Quantitative phase imaging with temporal kinetics predicts hematopoietic stem cell diversity

Abstract

Innovative identification technologies for hematopoietic stem cells (HSCs) have expanded the scope of stem cell biology. Clinically, the functional quality of HSCs critically influences the safety and therapeutic efficacy of stem cell therapies. However, most analytical techniques capture only a single snapshot, disregarding the temporal context. A comprehensive understanding of the temporal heterogeneity of HSCs necessitates live-cell, real-time and non-invasive analysis. Here, we developed a prediction system for HSC diversity by integrating single-HSC ex vivo expansion technology with quantitative phase imaging (QPI)-driven machine learning. By analyzing the cellular kinetics of individual HSCs, we discovered previously undetectable diversity that snapshot analysis cannot resolve. The QPI-driven algorithm quantitatively evaluates stemness at the single-cell level and leverages temporal information to significantly improve prediction accuracy. This platform advances the field from snapshot-based identification of HSCs to dynamic, time-resolved prediction of their functional quality based on past cellular kinetics.

This study introduces a label-free imaging and machine learning platform that predicts hematopoietic stem cell function based on their dynamic behaviors, offering new insights into stem cell diversity and potential clinical applications.

Introduction

Hematopoietic stem cell (HSC) research has advanced in parallel with the development of innovative identification technologies. Originally, the spleen colony-forming assay was employed to evaluate the differentiation capacity of HSCs¹. However, the advent of flow cytometry in the 1980s marked a paradigm shift in cell categorization methodology^2,3, unveiling the presence of murine HSCs in a subset defined as Thy-1^low Lineage^- Sca1⁺³. Subsequent investigations elucidated the existence of long-term self-renewing HSCs (LT-HSCs) and transient short-term HSCs (ST-HSCs)⁴. Ultimately, LT-HSCs were directly identified at the single-cell level within a bone marrow fraction characterized as Lineage^- cKit⁺ Sca1⁺ CD34^-5. Similarly, human HSCs have been identified within the CD34⁺ CD38^- CD90⁺ CD45RA^- CD49f⁺ fraction of human umbilical cord blood, also at single-cell resolution⁶. In line with these discoveries, we have previously reported on the exploration of HSC-specific marker genes and the development of methods for marking HSCs both in vitro and in vivo, establishing a technical platform^{7, 8–9}.

These advances in HSC identification techniques have driven the entire field of stem cell research forward, leading to the discovery of numerous stem cell-specific markers^{10, 11–12}. Moreover, single-cell analytical methods have uncovered the complex heterogeneity within stem cell populations^13,14. Specifically, in the field of hematology, this heterogeneity can be observed in the differential differentiation capacity of HSCs. Some HSCs are predominantly committed to myeloid or lymphoid lineages, while others are biased towards megakaryocyte differentiation^{15, 16, 17–18}. Additionally, heterogeneity exists in the metabolic states and clonal expansion capacities of HSCs^{19, 20–21}. Single-cell RNA sequencing (scRNAseq) and other technologies have emerged as established tools for elucidating the diversity among individual cells, significantly advancing research into the heterogeneity of HSCs²². However, these techniques are limited to evaluating the cellular state at a single temporal snapshot, thus failing to capture temporal dynamics. Cells are dynamic system, they undergo continual change and never retain a single state, HSCs are no exception^23,24. The comprehensive understanding of the temporal diversity of HSCs mandates an experimental system that allows for the real-time tracking and non-invasive probing of live cells for their analysis. The ultimate objectives include forecasting future cellular status based on past temporal dynamics and quantitatively predicting the functionality of HSCs at the single-cell level. Such predictive systems will be pivotal in further advancing the field of stem cell biology through enabling the establishment of a platform that predicts future stem cell diversity using their past behavior.

The long-term evaluation of phenotypic and kinetic behaviors of single HSCs firstly requires a culture system that allows for HSC expansion while maintaining function. We have previously established a long-term culture system for the expansion of both murine and human HSCs^{25, 26–27}. Conventional fluorescent imaging techniques may impair stem cell function due to the introduction of fluorophore or high-intensity illumination, potentially causing phototoxicity. Ptychographic quantitative phase imaging (QPI) techniques facilitate non-invasive and label-free monitoring of live cells across a wide field of view without the need for high-intensity light imaging^28,29. Furthermore, the meniscus compensation step during phase reconstruction actively senses the deformation of transmitted light, enabling fully quantitative and aberration-free imaging, even in U-bottomed culture wells, which allows high-throughput single HSC imaging during ex vivo expansion.

In this work, by integrating our single-HSC ex vivo expansion technology^25,26 and QPI-driven machine learning, we have developed a prediction system for HSC diversity. This achievement signifies a paradigm shift from the era of identification of HSCs through snapshot analysis to the era of prediction of HSC function based on temporal kinetics, thereby fundamentally altering the landscape of stem cell research.

Results

Quantitative phase imaging and single-cell kinetics analysis reveal infinite diversity of HSCs

To investigate the dynamics of HSCs during ex vivo expansion at a single-cell level, we combined our recently established single-cell expansion culture system for murine HSCs with QPI. We sorted a single CD201⁺CD150⁺CD48⁻KSL cell from a population of ex vivo expanded HSCs and monitored expansion for 96 h with time-lapse QPI (Fig. 1a). Quantitative analysis of cellular kinetics revealed remarkable diversity. After 96 h, marked differences in proliferation rate could be observed where 12.5% of HSCs underwent a rapid proliferation defined as producing more than 20 cells whereas other HSCs divided more slowly with 21.9% producing less than 4 cells (Fig.1b–f and Supplementary Movie 1). Morphologically, we observed significant variations in the output cells; 10.9% of HSCs produced cells with dry masses larger than 200 pg whereas 17.2% of HSCs produced cells with dry masses smaller than 100 pg (Fig.1g–k, Supplementary Fig.1a,b and Supplementary Movie 2). We further investigated the interval between the first and second cell divisions (Division Gap) and observed significant variations among cells; in 25.5% of cells, this interval exceeded 5 h, indicating the possibility of asymmetric cell divisions (Supplementary Fig.1c and Supplementary Movie 3). Interestingly, we also observed interrupted cytokinesis during the cell division process. To determine the frequency of these rare cell division processes, we conducted QPI of expanded HSCs (Supplementary Movie 4) over a wide field of view, measuring 1000 μm × 1000 μm. The imaging data captured 1243 cell division events within 36 h, with 91.3% exhibiting normal division patterns, while 8.21% showed interruptions in cytokinesis, and 0.48% exhibited abnormal division patterns, such as a single cell dividing into three cells during one mitosis. (Supplementary Fig.1d–g).

[See PDF for image]

Fig. 1

Quantitative phase imaging-based single-cell kinetic analysis of expanded HSCs.

a Protocol for analyzing single HSC time-lapse quantitative phase imaging (QPI). CD34^-CD201⁺CD150⁺KSL cells were isolated from fresh mouse bone marrow and cultured ex vivo for 7 days. Subsequently, a single CD48^-CD150⁺CD201⁺KSL ex vivo expanded HSC was isolated and monitored using time-lapse QPI for 96 h. Data represent n = 2 independent biological experiments. b, c Representative images of HSCs that underwent rapid or slow divisions over 96 h. Scale bar: 100 μm. d Number of cells produced by each HSC after 96 h of expansion, derived from experiments conducted following the protocol described in (a). Each ID represents a single HSC clone (n = 64). Source data are provided as a Source Data file. e, f Representative proliferation patterns of HSCs that underwent rapid or slow divisions over 96 h. Source data are provided as a Source Data file. g, h Representative images of HSCs that produce cells with high or low dry mass. i Dry mass of cells produced by each HSC after 96 h of culture, derived from experiments conducted following the protocol described in (a). Each ID represents a single HSC clone (n = 64). Source data are provided as a Source Data file. j, k Representative dry mass dynamics of HSCs that produce either high- or low-mass cells. Source data are provided as a Source Data file. l Protocol for classifying cells based on kinetic features derived from QPI data. Fifty CD34^-CD201⁺CD150⁺KSL cells were isolated from fresh mouse bone marrow. Subsequently, each HSC was tracked for 36 h, yielding a total of 11,512 cell images. From these images, 11 kinetic parameters (Dry Mass, Sphericity, Velocity, Volume, Mean Thickness, Perimeter, Radius, Area, Length, Width and Length/Width Ratio) were extracted using the Cell Analysis Toolbox within the Livecyte platform, followed by Uniform Manifold Approximation and Projection (UMAP) analysis and hierarchical clustering. Data represent n = 3 independent biological experiments. m UMAP analysis of expanded HSCs based on kinetic features obtained by QPI, colored by hierarchical clustering, derived from experiments conducted following the protocol described in (l). Source data are provided as a Source Data file.

