Assessing the clinical applicability of dimensionality reduction algorithms in flow cytometry for hematologic malignancies

Introduction

Flow cytometry is an essential tool in the diagnosis and monitoring of hematologic malignancies, allowing for rapid evaluation of the immunophenotype of target cells [1], 2]. Multiparameter flow cytometry (MFC) utilizes multiple light sources to simultaneously analyze several fluorochrome-conjugated antigens on a single cell, providing a significant advantage over immunohistochemistry [3]. In clinical laboratories, MFC panels are organized based on the immunophenotype of the target cell. Depending on various combinations of antigens and fluorochromes, MFC typically analyzes more than 8 to 12 markers, along with physical parameters, for hematologic malignancies such as acute leukemia, lymphoma, and multiple myeloma (MM) [4], 5]. In addition, measurable residual disease (MRD) analysis often requires the use of more than 10 markers to ensure accurate identification and quantification of residual abnormal cell population [6], 7].

The primary challenge in MFC analysis is the complexity of data interpretation [8]. While the increased number of simultaneously detectable fluorochromes allows for more detailed information to be obtained from each cell, the resulting data must still be visualized in two-dimensional plots. As the number of markers increases, so does the number of possible plot combinations, leading to a more complex analysis process that requires significant time and effort for accurate interpretation. Recently, Chen et al. reported a study in which they utilized a 24-color panel for the analysis of MRD in AML [9]. This highlights the usefulness of MFC as a highly informative test, but also one that requires a high level of expertise.

Dimensionality reduction (DR) algorithms are processes used to reduce data complexity by reorganizing high-dimensional data into lower-dimensional (typically two-dimensional) representations where similar data points are positioned closer together [10], 11], and these algorithms are increasingly being applied in MFC for analyzing immune cell subsets and diagnosing hematologic malignancies [12], [13], [14]. Although several studies have applied DR to MFC [12], 15], there remains a lack of systematic approaches for applying DR algorithms to MFC data for the analysis of hematologic malignancies.

t-Distributed stochastic neighbor embedding (t-SNE) and uniform manifold approximation and projection (UMAP) are widely used DR algorithms for visualizing high-dimensional data. t-SNE reduces dimensionality by converting pairwise similarities into conditional probabilities and mapping them onto a lower-dimensional space using Student’s t-distribution, effectively preserving local structures while potentially distorting global relationships [16]. In contrast, UMAP constructs a k-nearest neighbors (k-NN) graph to capture the data’s topology and optimizes the cross-entropy between high- and low-dimensional representations [17]. Compared to t-SNE, UMAP offers better global structure preservation and faster computation, making it advantageous for large-scale data visualization [18], and these algorithms have been integrated into MFC analysis software tools, such as FlowJo (Becton Dickinson, CA, USA) and Cytobank (Beckman Coulter, CA, USA) [19], 20]. This study evaluated the diagnostic applicability of t-SNE and UMAP for MFC in various hematologic malignancies by comparing their ability to qualitatively identify abnormal cell populations and quantitatively determine their proportions with traditional manual gating analysis.

Materials and methods

Study samples

This study included 237 samples from 225 patients who underwent bone marrow (BM) examination and flow cytometry for a diagnosis of hematologic malignancies between January 2020 and April 2024 (Figure 1). Flow cytometry standard (FCS) data files were collected from 167 samples which underwent the diagnostic MFC panels (80 samples with acute leukemia orientation tube (ALOT) panel for hematopoietic precursor cells, 42 samples which underwent the B-cell lymphoma (BCL) panel, and 45 samples which underwent the MM panel), and 70 samples which underwent the MRD panels (40 samples with MRD of B-ALL (B-MRD), and 30 samples with MM-MRD). The study was performed in accordance with the Declaration of Helsinki. The study was approved by the Institutional Review Board of Samsung Medical Center, Seoul, Korea (Approval No. 2024-05-003). Due to the retrospective nature of the study, the Institutional Review Board waived the requirement for informed consent.

Figure 1:

Study scheme of dimensionality reduction-based gating analysis for hematologic malignancies. The study included five MFC panels: acute leukemia orientation tube (ALOT, n=80), B-cell lymphoma (BCL, n=42), multiple myeloma (MM, n=45), B-cell acute lymphoblastic leukemia measurable residual disease (B-MRD, n=40), and MM measurable residual disease (MM-MRD, n=30).

