Description of the UPSIDE platform

UPSIDE is designed to be a versatile machine-learning pipeline for unsupervised exploration of cell morphological states in transmitted light images, and subsequent elucidation of their transition dynamics from movies (Fig 1A, see Methods section for detailed description of the pipeline). In this pipeline, cells are first segmented using a convolutional neural network that converts brightfield images of unlabeled cells into synthetic fluorescent images of cytoplasm for segmentation [22]. This neural network is trained using a set of images of cells stained for their cytoplasm (S1 Fig). This approach allows the network to autonomously tailor its parameters, and to accommodate a wide range of different cell types in order to optimize performance without human input. Dead cells and other debris are eliminated from identified cell sub-images through a convolutional classifier trained on cells from the same data set that were manually identified as being dead or alive (S2 Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Description of the UPSIDE workflow.

(A) Single cells are segmented directly from brightfield images and deep learning UNET architecture to predict synthetic fluorescent images [23]. Segmented cells are then pre-processed to generate separate mask and texture images, which are then used to concurrently train two variational autoencoders (VAEs). The shape and texture encodings learnt by these two VAEs are then concatenated and used for downstream data analysis. (B) Encoded latent vectors are then decoded into a shape and texture image to aid the interpretation of the encoded features.

https://doi.org/10.1371/journal.pcbi.1009626.g001

UPSIDE then learns morphological features of identified live cells using a variational autoencoder (VAE) architecture. To enhance the sensitivity of UPSIDE towards learning true cell-state or cell-type defining morphological features, we incorporated two computational procedures into our analysis pipeline. First, we rotated cells to align their major axes vertically, then reflected them to ensure an identical left-right skewing, to ensure cell encodings are invariant to rotation and reflectional transformations. Second, we trained two VAEs in parallel, one that takes a binary cell mask, and another that takes a normalized grayscale image containing cell textural features (see methods). We did so to ensure that both shape and textural morphological features of imaged cells are adequately utilized for feature encoding. The learned mask and textural encodings were then weighed using a coefficient, Ω, then concatenated for subsequent clustering and dimensionality reduction. We note that Ω can be adjusted depending on application to capture the relative importance of shape and textural features to underlying cell heterogeneity.

Latent representations of cells are then clustered using the Louvain method [23] and represented on a 2D plane using the uniform manifold and projection algorithm (UMAP) [24]. Finally, mask and texture vectors are subject to decoding through variation of magnitudes of specific features or groups of features, followed by generation of synthetic images in observable image space (Fig 1B). This approach allows latent features to be visually displayed for human inspection and interpretation.

UPSIDE uncovers distinguishing morphological features between heterogeneous blood cell types

We first tested UPSIDE’s ability to learn cell type-defining morphological features in a mixed dataset consisting of multiple blood cell types. To test the limits of UPSIDE’s ability, we chose four cell types that, despite having distinct size, shape and textural features, were similar in their gross morphologies (Figs 2A and S3A): a mouse T cell leukemia line (Scid.ADH2), a mouse macrophage cell line (Raw246.7), a human acute myeloid leukemia cell line (Kasumi-1), and primary patient-derived human acute myeloid leukemia stem cells (CD34⁺CD38^- AML LSC). Brightfield images from each cell population were captured, and cells were segmented using the neural network described above. Image crops of segmented cells were then mixed together and encoded into the latent space using UPSIDE’s VAE (S4A Fig). To quantify UPSIDE’s performance, we devised a cell type homogeneity score, which reflects how closely cells of the same type cluster together in their latent space (see Methods section). We ran the VAE for this dataset for different values of Ω, to optimize the relative weights of the learned mask and texture encodings to achieve a maximal mean homogeneity score across the four cell types (S5A and S5B Fig and Methods section). This parameter specifies the relative importance of shape and texture to phenotypic state identification, and thus provides the ability to adapt UPSIDE to analyses of cell types and questions that rely on different types of distinguishing features. To compare the performance of the VAE to other deep learning methods, we repeated this analysis with several alternative architectures such as a vanilla autoencoder (AE) [25], an adversarial autoencoder with latent dimension encoding trained to fit a normal distribution or mixed gaussian distribution [14] (1xAAE and 4xAAE, respectively), and the ClusterGAN architecture [26] (ClusGAN) (see Methods section).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. UPSIDE distinguishes morphologically-distinct blood cell types in a heterogeneous population.