Therefore, using QPI time-lapse imaging, individual HSCs were observed displaying unique behaviors even in pure phenotypic HSC fractions, highlighting the infinite diversity seen within the HSC population.

Gene-independent cell classification through kinetic features of murine and human HSCs

Next, we evaluated the potential for cell classification based on kinetic features obtained from QPI data. We tracked 50 murine HSCs over 36 h, and 11 parameters were extracted from a total of 11,512 cell images for Uniform Manifold Approximation and Projection (UMAP) analysis and clustering (Fig.1l and Supplementary Movie 5). This resulted in the identification of four distinct clusters, each with unique characteristics (Fig.1m). Notably, Cluster 3 exhibited cells with low dry mass, high sphericity and low velocity, while Cluster 4 comprised cells with high dry mass (Supplementary Fig. 2a,b). This UMAP is derived from time-lapse imaging data and includes actual time information, enabling the tracking of cells on the UMAP. For example, Cell ID No.1 and No.2 remained within Cluster 3 for 36 h without dividing, whereas Cell ID No.3 and No.4 underwent division and were found within Cluster 1,2 and 4. Cell ID No.5 remained in cluster 4 (Supplementary Fig. 2c). By referencing the temporal data, we observed a temporal progression from the top of Cluster 3 through Cluster 1 and 2, and towards Cluster 4 (Supplementary Fig. 2d,e). This suggests that cells at the top of Cluster 3 are the most immature HSCs.

Subsequently, we examined whether kinetic diversity is present in human HSCs. Human cord blood CD34+ cells were cultured using a recently developed HSC expansion system²⁷ and monitored with QPI. Similar to murine HSCs, human HSCs exhibited various kinetic features (Supplementary Movie 6). From 24 h of QPI time-lapse data, 11 parameters were extracted from 25,018 cell images, followed by UMAP analysis and clustering (Supplementary Fig. 3a). Consistent with findings in murine HSCs, this led to the identification of four distinct clusters, each with unique characteristics (Supplementary Fig. 3b–d). Notably, Cluster 2 included cells with a high length/width ratio, low sphericity (i.e., elongate cells), and high velocity, whereas Cluster 4 contained cells with low dry mass, low length/width ratio, high sphericity and low velocity (Supplementary Fig. 3b–d). Analysis using an imaging-enhanced cytometer revealed that elongated cells (Type A) or cells with protrusions (Type B) exhibited a lower proportion of CD201^high CD90^high HSCs compared to circular cells (Type C) (Supplementary Fig. 3e). Thus, similar to murine HSCs, human HSCs can be classified based on kinetic features, suggesting a relationship between kinetic features and stem cell potential.

To investigate the dynamics of human HSCs during ex vivo expansion at the single-cell level, Lin^-CD34⁺CD38^-CD45RA^-CD90⁺CD201⁺ HSCs were sorted individually from a population of ex vivo expanded human HSCs into a 96-well U-bottom plate and cultured for 21 days. As observed in murine HSCs, cell proliferation dynamics were heterogenous (Supplementary Fig. 3f). Type 1-4 cells exhibited high initial proliferation rates but differentiated and underwent early cell death, reaching their proliferation peak early and subsequently declining in number. In contrast, Type 5 cells showed lower initial proliferation rates but continued to proliferate over an extended period (Supplementary Fig. 3f,g). Type 5 cells constituted 34.4% of the pure phenotypic HSC fractions (Supplementary Fig.3g). Flow cytometry analysis revealed that cells maintaining over 90% Lin^- population at day 18 exhibited slower proliferation compared to other cells (Supplementary Fig. 3h). Among these, cells with lower proliferative capacity retained a phenotypic HSC pattern of CD90⁺CD201⁺ even after 18 days (Supplementary Fig. 3i,j). Thus, analyzing HSCs based on cellular kinetic features demonstrated the presence of heterogeneity even within the pure phenotypic human HSC fraction, lending further support to the relationship between kinetic attributes and stemness levels.

Hlf marks functional HSCs during ex vivo expansion

To further investigate the direct relationship between kinetic features and stemness levels, a system that simultaneously tracks stemness levels in real-time alongside time-lapse QPI is required. Despite the identification of several genes expressed in fresh bone marrow HSCs, there is insufficient information on genes that are expressed specifically in ex vivo expanded murine HSCs. To capture the differentiation process of murine HSCs during ex vivo expansion, we first analyzed expanded HSCs on days 1, 3, 5 and 7 with flow cytometry, revealing that CD201⁺CD48^- cells gradually transitioned to CD201⁺CD48⁺ and CD201^-CD48⁺ cells over time (Fig. 2a, b and Supplementary Fig.4a). Reanalysis of bulkRNA-seq data for each fraction showed that Hlf, Mecom and Fgd5 were highly expressed in the CD201⁺CD48^- fraction, with reduced expression in CD201⁺CD48⁺ and CD201^-CD48⁺ cells (Fig.2c and Supplementary Fig.4b)²⁶. Single-cell RNA sequencing of expanded HSCs revealed that Hlf was highly expressed in the HSC fraction and sharply decreased with differentiation (Fig. 2d and Supplementary Fig. 4c,d). Additionally, Hlf demonstrates a broader range of expression changes across differentiation stages compared to Mecom and Fgd5. This broader dynamic range makes Hlf more suitable for simple and quantitative evaluation of stemness using fluorescence imaging. Therefore, Hlf was determined to be an effective indicator for imaging-based quantitative evaluation of stemness during ex vivo expansion.

[See PDF for image]

Fig. 2

Hlf marks functional HSCs during ex vivo expansion.

a t-distributed stochastic neighbor embedding (tSNE) analysis of ex vivo expanded HSCs based on Fluorescence Activated Cell Sorter (FACS) data (CD201, CD48, CD150, cKit, Sca1, Lin). FACS data from three independent experiments were integrated for the analysis. b tSNE of expanded HSCs, overlaid with CD201 and CD48 markers. Data from three independent experiments were integrated for this analysis. c RNA-seq profiles of selected HSC-associated genes in CD48^-CD201⁺CD150⁺KSL cells (n = 3). Data reanalyzed from previously published source²⁶. Error bars represent standard deviation (SD). Mean of n = 3 independent cultures. Source data are provided as a Source Data file. d UMAP of single-cell RNA-seq data from 7-day expanded HSCs with 10 clusters and Hlf expression. e tSNE analysis of ex vivo expanded HSCs from Hlf-tdTomato mice. FACS data from days 1, 3, and 5 were integrated. Left: gating for high, middle, and low Hlf-tdTomato populations. Right: populations highlighted across expansion days. Data from n = 3 independent experiments were integrated. f Gating strategy for Hlf-tdTomato high and low expanded HSCs. g Mean donor peripheral blood chimerism in primary (n = 4, 5 mice per group) and secondary recipients (n = 3–5 mice per group). Source data are provided as a Source Data file. h Mean donor bone marrow chimerism in primary recipients (n = 4 mice in Hlf-tdTomato^high group, n = 5 in Hlf-tdTomato^low group). Error bars represent SD. Two-sided t-test, **P = 0.0025. Source data are provided as a Source Data file. i Schematic of hematopoietic dynamics following transplantation of expanded HSCs, classified by Hlf-tdTomato expression using AkaBLI. Luminescence recorded on day 9–28 post-transplantation. Experiments were conducted with n = 5 mice per group. j Representative IVIS images from the Hlf-tdTomato high/low Akaluc+ HSC transplantation. k, l Dynamics of total average luminescence intensity after transplantation of Hlf-tdTomato high/low expanded HSCs, derived from experiments conducted following the protocol described in (i). Experiments were conducted with n = 5 mice per group. All individual IVIS imaging data are presented in Supplementary Fig. 4h. Source data are provided as a Source Data file.

Based on these results, we utilized Hlf-tdTomato reporter mice to quantitatively track the stemness level of HSCs³⁰. High expression of tdTomato was observed in bone marrow derived CD34^-CD150⁺CD48^-KSL cells from Hlf-tdTomato reporter mice (Supplementary Fig. 4e). We validated whether the fluorescence intensity of Hlf-tdTomato was an effective indicator of the differentiation process of HSCs in expansion culture by analyzing cells from day 1, 3 and 5 post-expansion with flow cytometry. As expected, a strong correlation between changes in Hlf-tdTomato expression and temporal differentiation in HSCs was confirmed (Fig. 2e and Supplementary Fig. 4f,g). To determine whether Hlf-tdTomato expression levels act as a functional indicator of HSCs, tdTomato high and low expanded HSCs were transplanted into irradiated recipients, and long-term bone marrow reconstitution abilities were evaluated. HSCs with high tdTomato demonstrate robust long-term marrow reconstitution capabilities. Conversely, chimerism decreased in those who received tdTomato-low HSCs (Fig. 2f–h). Furthermore, the analysis of the spatiotemporal dynamics of early hematopoiesis post-transplantation revealed that tdTomato high HSCs gradually reconstituted systemic hematopoiesis, whereas tdTomato low HSCs appeared to induce rapid hematopoiesis in the spleen between day 14 and 21 (Fig. 2i–l, Supplementary Fig. 4h-j).