DR-based data analysis strategies

After DR processing, the total events (ALOT, BCL, and MM panels, Figure 2A) or primary-gated events (MRD panels, Figure 2B) were visualized in a two-dimensional dot plot with the X-axis and Y-axis representing the DR coordinates. A heatmap in color axis mode was used to visualize the signal intensity for each marker, with stronger signals represented from blue to red. Target clusters were initially gated based on this heatmap according to the each MFC panel. Subsequently, the association of these clusters with specific diseases was interpreted by reviewing the heatmaps of other markers in the panel. All target clusters were cross-checked against the CD45/SSC-A plot to confirm their position and exclude nonspecific signals, such as debris. A detailed description of DR-based gating for each panel is available in the Supplementary Material (Supplementary Methods and Supplementary Figures 1–5).

Figure 2:

Common gating strategy for DR-based gating against (A) total events for the ALOT, BCL, and MM panels, and (B) applying primary-gating strategy for MRD panels.

Results

Comparison of manual and DR-based gating

Diagnostic panels: ALOT, BCL, and MM

Among the 80 samples in the ALOT panel, t-SNE-based gating showed no discrepancies in lineage classification compared to the manual gating (Supplementary Table 2). UMAP-based gating showed one discrepant case where the difference in lineage interpretation (myeloblast vs. unknown lineage) stemmed from the interpretation of weak MPO positivity. Total agreement between manual gating and DR-based gating results was 100 % (kappa=1.000 [95 % CI, 1.000–1.000]) and 98.8 % (kappa=0.980 [95 % CI, 0.941–1.000]) for t-SNE- and UMAP-based gating, respectively (Table 1).

Table 1:

Total agreement, sensitivity, and specificity, PPV, and NPV of DR-based gating compared with manual gating.

	t-SNE-based gating			UMAP-based gating
	ALOT	BCL	MM	ALOT	BCL	MM
Agreement, % Kappa value (95 % CI)	100 1.000 (1.000–1.000)	95.2 0.927 (0.830–1.000)	100 1.000 (1.000–1.000)	98.8 0.980 (0.941–1.000)	95.2 0.927 (0.830–1.000)	100 1.000 (1.000–1.000)
Sensitivity, % (95 % CI)a	NA	100 (83.9–100 %)	100 (90.0–100 %)	NA	100 (83.9–100 %)	100 (90.0–100 %)
Specificity, % (95 % CI)a	NA	95.5 (77.2–99.3 %)	100 (69.2–100 %)	NA	95.5 (77.2–99.3 %)	100 (69.2–100 %)
PPV, % (95 % CI)a	NA	95.2 (77.8–99.1 %)	100 (90.0–100 %)	NA	95.2 (77.8–99.1 %)	100 (90.0–100 %)
NPV, % (95 % CI)a	NA	100 (83.9–100 %)	100 (69.2–100 %)	NA	100 (83.9–100 %)	100 (69.2–100 %)

1. ^aBased on positive and negative results. DR, dimensionality reduction; t-SNE, t-distributed stochastic neighbor embedding; UMAP, uniform manifold approximation and projection; ALOT, acute leukemia orientation tube; BCL, B-cell lymphoma; MM, multiple myeloma; PPV, positive predictive value; NPV, negative predictive value; CI, confidential interval; NA, not available.

Among the 42 samples in the BCL panel, both t-SNE- and UMAP-based gating showed concordant results with manual gating for all but two cases (Supplementary Table 3). Notably, both DR-based gating identified a CD5ￚ/CD10ￚ clonal B-cell population in one case that was not detected by manual gating (0.59 % in t-SNE-based gating and 0.43 % in UMAP-based gating) (Supplementary Figure 6). In the other discrepant case, a difference in immunophenotype interpretation arose due to the interpretations of CD5 positivity. In qualitative analysis, the total agreement between manual gating and DR-based gating results was 95.2 % (kappa=0.927 [95 % CI, 0.830–1.000]) for both t-SNE and UMAP, with sensitivity of 100 % (95 % CI, 83.9–100 %), specificity of 95.5 % (95 % CI, 77.2–99.3 %), PPV of 95.2 % (95 % CI, 77.8–99.1 %), and NPV of 100 % (95 % CI, 83.9–100 %).

For the 45 samples in the MM panel, both t-SNE- and UMAP-based gating results showed complete agreement with manual gating (Supplementary Table 4), with no discrepant cases observed. Across all diagnostic panels, the clusters identified by DR-based gating were spatially consistent with those defined by manual gating in the CD45/SSC-A plot.

MRD panels: B-MRD and MM-MRD

For the 40 samples in the B-MRD panel, both t-SNE- and UMAP-based gating results fully agreed with manual gating in identifying 18 samples with MRD-positive and 22 samples with MRD-negative (Table 2 and Supplementary Table 5).

