinded to the actual treatment received (e.g., SVF versus CSI, BMAC versus CSI, or cells versus CSI). As a single-blinded study, the site principal investigators were not required to be blinded; however, physicians and subjects were blinded to the injection. Blinding was maintained by using opaque coverings to prevent visualization of syringe contents, and by performing sham BMAC and SVF harvest procedures for patients in designated cohorts.

scRNA sequencing analysis

To promote consistency, standardized protocols were followed across the five study sites for cellular harvesting, product preparation, injection procedures, and CSIs. While BMAC and SVF samples were fresh autologous products, MSCs derived from UCT had been cryopreserved and prepared to cGMP standards at Duke University. In our earlier publication,⁸ 77 BMAC samples from the MILES cohort were used to compare healthy and disease conditions; an additional 57 samples in this study expanded the total to 134, following the same protocols. Single-cell suspensions from cryopreserved BMAC and SVF samples were prepared and assessed using the NC-200 Nucleocounter cell counter. Transcriptome analysis was conducted on both the BMAC and SVF cells, as well as the allogeneic UCT-MSC samples. Library preparation and sequencing were performed using the 10x Genomics Chromium platform with the 3′ V3.1 kit on an Illumina Novaseq 6000. BMAC samples were sequenced across four NovaSeq batches and two NextSeq batches, SVF samples across four NovaSeq batches, and UCT-MSC samples in a single NovaSeq batch. Importantly, libraries were not generated or sequenced in any site-specific or condition-specific order; all samples were processed in a fully randomized, blinded fashion to minimize confounding. Batch-specific differences were assessed during preprocessing, and integration approaches in Seurat (including rPCA) were applied to mitigate residual technical variability. Variance analyses confirmed that biological factors (e.g., clinical site) contributed more to variability than sequencing batch. The analysis pipeline is depicted in Figure 1C. Raw data were processed with Cell Ranger (10X Genomics) to generate count matrices, which were then analyzed with the Seurat (v.4.0) R package.³⁸ Cells were filtered by quality, requiring 500–10,000 Unique Molecular Identifier (UMIs) per cell and <20% mitochondrial gene content, with genes detected in fewer than 3 cells excluded. For BMAC samples, the median cell count per sample was 3,680 (SD, 2,218) and the median reads per cell were 45,282 (SD, 86.678). For SVF samples, the median cell count per sample was 4,869 (SD, 2,058) and the median reads per cell were 105,555 (SD, 3,746).

After normalization, PCA was used to reduce dimensionality. Among several normalization approaches tested, including robust rPCA,³⁹ canonical correlation analysis, and Harmony,⁴⁰ rPCA was selected for its ability to balance batch effects with biological signal retention. Clustering was performed using the Louvain algorithm, and after quality control comparison, we settled on parameters that identified 37 cell clusters.

Cell type annotation and abundance assessment

Identified clusters were annotated with cell types based on the expression of known marker genes from literature and databases. The DISCO⁴¹ database (https://www.immunesinglecell.org/ct_pub) was utilized for automated annotation, supplemented by manual review. For ambiguous clusters, additional marker gene analyses and literature review were performed to assign the most probable cell type. The results were visualized using Uniform Manifold Approximation and Projection (UMAP)⁴² highlighting the distribution and relationship of the annotated cell types within the samples.

A matrix of cell counts per identified cell type for each sample was prepared by aggregating the UMI counts for all genes within a cell, thereby assigning each cell to its annotated cell type based on the clustering and annotation steps previously described. Subsequently, the relative abundance of each cell type across different samples and conditions was quantified using the Propeller tool.⁴³ This approach computes proportions on a per-donor basis before statistical testing. BaselineProp values were defined as the mean of donor-level proportions across all samples, independent of site, while PropMean.1–PropMean.5 represent the mean donor-level proportions within each clinical site.