Consequently, Hlf-tdTomato expression levels are a useful indicator of stemness in murine HSCs during ex vivo expansion and serve as a potent predictor of both long-term and short-term hematopoietic dynamics post-transplantation.

Predicting Hlf expression levels from cellular kinetic features

Next, we analyzed whether kinetic features could predict stemness, using Hlf expression as a proxy indicator. We performed UMAP analysis using cellular kinetic features obtained from time-lapse imaging, excluding tdTomato fluorescence intensity, confirming multiple clusters resembling those observed in Fig. 2 (Fig. 3a, c, Supplementary Fig. 5a,b and Supplementary Movie 7). Then, the average fluorescence intensity of tdTomato was plotted onto the same plot, which revealed significant differences in tdTomato expression levels among clusters (Fig. 3b, d). In particular, we observed the highest tdTomato expression in Cluster 1 that gradually decreased from Cluster 2 through Cluster 5 (Fig. 3b, d). Cells with high tdTomato expression exhibited higher sphericity, lower dry mass, and lower velocity than those with low tdTomato expression (Fig. 3e, Supplementary Fig.6), mirroring the relationship observed in human HSCs and their kinetic features (Supplementary Fig. 3a–e). These findings suggest that Hlf-tdTomato expression levels can be predicted from kinetic features of HSCs. We further assessed the robustness of our approach under various stress conditions, including lentiviral transduction, inflammatory cytokine exposure, and gene editing (Supplementary Fig. 5c–j). Under these conditions, the enrichment of Hlf-tdTomato high cells across clusters was less prominent compared to wild-type HSCs cultured under our standard conditions, suggesting that the current method is optimized for a defined set of culture conditions. While not universally applicable across all stress conditions, our ex vivo expansion system²⁶, in combination with time-resolved QPI, enabled the prediction of HSC potency.

[See PDF for image]

Fig. 3

Predicting Hlf expression levels from cellular kinetic features.

a Protocol for classifying cells based on kinetic features derived from QPI data. Fifty Hlf-tdTomao⁺CD34^- CD150⁺KSL cells were isolated from fresh mouse bone marrow. Subsequently, each HSC was tracked for 36 h, yielding a total of 11,876 cell images. From these images, 11 kinetic parameters (Dry Mass, Sphericity, Velocity, Volume, Mean Thickness, Perimeter, Radius, Area, Length, Width and Length/Width Ratio) were extracted, followed by UMAP analysis and hierarchical clustering, and an overlay of Hlf-tdTomato expression levels based on fluorescence imaging. Data represent n = 3 independent biological experiments. b UMAP plot of expanded HSCs based on kinetic features, overlaid with Hlf-tdTomato expression level. Representative images of clusters are shown on the plot. derived from experiments conducted following the protocol described in (a). Source data are provided as a Source Data file. c UMAP plot colored by hierarchical clustering, derived from experiments conducted following the protocol described in (a). Source data are provided as a Source Data file. d Mean red intensity of each cluster, derived from experiments conducted following the protocol described in (a). This result was validated using other data sets (Supplementary Fig. 5c,d). Source data are provided as a Source Data file. e Violin plot comparing dry mass, sphericity, and velocity among Hlf-tdTomato high, middle, and low expanded HSCs. Hlf-tdTomato⁺ CD34⁻ CD150⁺KSL cells were isolated from fresh mouse bone marrow. After 7 days of ex vivo expansion, cells were sorted from the CD201⁺CD150⁺CD48⁻KSL fraction into Hlf-tdTomato high, middle, and low populations based on the histogram shown on the left, with 100 cells per group. Twenty-four hours after sorting, QPI was performed to extract kinetic features. Data represent n = 3 independent biological experiments. The median is indicated by bold dotted lines, and the quartiles by thin dotted lines. Statistical significance was assessed using one-way ANOVA with Tukey’s post-test (****P < 0.0001). Source data are provided as a Source Data file.

Diversity of HSCs based on single cell kinetic analysis with Hlf dynamics

To analyze individual cell dynamics more precisely, we performed index sorting on the CD201⁺CD48^-CD150⁺KSL fraction of expanded Hlf-tdTomato⁺ HSCs, followed by ex vivo expansion and 96-h QPI (Fig.4a). We observed diverse cellular dynamics between every single cell. By tracking changes in Hlf-tdTomato expression levels, we confirmed that diversity exists in fluorescence intensity dynamics among cells. After 96 h of expansion, 18.8% of HSCs produced high Hlf-tdTomato cells (>5000), whereas 14.1% of HSCs produced only low Hlf-tdTomato cells (<1000) (Fig.4b). The index sorting data revealed a positive correlation between the number of output cells and CD150 expression levels (P < 0.0001, R squared = 0.40) (Fig. 4l) and a weak negative correlation with Hlf-tdTomato levels (P = 0.0072, R squared = 0.11) (Fig. 4m). The average intensity of Hlf-tdTomato in output cells was positively correlated with Hlf-tdTomato (P < 0.0001, R squared = 0.53) (Fig. 4n) and other markers in input cells, such as CD201, Sca1 and cKit (Supplementary Fig.7a), while negatively correlated with the number of output cells (P < 0.0001, R squared = 0.38) (Fig. 4o). Cells that divided quickly showed a decrease in Hlf-tdTomato expression over time, while those that divided slowly maintained Hlf-tdTomato levels (Fig.4c, d, f,g, I, j, Supplementary Fig. 7b,c and Supplementary Movie 8). In fact, Hlf-tdTomato high HSCs predominantly produced HSC-like cells, whereas Hlf-tdTomato low HSCs produced more differentiated downstream cells (Supplementary Fig.7d-g). This aligns with the relationship observed in human HSCs, where slowly dividing human HSCs maintained their undifferentiated state, reflecting the connection between stemness levels and proliferation dynamics (Supplementary Fig.3f-j). Some cells maintained high Hlf-tdTomato expression and divided but eventually led to early cell death (Fig. 4e, h, k and Supplementary Movie 9). These cells showed lower levels of CD150 expression. To investigate the in vivo dynamics of these cells, CD150^high Hlf-tdTomato^high CD201⁺CD48^-KSL cells and CD150^dim Hlf-tdTomato^high CD201⁺CD48^-KSL cells were transplanted into irradiated mice. As a result, CD150^high cells showed long-term reconstitution capabilities, whereas CD150^dim cells exhibited short-term reconstitution abilities, and the chimerism gradually decreased over time (Supplementary Fig.7h-k). Thus, Hlf-tdTomato^high CD150^dim cells maintained Hlf-tdTomato expression ex vivo but exhibited short-lived characteristics, indicating short-term HSCs.

[See PDF for image]

Fig. 4

Single HSC kinetics analysis with Hlf-tdTomato dynamics.

a Protocol for analyzing single HSC time-lapse quantitative phase imaging. Fresh HSCs from Hlf-tdTomato reporter mouse bone marrow were cultured ex vivo for 7 days. Subsequently, a single CD48^-CD150⁺CD201⁺KSL ex vivo expanded HSC was isolated and monitored using time-lapse QPI for 96 h. Cellular kinetics were then quantified using the Cell Analysis Toolbox within the Livecyte platform. Data represent n = 2 independent biological experiments. b Red mean fluorescence intensity of each cell produced after 96 h, derived from experiments conducted following the protocol described in (a). Each ID represents a single HSC clone (n = 63). Source data are provided as a Source Data file. c–e Representative images of single sorted HSCs (input) and their produced cells (output) over 96 h. Scale bar: 100 μm. Hlf-tdTomato expression shown as red overlay. Representative dynamics of HSCs that maintained high Hlf-tdTomato levels and divided slowly (f, i), decreased Hlf-tdTomato levels and proliferated rapidly (g, j), or maintained Hlf-tdTomato levels but did not proliferate (h, k). Source data are provided as a Source Data file. l Correlation between the CD150 level of sorted single HSCs and the cell count after 96 h, derived from experiments conducted following the protocol described in (a). Simple linear regression (blue line) and 95% CI (black lines). Source data are provided as a Source Data file. m Correlation between the Hlf-tdTomato level of sorted single HSCs and the cell count after 96 h, derived from experiments conducted following the protocol described in (a). Simple linear regression (blue line) and 95% CI (black lines). Source data are provided as a Source Data file. n Correlation between the Hlf-tdTomato level of sorted single HSCs and the average Hlf-tdTomato level of produced cells after 96 h, derived from experiments conducted following the protocol described in (a). Simple linear regression (blue line) and 95% CI (black lines). o Correlation between the cell count and the average Hlf-tdTomato level of produced cells after 96 h, derived from experiments conducted following the protocol described in (a). Simple linear regression (blue line) and 95% CI (black lines).