Table 2:

Total agreement, sensitivity, specificity, PPV, and NPV of DR-based gating compared with manual gating in MRD panels.

	t-SNE-based gating		UMAP-based gating
	B-MRD	MM-MRD	B-MRD	MM-MRD
Agreement, % Kappa value (95 % CI)	100 1.000 (0.950–1.000)	96.7 0.934 (0.835–0.982)	100 1.000 (0.950–1.000)	96.7 0.934 (0.835–0.982)
Sensitivity, % (95 % CI)	100 (82.4–100 %)	100 (80.0–100 %)	100 (82.4–100 %)	100 (80.0–100 %)
Specificity, % (95 % CI)	100 (85.1–100 %)	93.8 (69.7–99.0 %)	100 (85.1–100 %)	93.8 (69.7–99.0 %)
PPV, % (95 % CI)	100 (82.4–100 %)	93.3 (81.8–98.9 %)	100 (82.4–100 %)	93.3 (81.8–98.9 %)
NPV, % (95 % CI)	100 (85.1–100 %)	100 (80.0–100 %)	100 (85.1–100 %)	100 (80.0–100 %)

1. DR, dimensionality reduction; t-SNE, t-distributed stochastic neighbor embedding; UMAP, uniform manifold approximation and projection; MRD, measurable residuals disease; MM, multiple myeloma; PPV, positive predictive value; NPV, negative predictive value; CI, confidential interval.

Among the 30 samples in the MM-MRD panel, 14 were identified as MRD-positive and 16 as MRD-negative through manual gating. Significantly, in one sample where manual gating failed to detect MRD, both t-SNE- and UMAP-based gating successfully identified an MRD population at a level of 0.009 % (Supplementary Figure 7). Total agreement between manual gating and DR-based gating results was 96.7 % (kappa=0.934 [95 % CI, 0.835–0.982]) for both t-SNE and UMAP. Compared with manual gating, both t-SNE- and UMAP-based gating showed sensitivity of 100 % (95 % CI, 80.0–100 %), specificity of 93.8 % (95 % CI, 69.7–99.0 %), PPV of 93.3 % (95 % CI, 81.8–98.9 %), and NPV of 100 % (95 % CI, 80.0–100 %).

Quantitative correlation

Compared with the results of manual gating, t-SNE-based gating demonstrated correlation coefficients of 0.990, 0.979, and 0.944 for the ALOT, BCL, and MM panels, respectively (all p<0.001) (Figure 3). UMAP-based gating showed correlation coefficients of 0.988, 0.970, and 0.947 for the ALOT, BCL, and MM panels, respectively (all p<0.001).

Figure 3:

Quantitative correlation of target populations measured by manual gating and DR-based gating.

For the B-MRD and MM-MRD panels, after conversion to a logarithmic scale, t-SNE-based gating achieved a correlation coefficients of 0.960 for both panels (all p<0.001), while UMAP-based gating achieved correlation coefficients of 0.976 and 0.982 for the B-MRD (range, 10^−5.30–10^−2.50) and MM-MRD (range, 10^−5.15–10^−0.78) panels, respectively (all p<0.001).

Discussion

In this study, we investigated the diagnostic applicability of DR algorithms in the flow cytometry for patients with hematologic malignancies. Our findings demonstrate that both t-SNE and UMAP are highly effective in reproducing manual gating results across the ALOT, BCL, MM, B-MRD, and MM-MRD panels, with overall agreement consistently exceeding 95 % and sensitivity reaching 100 %. Furthermore, quantitative correlations were excellent across all panels (ρ>0.94).

For the ALOT panel, discrepant cases between DR-based gating and manual gating were attributed to the interpretation of weak MPO positivity led to a difference between UMAP-based gating and manual gating in lineage classification. Determining positivity in flow cytometry can be often challenging when signal intensity is weak [25]. In the case of MPO, positivity is frequently determined when as little as 10 % of the cells show positivity based on isotype or internal control cut-offs [26]. In this study, marker positivity was not rigorously quantified based on signal intensity but rather assessed using a heatmap-based approach, which may have contributed to discrepancies in the interpretation of weak signal intensity. One discrepant case in BCL panel also showed CD5 weak-positive in manual gating. Future studies should aim to establish more objective criteria for positivity by calculating the median fluorescence intensity ratio of the target cluster relative to the internal control cluster (e.g. lymphocytes) [27], allowing for a more standardized and quantitative assessment of positivity.

Importantly, in the analysis of BCL and MM-MRD panels, both t-SNE and UMAP-based gating successfully detected two critical cases that manual gating failed to identify, revealing a clonal B-cell population in one case from the BCL panel and a clonal plasma cell population in another case from MM-MRD panel. The former case was later confirmed to have morphological evidence of diffuse large BCL BM involvement, while the latter case showed an increased serum free kappa light chain level, approximately twice the upper limit of the reference interval, with lambda light chain levels remaining within the reference range and no monoclonality observed in serum or urine protein electrophoresis. Cheung et al. previously suggested the potential of diverse automated algorithms (e.g. t-SNE, SPADE1, and FlowSOM), to uncover cell populations that may be missed by conventional sequential manual gating strategies [28]. Our study provides supporting evidence for this possibility, demonstrating that DR algorithms can reveal diagnostically relevant populations that are undetected through conventional manual gating.