(A) Images of four different blood cell types were mixed together and passed through the UPSIDE workflow. Resultant shape and texture images were used to train concurrent VAEs. Output latent encodings were weighted relative to each other, concatenated, then projected onto a 2D plane using UMAP. (B) Dot plots show distribution of each cell type projected on 2D UMAP space made by UPSIDE. (C) 2D UMAP projection of the VAE-generated encodings that have been grouped into different morphological clusters using Louvain clustering algorithm. Representative brightfield cell crop images from the different clusters were listed. Scale bar represents 5 μm. (D) Cell type fractional composition within each cluster. A fixed number of cells from each cell type were sampled, and the cluster-wised cell type composition was calculated from this pooled population.

https://doi.org/10.1371/journal.pcbi.1009626.g002

The VAE outperformed these other approaches, generating approximately 6% higher homogeneity scores compared to the adversarial autoencoders, 9% higher compared to PCA, and 26% higher than the ClusterGAN architecture (S5C Fig). Adversarial autoencoders performed better than the vanilla encoder though worse than the VAE, possibly because of difficulties in training the discriminator to fit the latent encoding to the desired distribution perfectly. Surprisingly, the ClusGAN architecture performed the worst, likely due to an inability to consistently generate direct, regularized encoded representations. Despite its superior performance compared to other architectures, the VAE achieved a maximal homogeneity score that was still significantly less than unity (~0.7, S5C Fig), indicating some degree of cell type mixing in latent space after learning. This was not unexpected given that some cell types – particularly AML LSCs and SCID.adh2C2 cells – appear visually similar or even indistinguishable in some cases, and cannot be separated by morphological characteristics observed from brightfield images alone. Nonetheless, these comparisons suggest that the VAE architecture is particularly well suited for learning morphological features for cell type discrimination.

To further visualize and analyze the representation of cells in latent space, we projected the encodings from the VAE into two dimensions using the UMAP algorithm [24] (Fig 2B). From the UMAP projection, we found that the cell types largely segregated into distinct regions in this two-dimensional space (Fig 2B). Raw264.7 macrophages occupied a region that was largely distinct from regions occupied by other three cell types, reflecting their markedly different size and shape distribution. The three other cell types occupied partially overlapping regions, reflecting greater similarities in morphology among these cells (S3A Fig). Interestingly, primary human AML stem cells (identified by their CD34⁺CD38^- surface marker phenotype) overlapped parts of the Scid.ADH2 region, suggesting some of Scid.ADH2 cells look quite similar to their AML counterparts. Despite these overlaps, there are substantial areas in the two-dimensional space occupied by these regions containing only one cell type, indicating the presence of morphological features that distinguish each of these three cell types from another and allow them to be identified in mixed populations.

To understand the morphological features driving cell type separation in this latent space, we clustered cell representations in the latent space using the Louvain method, then visualized cells and the morphological attributes that defined each cluster. Eight clusters were identified, with each enriched for different cell types (Figs 2C, 2D and S3B). Clusters C1-3 were highly enriched for Raw264.7 macrophages, phagocytic cells that are larger than their progenitor cells. Clusters C4 and C8 were highly enriched for Kasumi-1 cells, circular profile cells that contain dark granules, a unique distinguishing, observable feature of these cells. Cluster C5 was enriched for Scid.ADH2 cells, which are also circular, but lacked granules. Clusters C6 and C7 were enriched for both LSCs and Scid.ADH2 cells, both of which were small and lacked granules. Cells in Cluster C7 have darker interiors and less well-defined cell boundaries compared to Cluster C6 cells, indicating that they are flatter and may be more substrate-adherent. The morphological differences within these clusters indicate the existence of distinct morphological sub-states within individual cell types.