Despite the general trends observed, individual cell dynamics showed significant variations. Some cells maintained Hlf-tdTomato expression while producing many cells, indicating high self-renewal capabilities as HSCs (Supplementary Fig.7k and Supplementary Movie 10). However, these cells could not be identified using index sorting data. The size of output cells also varied, showing weak positive correlations with CD48 and CD150, and a weak negative correlation with Hlf-tdTomato, though these correlation coefficients were low, making predictions using index sorting data challenging (Supplementary Fig.7a). Additionally, no correlation was found between Division Gap and index sorting data (Supplementary Fig.7a). Thus, through the integration of QPI with temporal changes in Hlf-tdTomato expression, we were able to gain deeper insight into the complexity and diversity of HSCs beyond what can be captured by flow cytometry analysis alone.

Machine learning prediction of Hlf expression levels from live-cell behavioral dynamics

We have demonstrated the feasibility of predicting Hlf-tdTomato expression levels in HSCs using UMAP analysis with multiple cellular kinetic features extracted from QPI data. However, there are only a limited number of biological features used in the analysis. Therefore, we developed a system that can more accurately predict the Hlf expression levels of each cell by training a deep neural network with QPI datasets.

We tracked individual target cells and extracted QPI video data specific to the target cells, generating a dataset that matched the video data of each cell from Frame 1 to Frame n with the corresponding Hlf-tdTomato intensity at the final frame (Frame n), thereby excluding the influence of surrounding cells from the videos. Machine learning was conducted using 3D Residual Neural Network (ResNet) on this training dataset, and the model was validated using a separate dataset (Fig.5a). As a result, when predictions were made utilizing a sequence of 20 consecutive frames, the expression level of Hlf-tdTomato was predicted with remarkably high accuracy (correlation coefficient: 0.73, error: 0.02) (Fig.5c–e). Notably, by increasing the number of frames employed for learning from 5 to 20, the prediction accuracy significantly improved, more than doubling (Fig.5e). To evaluate how different model architecture might enhance prediction accuracy, we tested Convolutional Neural Networks (CNNs) and transformer-based models of different parameter sizes (Supplementary Fig.8). The correlation between measured and predicted Hlf-tdTomato expression increased with larger models, with significant gains for models under 30 M parameters – indicating smaller models lack sufficient expressiveness. Beyond 30 M parameters, gains diminished, suggesting these models have reaches the required expressiveness threshold. Although a larger-scale model offered a minor improvement, this advantage did not sufficiently outweigh the added computational complexity. Consequently, given its relatively simple architecture and robust benchmark performance, the 3D ResNet remained the most practical baseline for our current dataset.

[See PDF for image]

Fig. 5

QPI-driven machine learning prediction of Hlf expression levels.

a Protocol for QPI-driven machine learning to predict the Hlf-tdTomato level of live HSCs. Fifty Hlf-tdTomato⁺ CD34⁻ CD150⁺KSL cells were isolated from fresh mouse bone marrow. Subsequently, each HSC was tracked using QPI for 60 h; time-lapse imaging was performed every 4 min and 21 s. Individual cells were tracked, and their trajectories were extracted to generate a dataset matching video data of each cell from frame 1 to frame n with the corresponding Hlf-tdTomato intensity in the final frame (frame n). 3D Residual Neural Network (ResNet) was used on this training dataset. Dataset consists of videos from n = 3 independent biological datasets for training, and a separate, independent dataset was used for validation. b Illustration of representative datasets with fixed field-of-view images (above) and images centered on cell position (below). c QPI-driven machine learning with and without cell motility, comparing predicted Hlf-tdTomato expression levels (x-axis) and measured Hlf-tdTomato levels (y-axis). d Mean squared error of Hlf-tdTomato predictions with and without cell motility using 5, 10, and 20 input images. Source data are provided as a Source Data file. e Correlation coefficient of Hlf-tdTomato predictions with and without cell motility using 5, 10, and 20 input images. Source data are provided as a Source Data file.

To address the limitation imposed by the dataset size, we evaluated the effects of data augmentation and transfer learning. For data augmentation, random flipping and rotation were applied to each video, effectively increasing the dataset size by approximately tenfold. This approach substantially improved model performance, as evidenced by a reduction in root mean squared error (RMSE) and an increase in the correlation coefficient (Supplementary Fig 9). Additionally, we investigated transfer learning using several models, including S3D, 3D ResNet, and Swin3D Transformers (tiny, normal, large) with pretrained parameters from PyTorch, originally trained on the Kinetics-400 dataset. While transfer learning alone yielded modest gains, likely due to key differences between the pre-trained and our datasets, its combination with data augmentation resulted in a median performance improvement of approximately 1.80-fold, with the exception of an outlier observed with ResNet. To assess model reproducibility, repeated training runs were conducted and the coefficient of variation (CV) for RMSE values across interaction was computed. The incorporation of data augmentation reduced the median CV from 28% to 11%, a 2.3-fold improvement underscores enhanced reproducibility and robustness (Supplementary Table 1).

We also investigated the impact of cellular dynamics on prediction accuracy. Previous analysis had shown that less migratory HSCs tended to express high levels of Hlf (Fig.3). To assess the impact of HSC motility on prediction accuracy, we removed the X, Y positional information from the tracked cells, fixing the position of cells in the center of the video, and made prediction of Hlf-tdTomato expression levels (Fig.5b and Supplementary Movie 11). Consequently, when positional information was excluded, the prediction accuracy of Hlf-tdTomato expression levels significantly decreased and increasing the number of frames did not improve accuracy (Fig.5c–e). This indicates that the motility of HSCs during ex vivo expansion is an important indicator of stem cell properties.

In conclusion, integrating HSC ex vivo expansion technologies with QPI to continuously monitor HSC characteristics in a temporally dependent manner reveals HSC diversity previously undetectable by snapshot analysis. Furthermore, the use of a predictive model leveraging multidimensional analysis based on cellular kinetic features and a deep neural network has demonstrated that the introduction of temporal information significantly enhances the accuracy of predicting stem cell properties from QPI data. Together, these achievements highlight the importance of the temporal dimension in uncovering the dynamic behavior and functional heterogeneity of HSCs.

Discussion

In this study, we developed a system that continuously tracks the kinetic features and dynamics of HSCs by integrating our established HSC expansion culture system with QPI technology, allowing for high-precision, label-free prediction of HSC function. By increasing the number of video frames and considering the temporal mobility of cells, the system achieves a significant improvement in accuracy, highlighting the importance of evaluating cellular kinetic features over time. Focusing on temporal dynamics during ex vivo expansion culture, rather than snapshots³¹, kinetic features have been revealed as key biological parameters defining HSC characteristics. This technology revolutionizes the approach to HSC identification and establishes a next-generation label-free functional prediction platforms for expanded HSCs.

Through multidimensional analysis using extracted feature metrics, we have successfully classified cells into multiple categories, independent of genetic information, and have charted their continuous cell differentiation trajectories on UMAP. This strategy enables the analysis of authentic temporal shifts in the same cell, in contrast to estimated temporal data, such as pseudo time analysis obtainable through scRNAseq. Indeed, incorporating temporal information revealed diversities in HSCs that remained obscured in traditional snapshot analysis.

In the past, predicting HSC functionality through time-lapse imaging necessitated specialized equipment, such as custom plates^{32, 33–34}. In this study, we demonstrate that QPI allows for label-free cell tracking, enabling long-term imaging on U-bottom plates. Consequently, combining QPI with conventional 96-well plates enables the prediction of highly functional HSCs. This study serves as a pivotal resource, establishing a widely accessible and generalized system for HSC functionality prediction.

Previous research focused on predicting Hematopoietic stem and progenitor cells (HSPCs) lineage choice by examining the differentiation trajectories of HSPCs through imaging during differentiation³². In contrast, by employing our single-HSC expansion technology^25,26, we successfully captured the proliferation of HSCs while preserving their undifferentiated state, unveiling the heterogeneity within the top HSC populations as distinct variations in kinetic features. Consequently, we quantitatively predicted Hlf expression levels, thereby identifying highly functional HSCs. This advancement further allowed for the quantitative prediction of the hierarchy within functional HSC populations.

In UMAP analysis using kinetic features, our ex vivo expansion system in combination with time-resolved QPI reliably assesses HSC potency, although its accuracy varies with the applied stress conditions. For example, stress induced by IL1 or gene editing promotes cellular differentiation, resulting in an overall reduction in Hlf-tdTomato expression. This decrease likely compromises the predictive power of our system, particularly for cells exhibiting intermediate levels of Hlf-tdTomato. Notably, cells with a dry mass exceeding 300 pg were exclusively observed in wild-type HSCs expanded under standard culture conditions and were absent under stress conditions. Additionally, a lower sphericity in IL1-treated cells suggests that specific cytokine signaling influences cellular morphology. Collectively, these findings indicate that both the clustering patterns and the predictive accuracy of Hlf-tdTomato expression are modulated by the applied stress conditions, highlighting the need for tailored model training for each condition to ensure accurate classification and prediction.

