High-content imaging of primary chronic lymphocytic leukemia cells predicts patient cohorts with distinct cellular drug responses

Summary

Cancer precision medicine benefits from identifying biomarkers that can predict therapy response. However, within a population of chronic lymphocytic leukemia (CLL) patients, there is heterogeneity that is inherent to the disease and also between patients. This heterogeneity, usually explained at the level of genetic and epigenetic abnormalities, obscures conventional potential biomarkers. As an alternative, confocal microscopy of live primary CLL patient samples in a microenvironment model that mimics proliferation centers was used to identify morphological features that define cellular phenotypes that can be used as alternative biomarkers. Applying machine learning to micrographs of 133 patient samples revealed five stable patient clusters, not discernible by standard clinical methods. Within clusters, CLL patient samples responded similarly to drugs, suggesting that live cell imaging could be used to stratify patients and predict drug responses for rational treatment design.

Graphical abstract

Highlights

• •Morphological features identified by live cell painting define cellular phenotypes
• •Cellular phenotypes identify CLL patient cohorts with common drug responses
• •Cell stress predicts patient cells that will be killed by ruxolitinib and venetoclax

Cancer; Methodology in biological sciences

Introduction

Chronic lymphocytic leukemia (CLL) remains the most common type of adult leukemia in the Western world with a median age of diagnosis of 72 years.¹ Characterized by the accumulation of CD5-positive B cells in the blood, bone marrow, spleen, and lymph nodes, CLL patients present marked differences in their clinical behaviors. In addition to well-documented chromosomal abnormalities,² recent work on the characterization of the genetic landscape of CLL revealed more than 40 driver genomic aberrations in DNA damage repair, RNA processing, and MAPK signaling.^3,4

Only symptomatic patients are treated, while others are followed by active monitoring until evidence of progression. Treatment decisions involve an assessment of disease risk, using the conventional clinical staging systems or more recently, the CLL International Prognostic Index (CLL-IPI) that incorporates the prognostic values of genetic aberrations.^5,6 While these indices accurately predict disease severity, they do not inform on drug responsiveness of the cancer. Chemotherapeutic agents such as chlorambucil and its purine analog fludarabine have been the front-line choice for CLL for decades.⁷ In recent years, there has been a rapid expansion of agents that target pro-survival signaling in CLL cells and their surrounding protective microenvironment. Kinase inhibitors such as ibrutinib, which target B cell receptor signaling, have proven effective, but resistance eventually emerges.⁸ The Bcl-2 antagonist venetoclax engages the cell apoptotic machinery, and while effective, it has also not proven curative as a single agent.⁹ In contrast to the plethora of markers identified associated with disease progression, very few biomarkers are actionable for predicting patient drug responses.¹⁰

Previously we used primary cells from CLL patients to identify kinase inhibitors that exhibited increased cell killing when combined with the Bcl-2 inhibitor venetoclax. Despite screening a wide variety of combinations, no single combination was maximally effective at enhancing venetoclax-induced cell death in all the patient samples.⁹ The most likely explanation for this result is the complexity of the disease genetic landscapes. Therefore, there is an unmet need for a method to determine a priori which drug combination is likely to be the best for any individual patient. Using short-term cultures of primary patient cells, functional apoptosis assays identified two subgroups of CLL with different dependence on extracellular matrix signaling and sensitivities to FLT3 inhibition.¹¹ We hypothesized that a more actionable approach to patient precision medicine might be facilitated if, instead of relying on genetic markers, it was possible to identify phenotypically defined patient cohorts with predictable drug responses that were characterized beyond cell viability assessments. In this study, we used a microenvironmental model with high-content live cell microscopy of cells stained with non-toxic fluorescent dyes (live cell painting)¹² to identify five such cohorts within 133 ex vivo samples from CLL patients who were not on active treatments. The patient clusters, identified by machine learning, were not recognizable by conventional clinical staging systems. Nevertheless, patient samples from the same cluster responded to drugs similarly, even for drugs that differed in mechanism of actions. Analysis of the patient clusters with distinct cell phenotypes identified a patient cohort that was highly responsive to JAK inhibition in combination with the Bcl-2 inhibitor venetoclax. In contrast, another cohort responded to avadomide and venetoclax in combination, while still another was resistant to both treatment regimes. Thus, the identification of patient cohorts based on the phenotypic profiling of naive patient samples may equate to a predictive drug response marker useful to guide patient assignment to clinical treatment choices.

Results

High-content imaging of ex vivo primary CLL patient cells

The tumor microenvironment modifies CLL cell responses to therapeutics, contributing to selective pressures that drive drug-resistant cancer cells that eventually lead to disease relapse. In order to simulate niche effects, we and others demonstrated enhanced drug resistance in an in vitro CLL proliferation center model by culturing primary patient cells in serum-free conditions with interleukin (IL)-2 and the Toll-like receptor 7/8 agonist resiquimod.^9,13,14,15 Here, the in vitro model was further complemented with the addition of IL-4 to confer resistance to ibrutinib and enhance B cell receptor signaling.^9,16,17 Patient cells incubated in this media are referred to here as 3S CLL.

To determine if patient and disease heterogeneity was manifest in alterations in cellular morphology, CLL cells were isolated from patients who were not currently being treated for the disease, incubated in 3S media for 72 h, then stained with a fluorescent live cell dye ChromaLIVE and the nuclear stain Hoechst 33342, and imaged by automated confocal microscopy. ChromaLIVE is a non-toxic dye that reports on mitochondrial energetics (Figure 1, yellow) and stains multiple intracellular membranes of viable and dead cells (Figure 1, red), while Hoechst 33342 stains DNA (blue). In particular, ChromaLIVE exhibits both intensity shifts and staining pattern changes that reflect different physiological states of the cell.¹² Importantly, it has been applied in a method termed live cell painting,¹⁸ which facilitates the study of drug mechanisms of action and pharmacokinetics, in contrast to conventional cell painting approaches that typically rely on chemical fixation and washing steps that are difficult in 1,536-well imaging plates and that push non-adherent spherical CLL cells into the corners of the wells.¹² This optimized dye combination enables simultaneous labeling of multiple cellular components in live cells, without perturbing normal physiology. Live cell painting of the CLL patient cells revealed a variety of visually distinct cellular morphologies, (Figure 1A, identical shapes indicate visually similar morphologies, while the pentagon indicates a cluster of similar cells).

Figure 1.

Patient CLL cell heterogeneity

(A) 3S CLL cells stained with ChromaLIVE (red, yellow) and Hoechst 33342 (blue) were imaged by spinning disk confocal microscopy. Symbols indicate cells with visually similar morphologies. Scale bars, 100 μm.