To interpret and visualize the morphological features that separate cells into distinct groups in latent space, we performed hierarchical clustering on the averaged latent space representation for cells from different clusters (Fig 3A). This analysis revealed that each morphological cluster of cells is associated with a specific set of latent features, with magnitudes that are higher than population average. To decode these latent features, we transformed them back into synthetic images in visual space (Fig 3B and 3C, top). First, we generated a mean mask or texture vector by averaging over all cells in the dataset. From these mean vectors, we then selectively increased the magnitudes of the feature (or groups of features) of interest to generate a new vector. Using the VAE decoder module, we then transformed the feature-exaggerated vector and the mean vector into synthetic images for interpretation.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Cell-type specific morphological features can be interpreted by decoding the latent space cell representation.

(A) Clustergram of average z-scores for latent shape and texture features for different cell clusters (see Methods section for how z-scores values were calculated). (B) Five mask features with highest z-scores for each morphological cluster are decoded and visualized. Clustergram shows matrix of correlation coefficients for forty mask features having the highest standard variation in the dataset. Scale bar represents 5 μm. (C) Individual texture features were clustered into eight groups (T1-T8) according to their correlation with each other from the clustergram of texture features. Each group was decoded into brightfield difference images for interpretation (see Methods). Scale bar represents 5 μm.

https://doi.org/10.1371/journal.pcbi.1009626.g003

We first examined the synthetic decoded images from the five most enriched mask features for each morphology-defined cluster (Fig 3B, bottom). Clusters C1-4 contained large cells with large, round profiles. As expected, cluster C2 contained elongated cells with numerous elongated mask features. Clusters C5-8, in contrast, contained smaller cells enriched in features representing small, round profile cell shapes. These mask features are in general agreement with sizes and shapes for cells found within individual clusters (Fig 2C).

Unlike mask latent features, individual texture features in the latent space were not readily interpretable for this dataset. However, because magnitudes of the projections along individual texture dimensions correlated strongly with each other in distinctive groups (Fig 3C insets) – in contrast to those for individual mask dimensions (Fig 3B, inset) – observable texture features in the image space might be encoded not by individual latent features, but by groups of correlated latent features. Therefore, to visualize these observable features, we generated feature-dominated vectors by concurrently increasing groups of correlated latent features. We also calculated images representing the difference between feature-exaggerated decoded images compared to the mean texture decoded image for better visualization. From these synthetic difference images, we observed two overall texture pattern groups: one with darker cell interiors (T3, T6), indicative of a flatter morphology; and the other with lighter cell edges (T1, T2, T4, T5), indicative of a less flattened morphology (Fig 3C). The darker cell interior feature group is strongly enriched in Clusters C7 and C8, while the lighter cell edge group is significantly present in Clusters C1 and C3. Clusters C2, C4, C5, and C6 appear to have moderate enrichment in all these groups.

Taken together, these results reveal how UPSIDE separates cells into distinct morphological clusters by their size, shape and distinct textural features. This ability can be seen readily in Clusters C3 and C4, where cells of similar size and profile can be discriminated based on their cell edge texture features. Cells with similar textural features can also be discriminated using other features, e.g., Custers C7 and C8 are both enriched with dark cell interior textures, but differ in size with Cluster C7 cells larger on average than those in Cluster C8. These results demonstrate that UPSIDE can generate meaningful learned morphological features in an unsupervised manner, and these features can be effectively decoded into images to aid interpretability. This ability allows UPSIDE to extract valuable morphological properties by simply observing cells over time without prior manipulation or human annotations.

UPSIDE uncovers morphologically distinct cell states in patient-derived leukemic cells

LSCs play critical roles in AML disease propagation and drug resistance [27,28]. LSC and other AML cell subpopulations are typically identified and characterized by a combination of cell staining for granule content and cell surface markers as well as by their gene expression signatures [29,30]. All of these classification approaches can be further extended by transmitted light imaging and analyses to provide complementary information about leukemic cell types and states that is not readily obtainable through more conventional classification approaches. In particular, live cell movies that resolve phenotypic states over time and in response to pharmacological treatment could provide unique insights into cellular heterogeneity and responses that could better inform therapeutic decision-making.