In experiments observing post-transplant dynamics via imaging of Hlf-tdTomato High and Low HSCs, differential behaviors were noted between the two populations; however, inter-mouse variability was also apparent. Although the transplanted HSCs were sorted as a phenotypically homogenous population by flow cytometry, the significant variability in post-transplant hematopoietic dynamics among recipient mice underscores the intrinsic functional heterogeneity within the sorted HSC population. In the present study, due to the rarity of the Hlf-tdTomato High and Low HSCs, we were limited to transplanting only 100 cells per mouse, which may have accentuated the heterogeneity within the cell population. Despite this inherent variability, mice receiving Hlf-tdTomato Low HSCs tended to exhibit earlier and more robust splenic hematopoietic activity compared to those receiving Hlf-tdTomato High HSCs. This observation suggests that Hlf-tdTomato Low HSCs may be enriched for cells capable of initiating rapid splenic hematopoiesis, potentially by giving rise to a subpopulation of stem-like common myeloid progenitors with low CXCR4 expression, as identified in our recent work³⁵. Further investigation into the mechanisms governing post-transplant hematopoiesis, particularly by identifying and characterizing the factors responsible for spatial differences in hematopoietic activity, may enable more accurate prediction and control of the diverse behaviors exhibited by individual HSCs.

QPI is a label-free technique that leverages the phase shifts of light passing through a specimen to generate high-contrast, information-rich images. Unlike brightfield imaging, which offers minimal contrast for live cells, QPI requires no fluorescent labeling, thereby minimizing perturbation and enabling long-term observation. By collecting overlapping diffraction patterns and computationally reconstructing their relative phase shifts, QPI provides quantitative data on cell thickness and refractive index. This feature is especially valuable for studying sensitive cells such as HSCs, as it allows real-time analysis of their dynamics and properties without compromising viability. However, there are several current limitations to QPI. First, extending imaging beyond our present timeframe is difficult due to the need for regular medium change (every 48–72 h) to ensure successful HSC expansion; disrupting the culture to replace media halts imaging and displaces cells, thereby preventing continuous observation. Second, the U-bottom plate design, intended to keep cells centered for long-term tracking, complicates the assessment of migration. Although this setup suits our experimental goals, it remains unclear whether HSC motility in such wells is influenced by chemoattractant pathways or is entirely random. A further challenge lies in dataset generation, particularly because QPI necessitates extended imaging durations. While imaging multiple wells increases throughput, it lowers temporal resolution. Conversely, focusing on one well offers higher temporal resolution and capacity for larger sample sizes though it demands significant time and resources. Further improvements in model performance are likely to be achieved with larger datasets. We believe that future advancements in high-throughput quantitative phase imaging microscopy will enable the generation of significantly larger datasets, ultimately unlocking the full predictive potential of our approach. Finally, the ideal direct validation of our prediction system would involve single-cell isolation and transplantation, yet no current commercial platform can perform real-time QPI alongside such downstream functional assays. Developing integrated systems for QPI-driven single-cell isolation poses a considerable technical hurdle, but future innovations may ultimately allow direct proof of concept for this approach.

Ex vivo expanded HSCs display distinct marker profiles compared to bone marrow-derived fresh HSCs^25,26, prompting the need to re-evaluate and identify optimal markers for assessing stemness in this setting. Our scRNAseq and transplantation analyses demonstrate that Hlf expression quantitatively reflects temporal changes in HSC stemness during ex vivo expansion. Using Hlf-tdTomato reporter mice, we further show that kinetic features can accurately predict Hlf expression levels. During this process, we also found an association between CD150 expression and cellular proliferation patterns. Consistent with previous findings³⁶, CD150-high ex vivo expanded HSCs retain robust long-term bone marrow reconstitution, whereas CD150-low HSCs are largely confined to short-term reconstitution. Notably, CD150-low cells undergo transient proliferation but are prone to early cell death, leading to a diminished overall yield. These observations suggest that CD150 expression is tied to specific proliferation and survival pathways in ex vivo expanded HSCs.

Our analysis revealed that cells with high Hlf-tdTomato expression displayed higher sphericity, lower dry mass, and reduced motility compared to their Hlf-low counterparts. These findings suggest that Hlf-high HSCs represent a more quiescent state, whereas a decrease in Hlf expression is associated with morphological changes indicative of increased migratory potential. This correlation between Hlf expression and biophysical properties provides a valuable framework for guiding future experimental design. Specifically, further dissecting how these morphological features relate to stem cell identity and function may enable the identification of early transitional states during HSC differentiation. For instance, the observation that Hlf-low HSCs exhibit lower sphericity, and higher velocity raises important questions about whether these changes are passive consequences of differentiation or actively driven by external cues such as chemotactic signals. Elucidating these mechanisms could lead to the development of more targeted and efficient culture systems that preserve or enhance desirable HSC traits. Ultimately, integrating time-resolved morphological profiling with molecular characterization may refine our ability to prospectively isolate and expand HSCs with superior regenerative potential.

Our findings elucidate a relationship between stemness and kinetic features in both murine and human HSCs. While Hlf-tdTomato reporter mice enable direct tracking of HSC stemness, microscopy-based stemness tracking in human HSCs remains technically challenging. Given the elevated expression of HLF in human HSCs³⁷ and recent advances in knock-in techniques for fluorescent markers, there is significant potential for future studies to explore this direct relationship.

For predicting Hlf-tdTomato expression, we employed deep learning models capable of automatically extracting features directly from the input data, thereby eliminating the need for manual feature selection. This approach allows the model to identify and utilize patterns most relevant to the prediction task without explicit feature engineering. However, a key limitation is that it becomes difficult for researchers to determine precisely which features the model uses, since it inherently captures multiple patterns from the input data.

Ensuring quality control of HSCs is imperative for the advancement of regenerative medicine and gene therapy. Unfortunately, treatment with GPH-101 (nulabeglogene autogedtemcel or nula-cel), an investigational gene therapy for the treatment of sickle cell disease, resulted in pancytopenia and prolonged low blood cell counts in the initial patients, which necessitated the suspension of the trial in January 2023. This outcome suggests compromised HSC functionality. To guarantee both the safety and efficacy of these therapies, it is essential to select high-quality, functional HSCs. Our research demonstrates that real-time analysis using QPI provides a robust methodology for quantitatively predicting highly functional HSCs, presenting a clinically significant resource for the accurate prediction of high-quality HSCs.

Methods

Mice

C57BL/6NCrSlc (Ly 5.2, CD45.2) mice were purchased from SLC Inc., Japan. C57BL/6-Ly5.1 (Ly5.1, CD45.1) mice were obtained from Sankyo Labo Service Corporation, Inc., Japan. C57B/6 albino mice were purchased from The Jackson Laboratory Inc., Japan. All mice were obtained at the age of 8-10 weeks and maintained in a specific-pathogen-free environment with free access to food and water. Housing condition were temperature 22 ± 2 °C, humidity 55 ± 5%, light/dark cycle 12 h/12 h (8 a.m.–20 p.m. light). Hlf-tdTomato mice were kindly provided by Tomomasa Yokomizo (Department of Microscopic and Developmental Anatomy, Tokyo Women’s Medical University, Tokyo, Japan.). All animal studies were conducted in accordance with institutional protocols and were approved by the Animal Care and Use Committee of the Institute of Medical Science at the University of Tokyo and the Laboratory Animal Resource Center at the University of Tsukuba.

Murine HSC isolation

Male 8–10-week-old C57BL/6-CD45.1 mice were humanely sacrificed under isoflurane anesthesia. Pelvic, femur, and tibia bones were excised and crushed, and the resulting cell solution was filtered and the whole bone marrow cells were enumerated. Positive selection of cKit+ cells was accomplished using anti-APC magnetic-activated cell sorting (MACS, Miltenyi Biotec) antibodies after staining the cells with cKit-APC antibody (Thermo Fisher Scientific) for 30 min. Enriched cKit+ cells were then incubated with an anti-Lineage antibody cocktail (including biotinylated Gr1[LY-6G/LY-6C], CD11b, CD4, CD8a, CD45R[B220], IL7-R, TER119 (all from Thermo Fisher Scientific)) for an additional 30 min. This was followed by a 90-min incubation with CD34-FITC (Thermo Fisher Scientific), Sca1-PE (Thermo Fisher Scientific), cKit-APC, streptavidin-APC/eFluor (Thermo Fisher Scientific), CD150-PE/Cy7 (BioLegend) and CD48-PB (BioLegend) antibodies. Propidium iodide (PI) was employed to exclude dead cells. CD34^-CD150⁺ CD48^-cKit⁺Sca1⁺Lin^- (CD34^-CD150⁺ CD48^-KSL) cells were sorted via fluorescence-activated cell sorting (FACS) on Aria III cell sorter (BD) using a 100 µm nozzle with appropriate filters and settings. For Hlf-tdTomato reporter mice, cKit+ cells were incubated with an anti-Lineage antibody cocktail, followed by incubation with CD34-FITC, Sca1-BV605 (BioLegend), cKit-APC, streptavidin-APC/eFluor, CD150-PE/Cy7 and CD48-PB (BioLegend) antibodies. CD34^-CD150⁺CD48^-cKit⁺Sca1⁺Lin^-tdTomato⁺ (CD34^-CD150⁺CD48^-Hlf⁺KSL) cells were sorted via FACS on Aria III cell sorter.