(B) Cells stained as in (A) were treated with PDB (phorbol-12,13-dibutyrate, protein kinase C activation), 2DG (2-deoxyglucose, quiescence), LEN (lenalidomide, proteasome specificity), TN (tunicamycin, ER stress), DID (staurosporine + venetoclax, apoptosis), or DMSO control. Drug treatments induced distinct morphological phenotypes compared to control. Scale bars, 100 μm.

To facilitate comparison with compound-treated cells, control CLL cells were incubated in 3S media and an equivalent amount of DMSO to that added with compounds. No differences in cell morphology were detected in response to DMSO. The 3S control cells appeared larger than the unstimulated (UNS) cells, reflected in a larger nuclear area measured in control DMSO samples (Figures S1A–S1C). Cells in 3S media were also more metabolically active (ChromaLIVE 561, yellow; Figures S1A, S1B, and S1D). As expected from our previous results in the 2S system without and with added IL-4, 3S CLL cells were on average more resistant to chlorambucil and venetoclax, as measured by either average nuclear area (Figure S1C, chlorambucil, nuclear condensation is symptomatic of impending cell death) or loss of mitochondrial transmembrane potential (Figure S1D, venetoclax, ChromaLIVE 561, yellow) indicating metabolic shutdown. To have the best chance of capturing previously uncharacterized phenotypic changes in cells in a microenvironmental model of the niche (proliferation center) that confers therapy resistance, we focused on the analysis of cells in 3S media.⁹

Reducing patient sample heterogeneity into defined cell states using high-dimensional phenotypic features

Analyses of micrographs of CLL cells showed that simple morphological features, such as nuclear area or mitochondrial transmembrane potential, were sufficient to demonstrate drug resistance in 3S CLL cells, where therapy response was primarily assessed based on cell viability (Figures S1C and S1D). However, no single feature-based measurement is sufficient to depict a broader range of physiological states that may further inform drug mechanisms. To overcome this, we developed a novel image-based random forest classifier by training it on 3S CLL cells treated with carefully curated control drugs. The lack of ground truth has been a major limiting factor in the field; that is, the true relationships among perturbations (e.g., genes and compounds) are unknown especially in heterogeneous cancer cell populations. By testing a set of tool compounds with known mechanisms of actions over a mixed population of patient samples, we optimized the conditions to induce different cellular states in 3S CLL cells (Figure 1B). Treatment with control tool compounds perturbed 3S CLL cells from the untreated phenotypes observed in DMSO controls and adopted distinct phenotypes; however, within the same treatment condition, the inherent heterogeneity across the patient sample population was not eliminated (Figure 1B). Training with these mixed populations reduced the emphasis on features that varied due to the inherent heterogeneity. Image-based features were extracted from micrographs of over 2,000 cells per sample for 40 selected patient samples, representing a diverse background including male and female patients, varying IPI scores, and immunoglobulin heavy chain (IGHV) mutation status. To demonstrate visually the extent to which the phenotypes were separated by the 38 most differentiating image features (Table S3), the 200 cell images closest to the centroid for each phenotype were projected in 3-dimensional t-distributed stochastic neighbor embedding (t-SNE) space and then displayed in two dimensions (Figure 2A). As expected, the DMSO-treated controls (gray dots) were well separated from both the drug treatments and the two different classes of dead cells (green and black dots). To avoid cumbersome nomenclature, we refer to the cell states by the known main mechanism of action for each of the drugs. This approach also provides a convenient conceptualization of the major effect of each tool compound. Thus the phenotypes resulting from treatment with staurosporine and venetoclax (DID, triggers apoptosis, Dead In Drug), phorbol-12,13-dibutyrate (PDB, activation of protein kinase C, purple), 2-deoxyglucose (2DG, suppresses glycolysis leading to quiescence, blue), lenalidomide (LEN, alters the specificity of the proteasome, cyan), and tunicamycin (TN, inhibits glycosylation in the endoplasmic reticulum [ER] resulting in a strong stress response, yellow) are reasonably separated (Figure 2A).

Figure 2.

Measuring cell responses to drugs based on morphological changes identified by confocal fluorescence microscopy

Primary patient cells incubated in 3S media were exposed to the indicated drugs and stained with Hoechst to highlight nuclei and ChromaLIVE to image intracellular membranes. The drugs were selected based on known or expected effects on cells and used to identify feature sets that distinguish outcomes.

(A) High-dimensional morphological features (110 dimensions) were reduced with t-SNE, and representative cell images are color-coded by treatment: DMSO (gray), PDB (purple), 2DG (blue), DIM (black, dead in medium), DID (green, dead in drugs), TN (yellow), LEN (cyan).

(B) Random forest classification of 38 morphological features assigned each cell to one of 7 phenotypic classes. Patient-level profiles were calculated as the percentage of cells in each class.

(C) Consensus clustering (K = 5) of DMSO-treated patient samples based on phenotypic composition revealed 5 patient clusters. The color corresponding to the percentage of cells with that phenotype is indicated by the color scale.

(D) Distribution of cell phenotypes across clusters. For each phenotype (row) the mean value of the % of cells from that cluster (columns) assigned to that phenotype is shown for patient samples by a number within the heatmap and is illustrated according to the color scale.

This random forest classifier was then used to quantify the various cell phenotypes in a test set of patient samples. Since the 38-parameter classifier was trained on tens of thousands of cell images, overtraining is highly unlikely. However, the training sets are themselves heterogeneous. For example, ∼20% of all of the samples are made up of dead cells. These dead cells, termed “Dead In Media” (DIM), increased in number due to cytokine and Toll-like receptor agonist stimulation. Clearly DIM cells will not respond to compounds added to induce specific phenotypes, and therefore, DIM-tagged cells (black dots) clustered together irrespective of the treatment. Some cells treated with TN evidently died (yellow dots clustered with the DIM, black dots). Similarly, a few of the yellow dots are located adjacent to milder phenotypes (2DG and LEN), suggesting unsurprisingly, that there are variances in individual cell responses that are captured in the original high-dimensional space to the same concentration of TN. Thus, even at optimal drug concentrations in which most cells are expected to undergo ER stress, individual cell responses to TN varied from mild (2DG and LEN) to dead (DIM). This result was expected and reflects the inherent heterogeneity and the number of dead cells in the patient samples after 72 h incubation in 3S media. Consistent with this heterogeneity, when the classifier data were plotted as a confusion matrix for all 133 patient samples, the enrichment of correctly assigned treatments was clear but moderate (Figure S2A; Table S4). Nevertheless, this result also demonstrates that the classifier can be used to differentiate the cell phenotypes in micrographs.