Towards this end, we employed UPSIDE to profile primary human LSCs cultured under cytokine conditions promoting expansion and differentiation, and filmed using brightfield imaging (Fig 4A, left). We directly isolated CD34⁺CD38^- leukemic stem cells from an adult AML patient [27,31,32]. This population is associated with chemotherapy resistance and persistent disease³², and likely serves as a reservoir of drug resistant cells that fuel relapse after chemotherapy treatment. To profile the self-renewal and differentiation dynamics of these sorted cells, we then cultured LSCs with either IL-6 and thrombopoietin (TPO) to induce differentiation, or with Aryl hydrocarbon receptor inhibitors (AhRi) UM729 and SR1 to maintain stemness and suppress differentiation [33–35]. We then filmed these cells in the brightfield channel for ~4 days at high temporal resolution (3 minute intervals, Fig 4A). To determine the association between observed cell morphological states, stemness and differentiation, we also added fluorescently-labeled anti-CD34 and anti-CD38 antibodies in culture, and took fluorescent images every hour to follow expression of these markers in imaged cells (Fig 4A, top right). Such in situ antibody labeling allows real-time visualization of cell surface marker protein expression with minimal effects on cell viability [36]. UPSIDE is well-suited to facilitate these types of time course analyses and image-based profiling: the use of brightfield imaging obviates the need for genetic engineering of fluorescent reporters to allow a wider range of analyses to be performed on primary patient-derived cell samples. It also minimizes cellular phototoxicity, thus enabling long-term cell observation at high temporal resolution.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. UPSIDE identifies stem cell-associated morphological states from patient-derived AML leukemic cells.

(A) LSCs (CD34⁺CD38^-) from an acute myeloid leukemia patient were cultured in cytokines with or without AhR inhibitors (UM729 and StemRegnin1) filmed for ~5 days (left). Brightfield images were captured once every 3-5 minutes. αCD34-APC and αCD38-PE antibodies were added in situ, and fluorescent images were captured once every hour (top right). Still frames show representative time lapse images of AML cells (bottom right). Scale bar represents 10 μm. (B) UMAP 2D projection of the UPSIDE generated latent space cell representations. Individual morphological clusters were identified using the Louvain Clustering algorithm, then grouped manually based on their proximity to each other in the latent space (See S6B Fig). Representative cell images from each cluster were also shown. Scale bar represents 10 μm. (C) Clustergram shows Z-scores of the latent mask and texture encodings for each morphological cluster. (D) Decoded images of the four most enriched features for each morphological state. Texture features were visualized using difference maps that were zoomed in around the decoded cells. (E) Area, eccentricity, and edge strength for each cell were calculated and mapped to the UMAP latent space representation. (F) CD34 and CD38 levels were mapped onto the UMAP. (G) Violin plots show distributions of CD34 and CD38 expression levels in different morphological clusters (left). (H) Histograms showing log CD34 levels against CD38 levels for each morphological cluster (right).

https://doi.org/10.1371/journal.pcbi.1009626.g004

Resultant timelapse images revealed considerable heterogeneity in the morphologies of observed cells, with these cells differing in their sizes, shapes and textures, as observed by their contrast from transmitted light images (Fig 4A, right; S1 Movie). To better understand this morphological heterogeneity, we fed these images into the UPSIDE pipeline. To minimize batch effects, we analyzed cells from all time points both with and without AhRi treatment in a single encoding run. We segmented cells, then encoded using UPSIDE’s separate shape and texture VAEs (S4B Fig). We then clustered representations of cells in latent space using the Louvain method (S6 Fig). Based on proximity in the latent space, we further combined some of the clusters we obtained into larger clusters. To visualize these clusters in two-dimensional space, we then projected these latent representations onto a two-dimensional plane using the uniform manifold approximation and projection algorithm (UMAP). This projection revealed the locations of the discrete clusters, along with their overlap regions. In this two-dimensional visualization, some cell clusters showed considerable boundary overlap with others, reflecting the continuous nature of the latent features encoded by the VAE. Additionally, imaged cells occupied all clusters both with or without AhRi treatment (S7 Fig), though the distribution of cells in these populations varied between the two conditions, likely reflecting the result of AhRi treatment on the cellular state.