Murine HSC culture

Murine hematopoietic stem cells (HSCs) were cultured in Ham’s F12 medium (Wako), enriched with 10 mM HEPES (Thermo Fisher Scientific), recombinant cytokines murine TPO (100 ng/mL, Peprotech), and SCF (10 ng/mL, Peprotech), Soluplus (BASF), as well as insulin-transferrin-selenium (ITS, Thermo Fisher Scientific, 1:100 dilution) and 1% Penicillin-Streptomycin-L-Glutamine (PSG, Wako). Murine HSCs were cultured on untreated U-bottom or flat-bottom 96-well plates (TPP, for cultures starting with 100 cells). Cells were maintained in an incubator (Panasonic) at 37 °C with a constant CO₂ fraction of 5%, and medium changes were carried out every 2-3 days.

Human Umbilical Cord Blood Cells

Human umbilical-cord-blood-derived CD34+ cells were purchased from StemExpress. The products were stored at −150 °C in laboratory liquid nitrogen vessels and thawed appropriately in accordance with the manufacturer’s instructions.

Human HSC culture

Human cord blood CD34+ cell cultures were performed using IMDM (Life Technologies), 1% insulin-transferrin-selenium-ethanolamine (ITSX; Life Technologies), 1% penicillin–streptomycin–glutamine (P/S/G; Life Technologies), 0.1% soluplus (BASF), 5 µmol/L 740 Y-P (Scram), 0.1 µmol/L butyzamide (Shionogi) and 10 ng/ml FMS-like tyrosine kinase 3 ligand (FLT3, PeproTech) at 37 °C with 5% CO2. Single human HSC cell cultures were performed using StemSpan SFEM (Stemcell Technologies) supplemented with 50 ng/ml recombinant human SCF (PeproTech), 50 ng/ml FLT3 (PeproTech), 50 ng/ml recombinant human THPO (PeproTech), 50 ng/ml IL6 (PeproTech) and 1 µM SR-1 at 37 °C with 5% CO2.

Analysis of Human HSCs

Phenotypic analysis was performed by staining cells with CD34-APC (BioLegend), CD38-PE/Cy7 (BD Bioscience), Lin-FITC (CD2, CD3, CD4, CD7, CD8, CD10, CD11b, CD14, CD19, CD20, CD56 and CD235a) (BD Bioscience), CD45RA-BV510 (BioLegend), CD90-APC/Cy7 (BioLegend) and CD201-PE (BioLegend). The stained cells were then resuspended in 200 μl PBS/PI and analyzed using FACS Verse analyzer and FACS AriaIII cell sorter. Image-enhanced cytometer analysis was performed using Attune CytPix Cytometer (Thermo Fisher Scientific).

Quantitative phase imaging and analytical tools

Quantitative phase imaging was conducted with Livecyte (Phasefocus) on U-bottom/flat-bottom 96 well plate. Cells were maintained at 37 °C with a constant CO₂ fraction of 5%. Phase and fluorescence images were acquired in parallel for each well. In single-cell imaging, imaging was started approximately 8 h after sorting. In bulk-cell imaging, measurements were started approximately 15 h after sorting. Phasefocus’ Cell Analysis Toolbox software was used for cell segmentation, cell tracking, and data exportation in phase images. Segmentation thresholds were optimized using various image processing techniques, such as the rolling ball algorithm for background noise removal, image smoothing for detecting cell edges, and local pixel maxima detection to identify seed points for final consolidation. The feature tables output by the Phasefocus’ Cell Analysis Toolbox software for each imaged well were analyzed using R. For cell division pattern analysis, Trackmate ImageJ plugin was used to identify cell divisions and classify the division patterns. For the UMAP analysis based on cellular kinetic characteristics, parameters such as Volume, Thickness, Radius, Area, Sphericity, Length, Width, Length-width ratio, Dry mass, Velocity, and Perimeter were extracted from the feature table. The length and width of a feature are calculated using the smallest bounding box that encloses the pixels of the feature. The length is defined as the longer of the two axes of the bounding box. The area is determined by summing the total number of pixels within a feature’s segmented boundary and multiplying it by the pixel size of the image. The total volume of the feature is calculated by summing the optical volume of each individual pixel. Mean thickness is determined by dividing the volume by the area of the segmented feature. Dry mass refers to the total estimated non-water cellular matter within a feature. Radius is defined as the radius of a circle with an area equivalent to that of the segmented feature. The perimeter is the total length of the outer boundary of the segmented feature. Sphericity is a metric that measures how closely the surface area of the feature resembles that of a hemisphere. The velocity of the cell is calculated based on its displacement between the previous frame and the current frame. QPI analysis of HSCs under various stress conditions, including lentiviral transduction, inflammatory cytokine treatments, and gene editing for gene therapy, involved sorting 50 Lin^-cKit^-Sca1⁺CD150⁺CD201⁺CD48^-tdTomato⁺ HSCs (mNG+ cells in the case of lentiviral transduction and MHC Class I-negative cells for B2m knock out), culturing them for 36 h, and performing QPI and fluorescence imaging to extract kinetic features, excluding tdTomato fluorescence intensity.

Peripheral blood analysis

Peripheral blood was collected using a heparin tube. For chimerism and lineage analysis, erythrocytes were lysed in NH₄Cl solution. The resulting lysed blood cells were stained with Gr1-PB (BioLegend), CD11b-PB (BioLegend), CD4-APC (Thermo Fisher Scientific), CD8a-APC (Thermo Fisher Scientific), CD45R[B220]-APC/eFluor 780 (Thermo Fisher Scientific), CD45.1-PE/Cy7 (Tombo Biosciences) and CD45.2-BV421 (Thermo Fisher Scientific) for C57BL/6 mice samples. The stained cells were then resuspended in 200 μl PBS/PI prior to recording events on a FACSVerse (BD) analyzer using the appropriate filters and settings.

Cell counting and sample preparation for flow cytometry

Cell counting was carried out utilizing an automated cell counter (Countess II cytometer, Thermo Fisher Scientific). For hematopoietic stem cells in post-transplant bone marrow, cells were incubated with an anti-Lineage antibody cocktail consisting of biotinylated Gr1[LY-6G/LY-6C], CD11b, CD4, CD8a, CD45R[B220], IL7-R, and TER119. This was followed by staining with streptavidin-APC/eFluor, cKit-APC, Sca1-BV605 (BioLegend), CD150-PE/Cy7, and CD34-FITC. For sorting of ex vivo expanded HSCs after culture, cultured bulk cells were stained with an anti-Lineage antibody cocktail consisting of biotinylated Gr1[LY-6G/LY-6C], CD11b, CD4, CD8a, CD45R[B220], IL7-R, and TER119, followed by staining with streptavidin-BV421 (BioLegend), cKit-APC, Sca1-APC/Cy7 (BioLegend), CD150-PE/Cy7, and CD201-PE (Thermo Fisher Scientific) for cultured HSCs, and streptavidin-APC/eFlour, cKit-BV421, Sca1-APC/Cy7, CD150-PE/Cy7, CD201-APC (Thermo Fisher Scientific) and CD48-FITC for cultured Hlf-tdTomato HSCs. All antibodies were titrated to determine optimal staining conditions, starting from an approximate concentration of 1 μl per 10⁶ cells, and adjusted as needed for each experiment.　The stained cells were then resuspended in 200 μl PBS/PI and analyzed using FACS Verse analyzer and FACS AriaIII cell sorter. Image-enhanced cytometer analysis was performed on CD34+ human HSCs after 7 days of ex vivo expansion using the Attune™ CytPix™ Flow Cytometer.

Akaluc vector transduction

Cultured cells underwent transduction with a VSV-G pseudotyped lentiviral vector carrying an mNeonGreen-P2A-Akaluc transgene under the regulation of the human ubiquitin C (UbC) promoter at a multiplicity of infection (MOI) of 300. A medium change was performed one day after transduction, and further medium changes were performed every 2–3 days.