Having built a classifier that assigns individual cell images to one of seven high-level phenotypic cell states, our goal was to use this tool to reduce the patient sample level heterogeneity seen in Figure 1 into distributions of the defined cell states. For example, DMSO-treated patient samples were described by the distribution of >1,000 cell image assignments among the 7 cell states (Figure 2B). This method prevents overtraining and simplifies complex computations by converting individual image features into a smaller set of phenotypic cell states. These states can be directly linked to the biological changes caused by the specific compounds used, giving a more tangible and understandable view of how the cells respond to perturbations. For example, the fraction of cells assigned to the category TN (tunicamycin) can be viewed as an approximation of a measure of the amount of ER stress in that patient sample. As a multiparametric descriptor of patient samples, the accuracy of the class assignment of each individual cell is not as important as the separation of patient samples based on cell states. Therefore, this approach also minimized the noise in the data.

Cell-state profiling uncovers stable patient clusters in 3S CLL samples

After all patient samples from the “no drug treatment pool” (DMSO) were assessed by the 7-parameter phenotypic cell states (DMSO, 2DG, LEN, PDB, TN, DIM, and DID), considerable patient-level heterogeneity was observed (Figure 2C). Nevertheless, it was possible to detect patient cohorts using unbiased clustering. Analysis of K-means clustering of the 7-parameter patient samples by cluster stability (Figure S3), and separately by entropy, both resulted in the K value being 5 as the optimum number of clusters from the data (Figures 2C and 2D). Moreover, when the 7-parameter data were clustered using affinity propagation, the optimum cluster number was also 5. As expected, the patient samples were assigned to very similar clusters using either K-means or affinity propagation (Figure S2B). Thus, there is sufficient information in the 7 phenotypic parameters to robustly and reproducibly cluster the patient samples in DMSO into 5 clusters that vary in size from n = 9 (Cluster 4) to n = 43 (Cluster 2, Figure 2C). Displaying the patient sample features (% of cells assigned to phenotypes) as a heatmap enabled the basis for cluster enrichment to be visually appreciated (Figure 2C). The distribution of average phenotypes within clusters provides a simpler view of the aggregated data (Figure 2D). From this representation, it can be seen that cluster 2 consists of patient samples largely composed of cell images that resembled the untreated, viable DMSO controls (∼50%), with ∼15% of cells classified as dead (DIM + DID) and with smaller numbers of cells classified as other phenotypes. The patient samples in cluster 0 were similar to cluster 2 except they contained more cell images assigned to LEN and PDB phenotypes and therefore, less assigned to DMSO. Closer inspection revealed that patient samples in clusters 0, 2, and 3 all contained less than ∼20% of cells classified as dead. In contrast, cluster 4 consists of patient samples in which most of the cells died (DIM + DID) during the 72-h incubation (∼94%). Cluster 1 patient samples also had elevated dead cells (∼45%) after 72-h 3S stimulation. In contrast, DMSO-treated patient samples from cluster 3 exhibited the largest percentages of cells with phenotypes resembling cells treated with 2DG (∼quiescence, ∼24%) and TN (∼ER stress, ∼27%) (Figure 2D). Thus, as expected, the different clusters contain different distributions of cell phenotypes.

As an alternative way of scoring the patient-sample phenotypes present in the clusters, the enrichment of features was calculated rather than using a simple percentage. When Monte Carlo-derived phenotypic enrichment was calculated and Z scored for all of the cell phenotypes, the results were similar to those obtained using percentages, except that some of the less-common phenotypes were more evident (Figure S2C). This approach has the advantage that statistical significance can also be quantified but since the results are more complicated to display and the derived insights were not substantially different than those obtained using percentages, the latter was used for most of the analyses reported here.

Patient clustering reveals unique clinical profiles unrelated to conventional risk stratification

Patient clusters established based on the 7-parameter phenotypic cell states were examined for enrichment in clinical features such as age, disease severity, etc. For example, in cluster 4, in which the CLL cells mostly died during the 72-h 3S incubation (∼80% DIM), the patients were significantly younger than the average in other clusters (Figures 3A and S4A). This suggested that disease onset or progression might be reflected in patient sample clusters. As patients were sampled over a 3-year period, several were assayed at more than one time point. As expected, samples from the same patient tended to be assigned to the same cluster. However, we also observed that patients sampled at longer time intervals (> a year apart) were more likely to cluster differently (Figure S4B). Given the patients were untreated or sampled several months after treatment cessation, this suggested our approach is likely detecting the morphological differences of CLL patient cells related to disease progression. For this reason, each patient sample tested at different times of collection was treated independently in the following analyses.

Figure 3.

Patient cell-image-based clusters and their correlation to patient clinical features

(A) Age at diagnosis for patients in cell-image-based clusters. Patient codes were used to link patient ages from Table 1 to the appropriate cell-image-derived cluster. Patient samples were treated as individual entries. Patient clusters were color coded as shown in the boxplot. The middle hinge indicates the median (50% quantile) of each distribution; the box encompasses the lower and upper hinges (25% and 75% quantile). The whiskers stretch to the minimum and maximum values. p values were derived from a Kruskal-Wallis test corrected for multiple hypotheses using the Bonferroni method.

(B–E) Flow cytometry data reveal cluster-specific expression levels for CD38 and CD5. Flow cytometry histograms for (B) CD38 and (C) CD5 intensities of CD45⁺CD19⁺ B cells for selected representative patient samples from the indicated clusters. Flow cytometry histograms for (D) CD11c and (E) CD5 of CD45⁺ CD3⁻ CD56⁻ inferred B cells. Line plots below show the distribution of cells from each patient cluster (0, 1, 2, and 3) across deciles for the proteins indicated. For each query (CD38, etc.), the intensities for all cells of the patients were pooled and used to calculate deciles. Then the number of cells in each decile was calculated for individual clusters and expressed as a percentage of the cluster total on the line graphs (the colors indicate the cluster numbers as indicated). The black dashed line indicates the 10% of cells expected in each decile assuming the distribution within a cluster is the same as for the entire population.

Significant differences were not observed between the patient clusters and other available patient clinical records including β2-microglobulin , white blood cell count, and IgG levels (Table 1; Figures S4C–S4E). In a recent study, clusters of CLL patients identified based on cell viability measurements after cytokine treatments showed differences in the lymphocyte doubling times across clusters.¹⁶ However, in our study, where clusters were derived using high-content analysis of cellular morphologies of 3S CLL, we found no differences in lymphocyte doubling time measurements among the patient clusters (Figure S4F).

Table 1.