In order to gain insight into the features that drive the separation of the cell encodings into distinct clusters, we performed hierarchical clustering on averaged cell encodings from each group (Fig 4C), then decoded the specific mask or textural features with the highest z-scores in each group to generate feature-exaggerated synthetic images, as described above (Fig 3B and 3C). These synthetic images highlight significant morphological features that display coherence across cells within a cluster, but differ between cells in different clusters (Figs 4D and S7 and S8). Examples of coherent morphological features include size, with some having smaller cells (Clusters C1,C2) and others having larger cells (Clusters C6,C7,C8); cellular elongation or eccentricity, with some clusters displaying rounder cell profiles (Clusters C1,C8) and others have more elongated cells (Clusters C2 and C7); and the presence of finer morphological features, such as cytoplasmic protrusions from the cell body (Clusters C6 and C7). Another important morphological feature was the degree of contrast at the cell edge observed from brightfield images, an indication of the degree of cell flattening on its culture surface. Some clusters had cells with weak contrast at edges (C1, C2 and C6), indicative of a flattened morphology, whereas others had cells with stronger edge contrast (C3, C8), consistent with a rounder, less-flattened appearance (S7 Fig). Together, these results reveal the distinct, defining morphological features of cells that are found in different morphological states.

To verify that these differences in decoded features indeed reflect systematic morphological differences between cells in different clusters, we calculated cell area, eccentricity, and edge strength – defined by the maximum value of the cell’s gradient image – and then plotted these quantities onto the 2D projection of the latent space (Fig 4E). Indeed, regions occupied by the different cell clusters had area, eccentricity, and edge strength values consistent with what was generally observed in the decoded cell images: Clusters C1-3 resided in the region with small cell areas, whereas Clusters C6, C7 and C8 resided in the region with larger cell areas. Elongated cells in Clusters C2 and C7 resided in regions with high eccentricity, whereas cells with darker cell edges in Clusters C3 and C4 resided in regions with high edge strength. Together, this analysis shows that UPSIDE can elucidate defining shape and textural features of cells that can vary across a population.

Distinct morphological states are associated with different degrees of stemness

Cells in the different morphological states identified above may exhibit different degrees of AML cell stemness or differentiation. To test this idea, we investigated the relationship between these morphological states and the expression levels of CD34 and CD38, which together specify the stemness of these cells. To do so, quantified CD34 and CD38 expression levels for each cell, and mapped them onto the 2D projection of AML cell’s learned latent dimensions (Fig 4F and 4H). We also generated 2D heatmaps of these markers on a log-log axis, both for cells within individual clusters and for all cells (Fig 4H, right; and S8B Fig, left, respectively). From this analysis, we identified morphological clusters enriched for cells in either a stem cell state (CD34⁺CD38^-) or a more mature state (CD34^-CD38⁺). Specifically, morphological clusters C1 and C2 were enriched for the stem cell population (CD34⁺CD38^-). Cells in these clusters differed in their roundness, but were uniformly small, consistent with the quiescent nature of the stem cells. Remarkably, these cells were also flat on the 2D culture surface possibly reflective of the high expression levels of cell adhesion proteins in hematopoietic stem cells [37]. In contrast, morphological clusters C6, C7 and C8, were most enriched for the mature cell population (CD34^-CD38⁺). These cells were uniformly larger, consistent with the larger size of more differentiated cells, but had varying shapes and degrees of flatness. This diversity in shape and substrate adherence likely reflects the multiple differentiated cell types that can emerge under these in vitro differentiation conditions. Consistently, when we gated cells in either a stem state (CD34⁺CD38^-) or a mature state (CD34^-CD38⁺) and analyzed their morphological cluster composition, we found that the more immature (CD34⁺CD38^-) cell population had higher C1 and C2 cell fraction compared to the mature state (CD34^-CD38⁺), but a lower C6, C7 and C8 cell fraction (S8B Fig, right). Together, this analysis reveals distinct morphological features associated with different degrees of stemness or maturity in AML.