In vivo luminescence imaging

Mouse HSCs were transduced with the Akaluc gene as described previously³⁵. Following a 7‐day ex vivo expansion, 100 cells each of Hlf-tdTomato^high CD201⁺CD150⁺CD48^-KSL Akaluc⁺ and Hlf-tdTomato^low CD201⁺CD150⁺CD48^-KSL Akaluc⁺ populations were transplanted into lethally irradiated (8.0 Gy) mice. The mice were anesthetized with isoflurane and injected with 50 µL of TokeOni (15 mM, Kurogane Kasei Co., Ltd.) intraperitoneally. They were then placed in an IVIS in vivo imaging system (PerkinElmer). Images were taken after 5 min using appropriate binning and exposure settings. The luminescence signal intensities were analyzed with Living Image Software. To measure the luminescence signal intensity of the whole body, the ROI position for each mouse was kept consistent, and the temporal changes in average luminescence intensity at the same position were analyzed.

CRISPR/Cas9 gene editing

Gene editing for HSCs were performed as described previously²⁶. Recombinant S. pyogenes Cas9 (S.p. Cas9 Nuclease V3, IDT) was complexed with single guide RNA (sgRNA, synthesized at IDT) at a molar ratio of 1:2.5 for 10 min at 25 °C to form ribonucleoprotein (RNP) complexes. Sequences of sgRNA targeting murine B2m is mC*mU*mG*rGrUrGrCrUrUrGrUrCrUrCrArCrUrGrArCrGrUrUrUrUrArGrArGrCr UrArGrArArArUrArGrCrArArGrUrUrArArArArUrArArGrGrCrUrArGrUrCrCrGrUrUrArUrCrArArCrUrUrGrArArArArArGrUrGrGrCrArCrCrGrArGrUrCrGrGrUrGrCmU*mU*mU*rU (mN*: Phosphorothioated 2′-O-methyl RNA base; r: ribose backbone). HSCs were expanded and washed twice with PBS, pelleted, and resuspended in 20 ul electroporation buffer P3 (Lonza). The RNP duplex was gently added to the cells, and the suspension was transferred to a single 20 ul electroporation cuvette on a 16 well strip (P3 Primary Cell 96-well-Nucleofector Kit, Lonza). Electroporation was conducted using programs EO-100 on a 4D nucleofector device (Lonza). Cells were immediately recovered in pre-warmed medium and gently split-transferred into 24-well plates (Corning). One day after nucleofection, a medium change was performed, and further medium changes were performed every 2–3 days.

Transplantation assay

Cell transplantation via intravenous injection was conducted as previously reported^25,26. Cultured cells at the indicated doses were transplanted into irradiated (8.0 Gy) C57BL/6-CD45.1 recipient mice, along with 5 × 10⁵ C57BL/6-CD45.1/CD45.2 whole bone marrow competitor cells, using cultured cells at indicated cell doses. The cells were transplanted on the same day as the irradiation. Secondary bone marrow transplantations were performed by extracting WBM cells from the primary recipient and transplanting 1 × 10⁶ cells into lethally irradiated (8.0 Gy) secondary recipients. In the Hlf-tdTomato HSC transplantation experiments, CD34-CD150 + KSL cells were isolated from fresh mouse bone marrow and cultured ex vivo for 7 days. Subsequently, 100 cells each of Hlf-tdTomato^high CD201 + CD150 + CD48-KSL and Hlf-tdTomato^low CD201 + CD150 + CD48-KSL populations were transplanted. Peripheral blood samples were collected and analyzed at 4-week intervals.

Training dataset preparation through segmentation and tracking

Our methodology started by capturing single cell tracking videos with a quantitative phase microscope, which we then used as the input for our prediction model. Initially, we segmented each video frame using Voronoi Otsu labeling to identify each cell in every frame. Following this segmentation, we generated a tracking video for each cell. This step involved aligning the centroid of each cell across successive frames using the segmentation labels. We utilized Phase Focus’s live cell analysis software to create two types of videos: one that recorded cell movements within a static field of view (unchanged from the first frame) alongside their morphological changes and another that focused solely on morphological changes, cropping each frame to keep the cells centered. We manually verified every cell within the dataset to ensure data integrity.

We determined the prediction model’s output by measuring the cells’ cumulative fluorescence intensities. This process used the segmentation labels from the phase images as masks to measure the cumulative fluorescence intensity of cells in the video’s last frame, which we then used as the target for regression analysis.

Prediction model training

We employed the 3D ResNet18 model for training, coding the regression task with the Python-based library Pytorch Lightning to predict the fluorescence intensity of cells from video data. Our dataset consists of videos from three independent experiments for training, with videos from a fourth experiment set aside for validation. We conducted cross-validation within the training dataset, dividing it into three subsets using the Stratified K-Folds cross-validation method to ensure representativeness in each fold. We chose the mean square error as the loss function to optimize this model due to its effectiveness in regression tasks.

Single-cell RNAseq analysis

Murine bone marrow derived CD34^-KSL cells were cultured for 7 days and collected. Libraries were generated using Single Cell 3′ Reagent Kit v.3.1 (10x Genomics), and sequenced on Illumina Hiseq X. Sequence data were mapped to reference genome (mm10) using CellRanger. Subsequent analysis was performed using Seurat R package. We filtered out cells that had nFeature_RNA over 5000 or less than 200, and percentage of mitochondrial counts higher than 7.5%. Data was normalized and scaled using NormalizeData and ScaleData function. To reduce dimensions, Principal component analysis (PCA) and UMAP were performed using RunPCA and RunUMAP function. To identify clusters, we performed FindClusters function (resolution = 0.4).

Bulk RNAseq analysis

In our bulk RNA-seq analysis, we performed a reanalysis of data we had obtained previously²⁶. Briefly, after 7 days of ex vivo HSC expansion, the target cells were sorted into 1.5 mL tubes and lysed in 600 µL Trizol LS reagent (Thermo Fisher Scientific). Subsequently, RNA purification, library preparation, and next-generation sequencing were performed by Tsukuba i-Laboratory, LLC. Libraries were generated using the SMARTer cDNA synthesis kit (Takara) and the high-output kit v2 (Illumina), followed by sequencing on a NextSeq 500 sequencer (Illumina) with 2 × 36 paired-end reads. DESeq2 package in R was utilized for data normalization and comparative analysis.

Quantification and statistical analysis

Information concerning the statistical tests applied, the number of subjects and groups are mentioned in the figure legends. Student’s t-tests, one- and two-way analysis of variance (ANOVA) were executed using Prism (version 9.5, Graphpad). Standard deviations are represented by error bars. For statistical analysis related to RNAseq, R version 4 (R Core Team, 2020) along with the fitting package were utilized.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Acknowledgements

We thank the University of Tokyo Institute of Medical Science (IMSUT) Fluorescence Activated Cell Sorter (FACS) core laboratory for technical assistance. We thank Dr. Masafumi Muratani and Tsukuba i-Laboratory LLP for sequencing services. We further thank Dr. Martin Humphry of Phasefocus for providing expert knowledge regarding quantitative phase imaging. This work was supported by the Japan Society for the Promotion of Science (JSPS; #24K02478, #21F21108, and #20K21612 to S.Y., #24K19192 to T.Y. and #23K15315 to H.J.B.), the Japanese Agency for Medical Research and Development (AMED; #24bm1223011h0002, #23bm1223011h0001, #21bm0404077h0001, and #21bm0704055h0002 to S.Y. and JP25bm1123084 to T.Y.).

Author contributions

Conceptualization, methodology: T.Y. and S.Y.; investigation: T.Y., Y.I., H.J.B., T.K., R.I.; visualization: T.Y. and Y.I.; funding acquisition, project administration: S.Y.; supervision: T.Y., T.S., S.O., S.Y.; writing – original draft: T.Y.; writing – review & editing: T.Y., Y.I., H.J.B., A.S.F., T.K., T.Y., T.S., S.O., S.Y.

Peer review

Peer review information

Nature Communications thanks the anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Data availability

All data are available in the main text or the supplementary materials. Source Data are provided with this paper. All data have been deposited in a Mendeley Data repository and will be published upon acceptance of the manuscript (DOI: 10.17632/hzvpmrf5g3.1). The scRNAseq data have been deposited in the GEO under the accession codes GSE286255. Further information and requests for resources and reagents should be directed to and will be fulfilled by TY or SY.  are provided with this paper.

Code availability

Code for all our representation learning models is available at https://github.com/solabtokyo-org/QPI-HSCs

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-025-61846-3.

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

References

1. TILL, JE; McCULLOCH, EA. A direct measurement of the radiation sensitivity of normal mouse bone marrow cells. Radiat. Res.; 1961; 14, pp. 213-222.1961RadR..14.213T1:STN:280:DyaF3c%2FpslSgug%3D%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/13776896]

2. Visserc, JWM; Bauman, JGJ; Mulder, AH; Euason, JF; De Leeuw, AM. Isolation of murine pluripotent hemopoietic stem cells. J. Exp. Med.; 1984; 159, pp. 1576-1590.