CLL patient anonymized code and clinical information

PatientID	Age	Sex	Duration	Stage	WBC	β2-m	FISH	IGHV	Treatments	IgG	LDT
P1	52	F	6	4	150	3.3	Normal	NA	3	7.5	1.9
P2	80	F	8	4	200	5	13q	Mutated	1	5.6	2.5
P3	78	M	3	3	72	3.9	11q	Unmutated	2	5.2	1
P4	78	M	3	3	72	3.9	11q	Unmutated	2	5.2	1
P3	78	M	3	3	72	3.9	11q	Unmutated	2	5.2	1
P4	78	M	3	3	72	3.9	11q	Unmutated	2	5.2	1
P5	82	M	5	3	139	13.3	13q	Unmutated	0	6.8	0.8
P6	70	M	13	3	124	3	13q	NA	2	SCIG	NA
P7	80	M	4	4	101	4.8	13q	NA	3	SCIG	0.3
P8	74	M	20	1	76	2.8	p53 mut	Mutated	0	6.1	4
P10	58	M	7	3	209	2.4	13q	NA	0	5.6	2
P11	74	M	20	4	30	6.7	13,8p	Mutated	0	13.2	19
P12	46	M	1	0	34	2.4	NA	Unmutated	0	8.3	2
P13	66	M	14	4	220	3.6	NA	Unmutated	0	SCIG	10
P14	83	M	22	0	52	2.5	NA	Mutated	0	4	5
P15	78	F	12	0	59	2.7	NA	NA	0	6.2	8
P16	82	F	5	4	25	4	t12	Mutated	1	4.5	3
P17	62	M	19	0	27	2.3	NA	NA	0	8	21
P18	62	F	25	4	45	2.8	13q	Mutated	0	SCIG	20
P19	75	M	3	4	180	4.4	13q, 11q	Unmutated	1	7.3	NA
P20	59	M	24	4	40	6.9	NA	Mutated	0	SCIG	2
P21	69	M	25	2	46	2.7	13q,17p	Unmutated	6	SCIG	NA
P22	82	M	4	4	183	5.2	13q	Unmutated	0	9.1	1
P23	68	M	15	2	45	3.1	Normal	Mutated	0	6.04	17
P24	74	F	9	4	97	5.3	11q	Unmutated	2	9.4	1.8
P25	64	M	17	4	24	1.9	17p	NA	3	SCIG	0.4
P26	64	F	1	2	62	2.7	13q	Unmutated	0	6.1	0.5
P27	71	M	3	0	124	2	13q	Mutated	0	8	8
P28	58	F	13	3	208	4	13q	Mutated	0	7.1	2
P29	72	M	15	3	180	3	13q	Mutated	0	SCIG	12
P30	68	M	12	4	145	5.7	13q	Mutated	0	SCIG	5
P31	62	M	13	4	33	3.2	13q	Mutated	3	SCIG	4
P32	79	M	11	4	20	1.9	NA	NA	0	7.9	11
P33	64	M	11	0	46	2.5	NA	NA	0	9	6
P34	64	M	11	1	20	2.6	13q	Mutated	3	SCIG	5
P35	66	M	10	3	169	2.9	13q	Mutated	2	SCIG	NA
P36	53	M	15	3	430	3.2	13q	Mutated	1	SCIG	4
P37	71	F	15	0	14	2	11q	NA	1	11.8	12
P38	66	M	9	1	28	2.7	13q	Mutated	0	SCIG	10
P39	64	M	11	0	36	2	NA	NA	0	8.2	12
P40	70	M	9	4	75	6.3	17p	NA	3	SCIG	1
P41	78	M	11	4	70	6.2	13q	Unmutated	1	SCIG	10
P42	69	F	8	4	98	3.4	Normal	Mutated	0	6.8	6
P43	69	F	8	4	98	3.4	Normal	Mutated	0	6.8	6
P44	60	F	8	4	450	4.2	13q	Mutated	1	2.5	3.2
P45	79	F	9	4	67	6.6	13q	Mutated	0	5.2	3
P46	82	F	12	0	41	3.4	NA	NA	0	6.6	5
P47	57	M	7	4	104	1.9	Normal	Mutated	1	SCIG	6
P48	80	M	3	4	68	3.8	13q	Unmutated	0	10.7	1.3
P49	64	M	11	4	87	3.7	13q	Unmutated	1	6.5	1.1
P50	70	F	7	4	74	4.5	13q	Unmutated	0	SCIG	0.8
P51	48	M	5	2	66	3.7	11q	Unmutated	0	11.6	1.5
P52	66	M	10	0	45	2.2	NA	NA	0	10.5	5
P53	74	F	25	0	13	3.1	NA	NA	3	SCIG	7
P54	66	M	8	3	250	4	11q	Unmutated	2	SCIG	1.3
P55	71	M	6	3	156	3.3	13q,11q	Unmutated	1	9.1	0.5
P56	71	M	6	3	156	3.3	13q,11q	Unmutated	1	9.1	0.5
P57	73	M	5	3	65	6.3	13q	Mutated	1	14.2	1.5
P58	67	M	6	0	21	1.7	NA	Mutated	0	7.6	6
P59	61	M	7	4	21	2.8	NA	Unmutated	0	9.9	7
P60	70	F	5	3	95	3.7	NA	Mutated	0	8.1	2.6
P61	70	M	3	3	228	6.4	t12	Unmutated	1	6	0.9
P62	69	F	7	0	52	2.2	NA	Mutated	0	8.6	4.5
P63	52	F	7	4	76	3	Normal	NA	2	7.6	NA
P64	59	M	5	4	138	7.9	Normal	Unmutated	1	SCIG	1
P65	73	M	10	1	45	2.5	NA	NA	0	5.8	5
P66	63	M	4	4	97	6	NA	Unmutated	0	11.7	1.1
P67	54	M	3	1	47	1.8	NA	Mutated	0	6.7	2.2
P68	58	M	3	0	22	2.7	NA	Mutated	0	10.03	NA
P69	47	M	2	4	114	2.4	Normal	Mutated	0	5.5	1.9
P70	55	M	7	4	67	4	17p	NA	2	NA	NA
P71	72	F	1	0	16	1.9	NA	Mutated	0	10.4	NA
P72	67	F	3	0	70	3.2	NA	Mutated	0	10.6	1.2
P73	58	M	3	3	86	2.7	13q	Mutated	0	8.2	1.5
P74	87	M	5	4	44	4.4	t12	Mutated	0	6.2	NA
P75	66	M	5	3	172	3.9	Normal	Unmutated	1	9	0.6
P76	84	F	9	4	20	7.7	Normal	NA	2	SCIG	1
P77	82	F	8	4	212	7	17p	NA	5	9.2	0.9
P78	82	F	9	3	333	5.8	Normal	NA	2	SCIG	3
P79	94	M	25	4	62	10.9	t12 del8p	NA	5	6.6	3
P80	86	M	7	4	73	5.7	Normal	Unmutated	2	SCIG	1.2
P81	58	F	6	1	24	1.8	13q	Unmutated	0	SCIG	4
P82	94	F	5	0	33	3.1	NA	NA	0	7.7	6
P83	75	F	7	2	58	2.8	Normal	Unmutated	0	9.1	8
P84	87	F	12	3	58	6.5	13q	Mutated	0	6.1	12
P85	58	F	8	4	98	3.8	13q	Unmutated	0	SCIG	6
P86	73	F	5	3	65	11.9	13q	Mutated	1	3.4	2
P87	69	F	5	4	370	4.6	13q	Unmutated	1	6.2	2.2
P88	77	M	11	0	27	2.5	NA	NA	0	5.8	5
P89	70	M	7	0	41	1.7	NA	NA	0	8.4	4
P90	80	M	5	4	34	3.7	NA	Mutated	0	11.01	4.5
P91	56	M	3	1	90	2.1	NA	Unmutated	0	6.7	1
P92	75	F	7	3	62	2.8	NA	NA	0	9.5	3
P93	86	M	22	3	77	4.9	13q	Mutated	0	SCIG	3
P94	77	F	11	4	93	6.9	13q	Unmutated	0	5.48	2.9
P95	54	F	6	4	150	4.2	t12	Unmutated	1	6.4	1.6
P96	59	M	9	4	77	2.6	NA	Mutated	0	6.2	4.5
P97	78	M	3	4	150	4	17p	NA	3	6.1	0.3
P98	87	F	25	4	70	3.6	NA	Unmutated	1	8.4	6
P9	66	M	17	4	74	4	NA	Mutated	0	4.5	11