Population dynamics of cell morphological states

To gain insight into the population dynamics of cells in different morphological states, we examined how the numbers of cells in different clusters evolve over time, both with and without suppressing differentiation with AhR inhibitors (AhRi; Figs 5A and S9A). In the absence of AhRi, cell clusters enriched for stem cells states (Clusters C1 and C2) are progressively depleted, while those enriched for more mature cells (Cluster C8) expand, consistent with the maturation of LSCs into more differentiated cells over time. As expected, AhRi treatment increases the sizes of the clusters enriched for stem cells (C1 and C2) relative to untreated conditions, while decreasing the sizes of the clusters enriched for mature cells (Cluster 8). This reflects the known effects of AhRi in maintaining stem cell self-renewal. The fractions of cells in intermediate morphological states (Clusters C3-7) remain largely unchanged, regardless of the presence of AhRi, suggesting that AhR inhibition may affect intermediate state transitions without driving specific outcomes. At the same time, CD34 levels decrease whereas CD38 levels increased over time, with both these changes becoming less pronounced with the addition of AhRi (Figs 5B and S9B). Together, these results provide insights into the population dynamics of LSC self-renewal and differentiation, and how these dynamics are affected by pharmacological compounds that modulate self-renewal.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Population dynamics of identified morphological states.

(A) Population fraction dynamics over time for each morphological cluster with (right) or without (left) AhR inhibitors (top). Population fraction contribution of each cluster at the last time point of the culture (bottom). Comparisons of the population fraction with and without AhR inhibitor were performed using the Chi-Square test for the dependency between the AhRi treatment and a cell’s cluster identity. **: p < 0.001, *: p <0.05 (B) UMAP showing CD34 and CD38 expression levels at different time points, in the presence or absence of AhRi.

https://doi.org/10.1371/journal.pcbi.1009626.g005

Parallel experiments were used to further explore morphological changes that reflect LSCs maturation. We cultured LSCs (CD34⁺CD38^-) in parallel with live imaging experiments, and analyzed them after three days for the expression of CD34, CD38, and CD123, another common LSC marker [38] (S9C Fig). Compared to untreated samples, cells treated with AhRi show higher expression of CD34 and CD123. On the other hand, CD38 expression magnitude was higher in the untreated sample, indicating greater differentiation in this population. Of note, a population of cells expressed both CD38 and CD34; this result indicates that expression of these markers may not be mutually exclusive. Parallel live imaging experiments of AhRi-treated cells showed slower expansion of large round cell morphology clusters compared to their untreated counterparts (Figs 5A and S9A). Together, these results demonstrate that distinct cell morphological states identified using UPSIDE indeed reflect leukemic cells in different states of maturation.

Inference of morphological state transitions by cell linkage analysis

The high temporal resolution of the brightfield movie analyzed above enables tracking of individual cells from frame-to-frame and, in conjunction with UPSIDE, inference of the rates at which leukemic cells transition between different morphological states. Notably, this analysis enables inference of rates without generation of long cell trajectories, which is particular challenging due to rapid cell motility of AML cells under investigation. Here, we develop an analysis routine to automatically infer transition rates from brightfield movies. In particular, this UPSIDE-enabled analysis obviates the need for generating individual cell tracks, which are typically error-prone and require considerable manual intervention. We paired cells from adjacent frames together, based on their close proximity. We then identified the morphological states of linked cell pairs using the VAE above (see Methods section), and calculated state transition probabilities based on the frequencies of linked cell pairs with specific initial and final morphological states (Fig 6A). By repeating this calculation over all possible pairs of morphological states, we obtained a matrix, describing the transition probabilities between different morphological states (Fig 6B, left).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Calculation of morphological state transition probabilities by cell linkage analysis.

(A) Cell pairs found in proximity across on successive time points were linked (left). Cell linkages, along with assigned morphological states of linked cells, were used to calculate transition probabilities between all states. (B) Heatmap shows transition probability matrix between all morphological clusters (left); image montage shows representative cell tracks identified from the culture (right). Scale bar represents 10 μm. (C) Distribution of cellular velocity for linked cells for each morphological cluster. (D) Plot shows mean cell velocity against mean cell eccentricity for each morphological cluster.