3. Spangrude, GJ; Heimfeld, S; Weissman, IL. Purification and characterization of mouse hematopoietic stem cells. Science; 1988; 241, pp. 58-62.1988Sci..241..58S1:STN:280:DyaL1c3msVCjsw%3D%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/2898810]

4. Morrison, SJ; Weissman, IL. The long-term repopulating subset of hematopoietic stem cells is deterministic and isolatable by phenotype. Immunity; 1994; 1, pp. 661-673.1:CAS:528:DyaK2MXitlSjurs%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/7541305]

5. Osawa, M; Hanada, KI; Hamada, H; Nakauchi, H. Long-term lymphohematopoietic reconstitution by a single CD34- low/negative hematopoietic stem cell. Science; 1996; 273, pp. 242-245.1996Sci..273.242O1:STN:280:DyaK283nvFKnsg%3D%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/8662508]

6. Notta, F et al. Isolation of single human hematopoietic stem cells capable of long-term multilineage engraftment. Science; 2011; 333, pp. 218-221.2011Sci..333.218N1:CAS:528:DC%2BC3MXotlGmtL8%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/21737740]

7. Chen, JY et al. Hoxb5 marks long-term haematopoietic stem cells and reveals a homogenous perivascular niche. Nature; 2016; 530, 223.2016Natur.530.223C1:CAS:528:DC%2BC28XisVyku7k%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/26863982][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4854608]

8. Tajima, Y et al. Continuous cell supply from Krt7-expressing hematopoietic stem cells during native hematopoiesis revealed by targeted in vivo gene transfer method. Sci. Rep.; 2017; 71, pp. 1-10.

9. Ito, K et al. Gene targeting study reveals unexpected expression of brain-expressed X-linked 2 in endocrine and tissue stem/progenitor cells in mice. J. Biol. Chem.; 2014; 289, pp. 29892-29911.1:CAS:528:DC%2BC2cXhvVSmt7rE [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/25143383][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4208000]

10. Goodell, MA; Nguyen, H; Shroyer, N. Somatic stem cell heterogeneity: diversity in the blood, skin and intestinal stem cell compartments. Nat. Rev. Mol. Cell Biol.; 2015; 165, pp. 299-309. 16

11. Purton, LE. Adult murine hematopoietic stem cells and progenitors: an update on their identities, functions, and assays. Exp. Hematol.; 2022; 116, pp. 1-14.1:CAS:528:DC%2BB38XivVWnsLnF [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/36283572]

12. Rix, B; Maduro, AH; Bridge, KS; Grey, W. Markers for human haematopoietic stem cells: The disconnect between an identification marker and its function. Front. Physiol.; 2022; 13, 1009160. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/36246104][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9564379]

13. Laurenti, E; Göttgens, B. From haematopoietic stem cells to complex differentiation landscapes. Nature; 2018; 553, pp. 418-426.2018Natur.553.418L1:CAS:528:DC%2BC1cXhs12itbs%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/29364285][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6555401]

14. Donati, G; Watt, FM. Stem cell heterogeneity and plasticity in epithelia. Cell Stem Cell; 2015; 16, pp. 465-476.1:CAS:528:DC%2BC2MXntFKitbo%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/25957902]

15. Carrelha, J et al. Hierarchically related lineage-restricted fates of multipotent haematopoietic stem cells. Nature; 2018; 554, pp. 106-111.2018Natur.554.106C1:CAS:528:DC%2BC1cXitVCmsr0%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/29298288]

16. Sanjuan-Pla, A et al. Platelet-biased stem cells reside at the apex of the haematopoietic stem-cell hierarchy. Nature; 2013; 502, pp. 232-236.2013Natur.502.232S1:CAS:528:DC%2BC3sXht1CgsLjN [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/23934107]

17. Rodriguez-Fraticelli, AE et al. Clonal analysis of lineage fate in native haematopoiesis. Nature; 2018; 553, pp. 212-216.2018Natur.553.212R1:CAS:528:DC%2BC1cXntVGrtQ%3D%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/29323290][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5884107]

18. Yamamoto, R et al. Clonal analysis unveils self-renewing lineage-restricted progenitors generated directly from hematopoietic stem cells. Cell; 2013; 154, pp. 1112-1126.1:CAS:528:DC%2BC3sXhtlGru73O [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/23993099]

19. Mansell, E et al. Mitochondrial potentiation ameliorates age-related heterogeneity in hematopoietic stem cell function. Cell Stem Cell; 2021; 28, pp. 241-256.e6.1:CAS:528:DC%2BB3cXitFGrt7rN [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/33086034]

20. Vannini, N et al. Specification of haematopoietic stem cell fate via modulation of mitochondrial activity. Nat. Commun.; 2016; 71, pp. 1-9.

21. Zhou, H et al. Label-free metabolic optical biomarkers track stem cell fate transition in real time. Sci. Adv.; 2024; 10, 6770.

22. Nestorowa, S et al. A single-cell resolution map of mouse hematopoietic stem and progenitor cell differentiation. Blood; 2016; 128, pp. e20-e31.1:CAS:528:DC%2BC28XhslOktrjE [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/27365425][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5305050]

23. Loeffler, D; Schroeder, T. Understanding cell fate control by continuous single-cell quantification. Blood; 2019; 133, pp. 1406-1414.1:CAS:528:DC%2BC1MXptlSqtLg%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/30728141]

24. Kull, T. & Schroeder, T. Analyzing signaling activity and function in hematopoietic cells. J. Exp. Med. 218, e20201546 (2021).

25. Wilkinson, AC et al. Long-term ex vivo haematopoietic-stem-cell expansion allows nonconditioned transplantation. Nature; 2019; 571, pp. 117-121.2019Natur.571.117W1:CAS:528:DC%2BC1MXhtVOgtL7L [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31142833][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7006049]

26. Becker, HJ et al. Controlling genetic heterogeneity in gene-edited hematopoietic stem cells by single-cell expansion. Cell Stem Cell; 2023; 30, pp. 987-1000.e8.1:CAS:528:DC%2BB3sXhtlCnurbP [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/37385251][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10338855]

27. Sakurai, M et al. Chemically defined cytokine-free expansion of human haematopoietic stem cells. Nature; 2023; 615, pp. 127-133.2023Natur.615.127S1:CAS:528:DC%2BB3sXjslGisL0%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/36813966]

28. Rodenburg, JM; Humphry, MJ; Maiden, AM. Optical ptychography: a practical implementation with useful resolution. Opt. Lett.; 2010; 35, pp. 2585-2587.2010OptL..35.2585M [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/20680066]

29. Maiden, AM; Rodenburg, JM. An improved ptychographical phase retrieval algorithm for diffractive imaging. Ultramicroscopy; 2009; 109, pp. 1256-1262.1:CAS:528:DC%2BD1MXpvFyiu7s%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/19541420]

30. Yokomizo, T et al. Hlf marks the developmental pathway for hematopoietic stem cells but not for erythro-myeloid progenitors. J. Exp. Med.; 2019; 216, pp. 1599-1614.1:CAS:528:DC%2BC1MXhvVOjt7nE [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31076455][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6605751]

31. Wang, S et al. Deep learning-based predictive classification of functional subpopulations of hematopoietic stem cells and multipotent progenitors. Stem Cell Res. Ther.; 2024; 15, pp. 1-13.

32. Buggenthin, F et al. Prospective identification of hematopoietic lineage choice by deep learning. Nat. Methods; 2017; 14, pp. 403-406.1:CAS:528:DC%2BC2sXis1ylsbc%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/28218899][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5376497]

33. Dykstra, B et al. High-resolution video monitoring of hematopoietic stem cells cultured in single-cell arrays identifies new features of self-renewal. Proc. Natl. Acad. Sci. USA; 2006; 103, pp. 8185-8190.2006PNAS.103.8185D1:CAS:528:DC%2BD28Xlt1ChtLs%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/16702542][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC1461403]

34. Lutolf, MP; Doyonnas, R; Havenstrite, K; Koleckar, K; Blau, HM. Perturbation of single hematopoietic stem cell fates in artificial niches. Integr. Biol.; 2009; 1, pp. 59-69.1:CAS:528:DC%2BD1MXhtFaqsLY%3D

35. Yogo, T et al. Progenitor effect in the spleen drives early recovery via universal hematopoietic cell inflation. Cell Rep.; 2025; 44, 115241.1:CAS:528:DC%2BB2MXit1artb4%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/39864058]

36. Kiel, MJ et al. SLAM family receptors distinguish hematopoietic stem and progenitor cells and reveal endothelial niches for stem cells. Cell; 2005; 121, pp. 1109-1121.1:CAS:528:DC%2BD2MXmtF2ntro%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/15989959]

37. Lehnertz, B et al. HLF expression defines the human hematopoietic stem cell state. Blood; 2021; 138, pp. 2642-2654.1:CAS:528:DC%2BB38XitVekug%3D%3D [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/34499717]