Pseudobulk differential gene expression analysis

To investigate the relationship between differential gene expression and clinical response in the trial arms, pseudobulk⁴⁴ profiles were generated with customized scripts by aggregating scRNA-seq counts across annotated cell types within each sample. Initial quality control included PCA and hierarchical clustering on log-transformed or variance-stabilized data to assess sample similarity and detect outliers. To mitigate site-specific effects across clinical centers, batch correction was applied to the pseudobulk data using ComBat-Seq.⁴⁵ Identification of differentially expressed genes (DEGs) was then performed using DESeq2.⁴⁶ For BMAC, we utilized LRTs⁴⁷ to compare full and reduced models (excluding age, sex, and BMI) to identify enriched sets of DEGs. In parallel, variance component partitioning was performed with dreamlet^48,49 (mixed-effects pseudobulk per cell type). For the SVF dataset, additionally to the pseudobulk analyses, scRNA-seq data were also analyzed with scITD, which uses Tucker tensor decomposition to capture gene sets co-regulated across multiple cell types. All results were adjusted for multiple comparisons using the Benjamini-Hochberg method with a 5% FDR⁵⁰ threshold.

Cell-to-cell communication analysis

CellChat⁵¹ and NicheNet⁵² were used to infer intercellular communication networks from scRNA-seq data. CellChat analyzes and visualizes communication probability between cell types based on curated, experimentally validated ligand-receptor interactions, generating global signaling networks and highlighting sender and receiver populations.

NicheNet was applied to further prioritize ligands most likely to explain observed gene expression differences in receiver cells. Briefly, NicheNet correlates ligand expression in sender cell types with differential expression of predicted target genes in receiver cell types, integrating prior knowledge of ligand-target relationships. Ligand activity scores were defined as the Pearson correlation between the predicted target set and the observed gene expression ranking. Ligand performance was summarized using area under the precision–recall curve, which reflects the ability of a ligand to predict expression changes in its target genes.

Ligands were filtered using the following thresholds: |log2 fold-change (responders−non-responders) | ≥ 0.25, a difference in percent expressing cells ≥10%, and FDR < 0.1. Visualization outputs include dot plots, where dot size corresponds to the percentage of sender cells expressing a ligand and dot color indicates log2 fold-change between responders and non-responders, and heatmaps, which show normalized ligand expression in sender populations along with predicted regulatory potential of target genes. All analyses were performed using the standard workflows provided by the developers, with default parameters unless otherwise noted. Because CellChat and NicheNet address complementary aspects of cell-cell communication, their detected ligand sets do not fully overlap. CellChat identifies global signaling topology across all cell types using curated ligand-receptor interactions, highlighting pathways such as TGFβ and SEMA4. NicheNet instead ranks ligands according to their ability to predict target gene programs in receiver populations, which emphasizes sender-receiver regulatory potential rather than global signaling abundance. Accordingly, ligands highlighted by CellChat (e.g., SEMA4) may not appear among the top-ranked ligands in NicheNet’s MSC-focused prioritization.

Pathway enrichment analysis

Pathway enrichment analysis was conducted using a combination of GSEA⁵³ and its optimized counterpart, fast GSEA,⁵⁴ to rank genes based on their differential expression and assess the enrichment of predefined gene sets. The ESCAPE R⁵⁵ package was employed for the integration of pathway databases, enabling comprehensive visualization and interpretation of enriched pathways. Statistical significance was determined using nominal p values and adjusted for multiple comparisons also with a 5% FDR threshold.

Data and code availability

The datasets used in this manuscript are available with a reviewer link to GEO: GSE292453 and will be made public upon publishing this paper. Code is available at GitHub.

Acknowledgments

We wish to acknowledge the Cellular Analysis and Cytometry Core and the Molecular Evolution Core facilities at the Parker H. Petit Institute for Bioengineering and Bioscience at the Georgia Institute of Technology for the use of their shared equipment, services, and expertise. Funding for this work was provided by The Billie and Bernie Marcus Foundation to Emory University and the Georgia Institute of Technology, through the Marcus Center for Therapeutic Cell Characterization and Manufacturing (MC3M). K.R. was supported by the Robert A. Milton Endowed Chair. MC3M is also supported by the Georgia Tech Foundation and the Georgia Research Alliance.

Author contributions

P.C. performed scRNA-seq experiments and analyzed and interpreted the data with guidance from G.G. H.Y.S. and L.E.K. assisted with sample processing and quality control. S.D.B., K.M., and H.D. managed the MILES trial and conceived the overall experimental design. C.Y., H.D., and K.R. also contributed to data interpretation. All authors reviewed and approved the final manuscript, which was initially drafted by P.C. and G.G.

Declaration of interests

The authors declare no competing interests.