The table summarizes clinical details of CLL patients included in this study, including anonymized patient identifiers (PatientID), age of disease diagnosis (age), patient gender (sex), disease duration at the time of sampling in months (duration), disease stage, WBC count expressed in ×10⁹/L, β2-m levels in mg/L, FISH status, IGHV region gene mutation status, SCIG replacement, LDT in months, and other relevant clinical features. All patient codes are anonymized.

β2-m, β2-microglobulin; FISH, fluorescence in situ hybridization status; SCIG, subcutaneous immunoglobulin; LDT, lymphocyte doubling time; NA, not available.

To further investigate associations between patient clusters and other known disease risk factors, we investigated the patient IGHV status and computed the CLL-IPI. The mutational status of the variable region of the IGHV gene has been shown to be a powerful predictor of duration of response and overall survival with chemoimmunotherapy.¹⁹ CLL-IPI has been adopted as the staging method for modeling the CLL prognosis by combining the genetic, biochemical, and clinical parameters.⁶ By comparing the frequency of patient distribution in each cluster to IGHV mutation status or different IPI risk categories, no significant association was detected among high-content screening (HCS)-derived clusters (not shown). However, the patient numbers in each cluster are small enough that only strong correlations would reach statistical significance. Nevertheless, the patient sample clusters based on drug-induced phenotypes appear to be mostly independent of conventional risk parameters.

For 38 patients, we were able to retrieve clinical flow cytometry data performed on peripheral blood. An initial assessment based on conventional 2-dimensional dot plots suggested intensities for a variety of markers were different between patient sample clusters; however, the within- and between-patient heterogeneity and small numbers of patients for which flow data were available (typically ∼10 patient samples per cluster) precluded establishing statistical significance based on these criteria. For visual examples of mean intensity distributions of CLL cells from patients in clusters 0–3, see cell count vs. intensity plots in Figure 3B–3E (upper rows). As an alternative approach for selected markers, the distributions of the cell-level intensity data for all of the patient samples were pooled and divided into intensity deciles. Comparing the intensity distributions for individual patients within clusters revealed that for CD38, CD5, and CD11c, most patients within a cluster had similar intensity distributions. For example, all of the patients in cluster 0 had similar intensity distributions for CD11c, and 9 of 12 patients in the same cluster had similar intensity distributions for CD38. The other 3 patients had distributions similar to each other but different from the 9 (Figure S5). Similarly, for the 10 patient samples from cluster 2, the CD5 and CD38 levels were very similar to each other for two groups of 5 patient samples.

To analyze the differences in markers between clusters quantitatively, we compared the distributions of marker intensities using a Kolmogorov-Smirnov statistic. This analysis revealed that for B cells (CD45⁺ CD19⁺ cells) there were highly statistically significant differences between clusters in the distributions of CD38 intensities, with all four clusters showing different intensity distributions in deciles 7–10 (Figure 3B). While statistically significant, comparison of the flow cytometry histograms suggests that these differences are too small to reliably assign patients to clusters.

In addition to the differences observed for CD38, cellular CD5 intensities (associated with lymph node egress²⁰ and clinical outcomes⁴) varied across all deciles. For example, patients from cluster 0 had few cells in deciles 1–3 and a larger percentage of cells in the higher deciles (7–10). In contrast, samples from cluster 2 had more cells in deciles 1–2 and less in decile 10 (Figure 3C). As expected, when the analysis was carried out in parallel, for samples of “inferred B cells” (CD45⁺ CD3⁻ CD56⁻), the distribution of CD5 intensities across deciles for each cluster in Figure 3E was similar to that seen for CD45⁺ CD19⁺ B cells in Figure 3B. This result indicates that, as expected, the findings were independent of the gating strategy employed. In contrast, no between-cluster differences were observed for CD11c intensities for the inferred B cells (Figure 3D). In summary, clustering patients based on morphological measurements of 3S CLL reveals significant distribution differences in clinically established phenotypic markers such as CD38, which would not otherwise have been recognized. These results also support the validity of refining patient CLL cells into cohorts beyond clinically CD38-positive/negative categories.

Patient clusters exhibit differential responses to therapeutic drugs

As the features used to cluster patient samples are related to cell phenotypic responses to the tool compounds, we hypothesized that cluster membership might predict some level of commonality in drug responses. An exception was cluster 4, in which most of the cells died during the incubation in control, DMSO-containing, 3S media. Dead cells cannot respond to drugs; therefore, patient samples from cluster 4 were not analyzed for drug responses. Patient samples from clusters 0 and 2 exhibited limited morphologic changes in response to the addition of cytotoxic agents (Figures 4 and S6). This result along with lack of killing in response to 30 nM ibrutinib for patient samples from all of the clusters is consistent with previous data for 2S media with IL-4, a media composition essentially identical to the 3S CLL used here.⁹ However, for patient samples from cluster 0, the addition of ibrutinib reduced the percentage of cells assigned the PDB (stimulated) phenotype, an effect not seen for samples from cluster 2. The main phenotypic change in response to frontline therapeutic agents for patient samples from clusters 0 and 2 was an increase in the percentage of cells assigned to the LEN phenotype. This was particularly true in response to 1 μM avadomide, in which the percentage of cells scored as LEN increased from 16% to 27%, and from 11% to 27% of cells, for clusters 0 and 2, respectively (Figure 4). This result was expected due to the similar mechanisms of action of the two drugs. To our surprise, the patient samples from cluster 1 were most sensitive, while samples from cluster 2 were the least sensitive to 10 nM venetoclax (DID and DIM bins combined ∼61% and 28% dead, respectively, Figure S8A). At 100 nM venetoclax, all the patient samples were 50%–60% sensitive at the 72 h time point (DID and DIM bins combined), irrespective of cluster membership (Figure S8B). Furthermore, combining 10 nM venetoclax with the drugs shown in Figure 4 did not increase the cytotoxicity appreciably. However, for patient samples from clusters 1 and 3, there was a modest increase in cell killing when 100 nM venetoclax was combined with 1 μM avadomide (Figure S8B).