https://doi.org/10.1371/journal.pcbi.1009626.g006

This analysis revealed that cells largely transition between different morphological states in a highly dynamic and reversible manner, but exhibit state transition preferences that reflect underlying longer-term differentiation trends. Transition probabilities generally range between ~0.05 to 0.5 (/3 minute frame) (S1 Table), implying transition timescales of tens of minutes or less. However, these probabilities are not uniform, but are higher for specific transitions, such that the numbers of cells in different states changed progressively over time. Specifically, cells in the most immature states (C1 and C2), as defined by CD34 and CD38 expression (see above), tend to either inter-convert within these states, or switch preferentially to some of the more intermediate states (C5 and C6). In contrast, cells in the more intermediate states (C3, C4 and C5) tend to switch to one of the more mature states (C6, C7 and C8), though they also transition back into the an immature state with lower probability (e.g. C3 → C1; C5 → C2). Interestingly, intermediate states appear to inter-convert in a more dynamic manner, such that probabilities of maintaining the same state over time were lower for these intermediate states than they were for either the immature (C1, and C2) or more mature (C6, C7 and C8) states. Finally, cells having the most mature states (C6, C7 and C8) tend to stay within these states, and more stably maintained these states compared to other cell populations. A similar maturation trajectory was observed when transitions were visualized with averaged vector fields on a UMAP projection of the latent space (S10C Fig). These vector fields emanated from the most immature states (C1 and C2), flowed through the intermediate states (C3-C5), then converged upon the most mature states (C6-C8). Moreover, the magnitude of these vector flows are higher within the immature and intermediate states (e.g. C1, C3 and C4) but lower for the mature states (C6 and C8), consistent with the stability of these states observed in the transition matrix. Differences in trajectory stability were also directly observed when following single cells; mature cells (those starting in state C7 and C8) tend to remain in the same state (Fig 6B right, tracks 1 and 4), whereas more immature cells (C1, C3 and C5) are highly dynamic, switching from one state to another rapidly between successive frames (Fig 6B, tracks 2 and 3). Together, these data indicate that cells transition rapidly and reversibly between alternate morphological states, but do so in a biased manner, such that they progress from more stem cell-associated states (C1 and C2) toward more mature cell states (C8).

What is the basis of these dynamic morphological state transitions? The AML cells we filmed move rapidly on the culture surface (S1 Movie), and polarize as they move; thus some of the observed morphological state transitions could reflect transitions from a stationary to a motile state. To test this hypothesis, we derived the instantaneous velocities of cells in different morphological states, by calculating the displacement between successive frames for each state (Fig 6C). From this analysis, we found that cells with elongated morphologies, such as those in states C2, C4, C5, and C7, show higher movement velocities compared to other states. Consistently, there was a strong correlation between instantaneous velocity and cell eccentricity, averaged over all cells in individual clusters (Fig 6D). Thus, though morphological transition probabilities are biased by differentiation, the transitions themselves involve the rapid switching between stationary and mobile states (Fig 6B, left, tracks 2-3).

If morphological state transition probabilities calculated above are biased by the directionality of cell differentiation, they would change when cells are subject to perturbations that modulate stem cell differentiation and self-renewal. To test whether this was the case, we repeated this state transition analysis for cells treated with AhR differentiation inhibitors (S10A Fig and S1 Table). In the presence of AhRi, there is a decrease in the transition probabilities into the more differentiated states (C6, C7 and C8), together with an increase in the transition probabilities into and amongst intermediate cell states (C3, C4 and C5). This suggests that AhRi blocks movement from a stem cell-associated morphological state to one associated with mature cells by stalling cells in an intermediate state. Consistently, single cell track analyses showed that cells often transition rapidly from an intermediate state (C5) to more mature state (C8) without AhRi, but stall at an intermediate state (C5) when AhRi is present (S10B Fig). Thus, these observations show that changes in cell differentiation dynamics can influence the probabilities at which cells transition between different morphological states.

Ethics statement

Human blood and bone marrow samples were obtained by written informed consent on a protocol approved by the University of Washington/Fred Hutchinson Cancer Research Center Cancer Consortium Institutional Review Board. The samples were deidentified in the laboratory.

Experimental techniques

Image analysis

UPSIDE computational pipeline is designed to analyze the morphological diversity of cells from timelapse brightfield images. The method consists of four main modules: 1) Label-free prediction, 2) Image segmentation, 3) Live cell classification, and 4) Unsupervised feature learning. The following section describes each of the modules in further details.

Algorithms and quantitative analysis