In the MRD panels, we applied a primary-gating strategy to reduce DR processing time. While the sensitivity evaluation using total events from the B-MRD panel, neither t-SNE nor UMAP could detect MRD levels below 10^−5.00, the implementation of CD19-based primary gating strategy enabled successfully MRD detection at levels as low as 10^−5.30.

Regarding DR processing time, t-SNE required 1ￚ3 minutes per sample for ALOT, BCL, and MM panels, each containing approximately 200,000 total acquisition events. In contrast, for the B-MRD panel, which included up to six million total acquisition events, the processing time extended to 1ￚ1.5 h per sample. While UMAP showed shorter processing time compared to the t-SNE, as previously reported [29], we did not thoroughly evaluate these differences in our study. Although processing times increase with the number of analyzed events for both algorithms [20], UMAP may be more suitable than t-SNE for MRD analysis from a computational time perspective, particularly for very large datasets containing millions of cell events. Processing time can also vary depending on the performance of the analysis computer, as well as algorithm parameters such as iteration time and perplexity. Utilizing primary-gating strategies significantly reduced the processing time and enhanced sensitivity, which was accomplished through both the reduction of events in primary-gated populations and more refined cluster discrimination.

The advantage of DR-based gating is its ability to streamline the data interpretation process by reducing the number of two-dimensional plots that need to be reviewed [28]. In DR-based gating, only single t-SNE or UMAP plot needed to be checked after primary gating on the CD19-positive population. Any nonspecific signals or non-MRD cells overlapping with the target population could be verified using the CD45/SSC-A and CD19/SSC-A plots. For other panels without primary gating, the target clusters could be identified in the DR plot after debris exclusion, followed by verification of their expected locations in the CD45/SSC-A plot.

DR-based gating can be applied in various ways. In this study, while the DR algorithm was applied exclusively to immunophenotypic markers, incorporating physical parameters like FSC and SSC could enhance cluster accuracy by improving cell classification [30]. Additionally, although we manually classified CD19-positive or CD138-positive populations during primary gating in MRD panels, this step could be further automated using algorithms in the future, implementing a multi-step automated gating strategy for increased efficiency and consistency. Furthermore, combining multiple clustering algorithms may be more effective for detecting very small MRD populations or unexpected cell clusters. For example, Botta et al. reported a FlowCT semi-automated workspace that integrates flowAI (for removing low-quality events), PARC and FlowSOM (for automated clustering), PCA, t-SNE, and UMAP (for DR), and other machine learning algorithms [29]. Importantly, a prerequisite for such automated gating is not solely dependent on DR algorithms but primarily on achieving high-quality data output and well-defined cell clustering with appropriate reagents [28].

Our study has a few limitations. First, we utilized the default settings provided by the FlowJo software for DR algorithms. The analytic performance of DR-based gating can vary significantly depending on the specific parameters used. Shadi et al. suggested that increasing parameters like iteration number and perplexity can lead to greater separation between distinct cellular lineages with discrete marker distribution [12]. However, excessive adjustment of these parameters may result in less distinct clustering, potentially obscuring the data interpretation. Ultimately, there is a need for efforts to customize the parameter settings to fit the specific panel and laboratory conditions. Second, the demarcation of clusters in DR plots is not always well-defined, leading to a certain degree of subjectivity in cluster identification, especially in myeloblasts. Due to the heterogeneous and complex nature of myeloblasts [31], clearly gating the target population was often challenging both in manual gating and DR-based gating. This issue was particularly evident with t-SNE, which effectively separates distinct clusters but has limitations in resolving populations with continuous marker expression, potentially leading to qualitative or quantitative differences in gating results [32]. When clusters did not show clear boundaries, we determined the division points where the clusters began to narrow. Lastly, sensitivity and precision were assessed for only a subset of panels (BCL, B-MRD). To ensure the reliability and applicability of these algorithms in routine diagnostic practice, it is essential to validate these parameters across all MFC panels used in clinical workflows.

In conclusion, DR-based gating demonstrates significant potential as a valuable tool in flow cytometric analysis of hematologic malignancies, showing high concordance with manual gating and applicability to MRD analysis. DR-based gating is expected to facilitate streamlined data interpretation by clustering events with similar immunophenotypic characteristics while reducing the complexity of traditional interpretations and minimizing overlooked populations. Furthermore, integration with other automated tools could further enhance these capabilities, improving both efficiency and rare event detection while expanding its utility in both research and clinical settings.