Figure 4.

3S CLL from cluster 0 and 2 are resistant to frontline therapeutic drugs

Patient samples were treated with DMSO as a solvent control or 3 μM atovaquone, 1 μM avadomide, 3 μM chlorambucil, 1 μM dexamethasone, 1 μg/mL fludarabine, or 30 nM ibrutinib as indicated to the left. Cell images for patient samples from clusters 0 and 2 were assigned to the phenotypes indicated above the columns. The percentages of cells assigned to each phenotype are indicated numerically and colored according to the color scale to aid visual interpretation. Cells/micrograph and var indicate the average and the variance of the number of cells per micrograph for the two micrographs analyzed for each patient sample and treatment (green). That the number of cells/micrograph is not different from the DMSO control verified that cells were not lost due to drug treatment.

To better understand the wide range of compounds tested in the study with different mechanisms of action, we grouped them based on phenotypic similarity. They were assessed by an XGBoost model using the change in distributions of patient cells in the 7-parameter phenotypic cell states across all patient samples. As expected, the internal controls, avadomide,²¹ and lenalidomide²² clustered together in a phenotypic similarity matrix (Figures 5A and S6). Unexpectedly, atovaquone, while separated from the rest of query drugs, had avadomide and lenalidomide as nearest neighbors. Atovaquone was included in our experiments because of its effects on STAT3 that transmits JAK signals and not as a conventional therapeutic for CLL.²³ The cytotoxic drugs, chlorambucil, dexamethasone, and fludarabine clustered together along with the potent and selective TYK2 inhibitor NDI 031301. Compared to the frontline therapeutic agents, the other janus kinase (JAK) inhibitors (JAKins) are more similar among themselves and distant to non-JAKins when the treatments are clustered based on responses of the patient cells (Figures 5A and S6). Moreover, these clustering results segregated cerdulatinib and peficitinib, which target multiple JAK isoforms, and also tyrosine kinases from filgotinib, ruxolitinib, and tofacitinib, which target selective JAK isoforms (Figures 5A and S6).^24,25,26

Figure 5.

Phenotypic similarity between drugs and patient responses

(A) Phenotypic similarity values between drugs were used to cluster drugs by hierarchical clustering. Phenotypic similarity represents the percentage of testing drug (row ids) sample wells that are predicted as query drug (column ids). Thus, each value in the hierarchical heatmap represents the phenotypic similarity between corresponding drugs (see “phenotypic similarity between drugs” in STAR Methods for details). Avadomide and lenalidamide were included as internal controls as the drugs share very similar mechanisms of action. Drug concentrations are 1 μM, except for atovaquone (3 μM), chlorambucil (3 μM), and fludarabine (1 μg/mL).

(B) Phenotypic responses of patient samples from cluster 3 in response to JAKins alone and in combination with 10 nM venetoclax. Patient samples were treated with DMSO as a solvent control or 1 μM of the JAKins indicated to the left. Cell images for patient samples from cluster 3 were assigned to the phenotypes indicated above the columns. The percentages of cells assigned to each phenotype are indicated numerically and colored according to the color scale to aid visual interpretation. (Upper) Addition of JAKins increased cell stress as judged by the increase in cells classified as the tunicamycin phenotype (TN). (Lower) Addition of 10 nM venetoclax (+VEN) resulted in a large increase in the percentage of cells classified as DID (dead in drug) that was further augmented by co-treatment with JAKins.

Despite the similarity in responses across patient samples when treated with JAKins being sufficient to cluster the drugs, the within-cluster responses were still more similar than the responses between clusters. Patient samples from different clusters were not identical following treatment with JAKins alone or in combination with 10 nM venetoclax. For example, incubation of patient samples from cluster 3 in JAKins resulted in a modest decrease in the percentage of cells assigned to DMSO and a corresponding increase in the percentage of cells assigned to TN (stress) (Figure 5B). However, in combination with 10 nM venetoclax, the addition of JAKins reduced the percentage of cells assigned as DMSO, 2DG, and to a smaller extent, PDB, while increasing the percentage of cells assigned to the DID (dead) category (Figure 5B). This suggests that patient samples from cluster 3 are particularly sensitive to this drug combination. In combination with 100 nM venetoclax, the patient samples from cluster 3 (and cluster 1) were particularly sensitive to the addition of the JAKin momelotinib (Figure S8B).

Similar to the results obtained with the frontline therapeutic agents (Figure 4) patient samples from cluster 2 had the lowest percentages of dead cells in response to JAKins without and with 10 nM venetoclax. However, patient samples from cluster 2 did not respond equally to all of the JAKins (Figure S8). Patient samples from this cluster were most responsive to incubation with the JAKin cerdulatinib, which reduced the percentage of cells assigned to the DMSO phenotype by almost 50%, a much greater decrease than for the other JAKins. The largest corresponding increase was in the cells assigned to TN with some also in the dead category (DIM + DID). This suggests that patient samples from cluster 2 may respond to cerdulatinib by undergoing a stress response. In combination with 10 nM venetoclax, treatment with cerdulatinib resulted in most of the cells lost from the DMSO phenotype being assigned as dead, as indicated by assignment to the DIM and DID phenotypes, suggesting that venetoclax killed the cells assigned as TN as a result of incubation with only a JAKin (Figure S8A). Interestingly, the other two JAKins, peficitinib and gandotinib, increased cell death in combination with venetoclax, clustered adjacent to one another, and have also been reported to inhibit tyrosine kinases (Figure 5A). This observation may suggest that for patient samples in cluster 2, selective tyrosine kinase inhibition could be a beneficial addition to venetoclax.^25,27

Unexpectedly, patient samples from cluster 1 seemingly underwent more cell divisions than patient samples from any of the other clusters, as indicated by the average cells per micrograph column in Figure S8A. Furthermore, these patient samples exhibited only small changes in response to the addition of JAKins, most notably a small increase in the percentage of cells assigned to TN. This was accompanied by a decreased percentage of cells assigned PDB and a small increase in cells assigned DID phenotype (dead). For these samples, the addition of venetoclax resulted in a marked increase in cells assigned to the DID category primarily at the expense of the DIM category, suggesting that incubation in this drug combination changed the mechanism of cell death to be more apoptotic consistent with the known mechanism of action of venetoclax. However, as described above, samples in cluster 1 were also sensitive to avadomide in combination with 100 nM venetoclax (Figure S7B).

Patient cluster membership predicts responses to drug treatments

Building on the finding that drug responses within a patient cluster were more similar to each other than to responses in other clusters, we next investigated whether a patient’s cluster membership obtained from untreated samples could be used to predict their response to various drugs. To this end, patient samples treated with the drugs were clustered independently using K-means methods described previously. For patient samples treated with ruxolitinib (cluster number prefixed B), the optimal value of K was 6, even though the one additional cluster contained only two patient samples. For patient samples treated with both ruxolitinib and venetoclax (cluster number prefixed C) the optimal value of K was 5. After clustering the patient sample data for each of the treatments, the results were compared using a Sankey plot (Figure 6). In this way it is possible to track the relationships between clusters via a direct linkage map. To provide a simple visual summary of the cell phenotypes within each cluster we used pie charts in which the cell phenotypes were combined into four categories: live, quiescent, stressed and dead. This analysis revealed that the initial assignment of untreated patient samples into clusters was predictive of both responses to ruxolitinib and the combination of ruxolitinib and venetoclax. It was not surprising that when treated with drugs, the 9 patient samples in cluster 4, composed primarily of dead cells, were reassigned to clusters of dead cells. The results for patient samples from cluster 2 (Figure 6) were mostly unchanged by treatment with ruxolitinib alone or in combination with venetoclax. As a result, these patient samples remained in the clusters with the largest percentage of live “DMSO”-type cells (B_13 and C_11), as expected from the data in Figure 4. As indicated by the pie charts, the response to ruxolitinib and venetoclax in combination was an increase in the fraction of “dead” cells in cluster C_11 compared to cluster 2. The small increase in patient samples from 43 in cluster 2 to 50 in C_11 is due to an influx of relatively resistant patient samples from other clusters. That most of the cluster 2 patient samples end up in the cluster with the smallest fraction of dead cells confirms the data presented above, suggesting that cluster 2 identifies patient samples that are the most resistant to both conventional and targeted therapies.

Figure 6.

Sankey plot illustrating the relationship between patient clusters and their phenotypic responses to ruxolitinib alone and in combination with venetoclax

The pie charts illustrate the simplified distribution of phenotypes within the clusters. Each line represents one patient sample. Clusters were independently generated for DMSO, ruxolitinib, and ruxolitinib plus venetoclax with values of K = 4–6 to determine the value of K that generated the highest stability clusters. Clusters 0–4, the original 3S CLL patient samples clusters based on the DMSO solvent control and K = 5. Clusters B_10–B_15, patient samples clusters for 3S CLL exposed to 1 μM ruxolitinib clustered by K-means with K = 6. Clusters C_10–C_13, patient samples exposed to 1 μM ruxolitinib plus 10 nM venetoclax clustered by K-means with K = 4. In all cases, the number after the colon indicates the number of patient samples in the corresponding cluster.

The patient samples from cluster 3 were particularly revealing. These patient samples were split almost evenly by treatment with ruxolitinib alone, with about half of the cells joining a cluster (cluster B_14) exhibiting minimal change, while the other half joining a cluster that showed a marked increase in TN “stressed” and a moderate increase in 2DG “quiescent” cells (cluster B_10). However, when 10 nM venetoclax was added in addition to ruxolitinib, almost all of these patient samples ended up in the same cluster (C_13) that is composed of ∼80% dead cells along with ∼20% of cells assigned as “stressed.” Thus, cluster 3 membership was highly prognostic for a large increase in dead cells in response to this combination treatment. Similarly, patient samples from cluster 1 split into two different clusters in response to ruxolitinib, one of which was similar to the untreated patient sample (B_15) and the other in which the percentages of quiescent and stressed cells increased (B_14). In combination, venetoclax and ruxolitinib increased the percentage of dead cells dramatically (Figure 5B).

Finally, only for patient samples from cluster 0 was cluster membership unstable with the addition of ruxolitinib and even less so in response to the combination of ruxolitinib and venetoclax. However, the fraction of cells assigned to the dead category increased from roughly 25% with ruxolitinib to ∼45% with the addition of 10 nM (Figure 6).

Taken together, the linkage map data strongly suggest that there is a cohort of patient samples (cluster 3) that includes ∼20% of all patient samples, which responded strongly to the combination of ruxolitinib and venetoclax. This cohort was identifiable based on the relatively large percentages of TN (stressed) and 2DG (quiescent) cells in untreated samples (cluster 3). In addition, another cohort of patient samples (cluster 1) that was effectively killed (65%) by the combination of ruxolitinib and venetoclax was identifiable due to the induction of quiescence and stress by ruxolitinib alone (Figure S8). Taken together, these results suggest that drugs that induce a stress response in 3S CLL cells sensitize them to killing by venetoclax.

To further support these conclusions, we also compared frontline treatments exemplified by chlorambucil alone and in combination with 10 nM venetoclax or 30 nM ibrutinib (Figure S8). Again, the initial untreated patient samples clusters were predictive of both responses to chlorambucil and the combination of chlorambucil and venetoclax/ibrutinib. The linkage map displayed a consistent trend where most patient samples continue to cluster together throughout independent clustering. In the case of chlorambucil and venetoclax, the drugs increased the proportion of dead cells slightly but not enough to dramatically change cluster membership (Figure S8A). As expected, the addition of ibrutinib had very little effect on cell phenotypes since the cells were grown in media containing IL4 (Figure S8B). Thus, cluster 0 and 2 remained resistant in a “live”-dominant phenotype in response to chlorambucil or combination therapies. Clusters 1 and 3 were resistant to chlorambucil as noted earlier in 3S conditions (Figures S1C and S1D). There was no marked increase in TN “stressed” phenotype, and consistent with our hypothesis that venetoclax effectively kills stressed cells, the addition of 10 nM venetoclax to chlorambucil did not increase cell death markedly beyond what could be accounted for by the stressed cells in DMSO (Figure S9A C_10).

Discussion

CLL, like many other types of cancers, is a highly heterogeneous disease. As a result, patients differ in disease progression and also in responses to treatment. Phenotypic profiling as shown here helps in identifying subgroups of patients (cohorts) who are likely to benefit from specific treatments. While conventional classical biomarkers and even the more comprehensive whole-genome or proteogenomic analysis are useful in stratifying patients for disease progression based on risk profiles, they do not predict patient treatment responses.^28,29 As a result, there are frequently no actionable criteria upon which to make decisions on the best drug options.

High-content phenotypic profiling (or “cell painting”³⁰ and “live cell painting”¹²) have become increasingly popular to quantitatively appraise cell responses to chemical perturbations, owing to the advancement in automated instrumentation, image analysis techniques, and machine learning algorithms. We previously leveraged the power of high-content phenotypic screening to identify venetoclax resistance in activated CLL cells and demonstrate that resistance was due to pro-survival kinase signaling from the micro-environment.^9,13 Here, we extended the method to identification of patient cohorts predictive of drug response profiles. We discovered stable cell image-based cohorts based on the phenotypic characteristics of untreated patient cells (Figures 2C and 2D). Importantly, while numerous studies have equated cell viability as a measure of drug response, in our study we employed seven phenotype descriptors as features to profile drug responses. Post hoc analysis revealed that no single phenotypic bin was sufficient to drive clustering or to predict all of the drug responses. Moreover, while the DID and DMSO phenotypes differentiated cluster 2 from the other clusters, the DID phenotype was not nearly as effective as the DMSO phenotype (Figure 2D). Indeed, the dead cell phenotypes (DIM and DID) predominantly contributed to identifying clusters 1 and 4 (Figure 2D), highlighting the importance of features other than cell viability measures to gain a more comprehensive understanding of patient sample responses.

The patient clusters identified using multiple phenotypic features from treatment-naive samples were not identifiable by conventional clinical criteria. Furthermore, with the exception of cluster 4 patients regarding their age at disease diagnosis (Figure 3), the clusters did not segregate clinical correlates. Considering that cluster 4 patients are significantly enriched in cells that died during incubation in vitro (DIM and DID), the clinical correlation to early onset might point to the presence of a subpopulation of patients with aggressive disease that is hard to detect based on current CLL clinical protocols. In addition, clinical flow cytometry data demonstrated differences in the expression patterns for both CD38 and CD5 (Figure 3B). Expression of these proteins has been linked to CLL disease diagnosis, progression, and tumor microenvironment interaction.^31,32,33 However, these changes cannot be used predictively and were not detectable using mean intensities to analyze the flow data because the differences were in the tails of the distributions. Although patient clusters differ in their CD38 and CD5 expression profiles, the differences are subtle and alone or in combination could not be used to establish patient clusters. In parallel, we also explored using the seven high-level phenotypic features to predict disease severity and risk status. To extend this analysis, we selected 40 image features associated with risk status and calculated the mean Wasserstein distance between patients thereby, producing a similarity matrix that was used to hierarchically cluster patient samples or as an input for nine different machine learning models. All these models showed low performance as assessed by accuracy, precision recall, and F1 score. We conclude, therefore, that the phenotypic features we derived are only effective for predicting drug responses. This result is consistent with our use of phenotypic features derived from drug responses and aligns with the underlying hypothesis that drug response predictions are independent of disease prognosis.

While there have been significant advancements in the development of small molecule inhibitors targeting the Bruton tyrosine kinases (e.g., ibrutinib and newer variants) and the inhibitor of apoptosis, Bcl-2 (such as venetoclax), current treatment approaches for CLL remain non-curative. This persistent challenge has spurred an ongoing quest for novel therapeutic strategies. One noteworthy avenue is the exploration of pro-survival signaling originating from the microenvironment. Ruxolitinib has been tested in over 30 CLL patients as a single agent or in combination with ibrutinib with or without the cytotoxic glucocorticoid dexamethasone.^34,35,36,37 Ruxolitinib appears to generally activate CLL cells in vivo in part by interruption of tumor suppressor signals from microenvironmental cytokines like IL-10.³⁴ Ruxolitinib has not yet been tested in vivo in combination with venetoclax. Given the importance of the JAK/STAT pathway in IL-4 signaling, we investigated JAK inhibition coupled with venetoclax. Across all patient clusters, cluster 1 patient samples are most sensitive to JAKins alone (Figure S8). In combination, JAKins sensitized 3S CLL to venetoclax to various degrees, in contrast to chlorambucil, which did not (Figure S7). However, no single JAKin was able to sensitize all patients to venetoclax. Cluster 3 patient samples, enriched in TN phenotype cells, were most sensitive to the combination of JAKin and venetoclax (Figure S8). This was followed by patient samples from cluster 1, in which JAK inhibition triggered an increase in TN (stress) response phenotypes. In contrast, patient samples from cluster 2 with the highest percentage of live DMSO-type cells were most resistant (Figure S8). This result is consistent with variability in patient responses to “one-size-fits-all” treatment approaches and speaks to the need for a precision approach. Consistent with our hypothesis, when 3S cells treated with individual JAKin or JAKin + venetoclax were clustered independently, the patient clusters and their phenotypes are largely preserved. The bulk of cluster 2 patient samples remained together with a DMSO “Live” phenotype, while cluster 3 patient samples mostly ended up together with a “Dead” phenotype (Figure S8). Similarly, when treated with chlorambucil, the patient sample clusters maintain their distinct drug response tendencies and overall phenotypic profiles. This suggests that the cluster membership can be considered to be a predictive drug response marker for patient stratification.

The fundamental approach of using live cell imaging to extract high-level phenotypic features from treatment-naive samples, followed by identifying patient subgroups predictive of drug responses, is not limited to CLL. Despite the inherent heterogeneity of cancer, this strategy could be broadly applicable to other malignancies where conventional biomarkers fail to capture functional drug sensitivity. By leveraging dynamic, quantitative measurements of cellular morphology and behavior under controlled conditions, phenotypic profiling can provide a functional readout of drug response potential, offering a precision medicine framework adaptable across diverse cancer types.

Limitations of the study

We acknowledge four limitations of this study. First, the analysis was conducted using ex vivo primary CLL cells cultured for 72 h in a simulated microenvironment model (3S media), which may not fully recapitulate the complexity and heterogeneity of the in vivo tumor microenvironment and its dynamic interactions with surrounding tissues. Second, patient cohorts were from only 133 patient samples. Close examination of the clustering data suggests that if a larger number of patient samples were analyzed, some of the clusters, particularly cluster 0, may have split into two. While these clusters demonstrated predictive value for drug responses in vitro, clinical validation through prospective trials is essential to confirm their utility in guiding treatment decisions. Third, the goal here was to identify drug combinations that would augment patient responses to venetoclax. This limited the number of drug combinations tested and restricted our understanding of whether the phenotypic clusters can predict responses to other emerging therapies. Finally, future studies should address the molecular and genetic mechanisms underlying the distinct phenotypic clusters. At present, these remain incompletely characterized, potentially limiting our understanding of why certain patient cohorts respond differently to specific drug combinations.

Resource availability

Lead contact

Requests for further information and resources should be directed to and will be fulfilled by the lead contact, David W. Andrews (david.andrews@sri.utoronto.ca).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •All of the data extracted from the images are available in tabular form at Borealis: https://doi.org/10.5683/SP3/ZOLWB3.
• •The patient data supporting the findings of this study are provided in the anonymized Table 1.
• •Code used to generate the analyses can be found at Github: https://github.com/Glauber234/Manuscript_code. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgments

Yonghong Shi and YuXuan Luo helped purify cells. Santosh Hariharan helped with generating data analysis tools, and Justin Kale helped with data management. Funding was provided by the Canadian Institutes of Health Research Foundation grant FDN143312 to D.W.A. Additional funding included an Artificial Intelligence Catalyst Grant to D.W.A. and an operating grant to D.E.S. from the Leukemia and Lymphoma Society of Canada.

Author contributions

M.X.L., G.C.B., D.W.A., and D.E.S. designed the research, interpreted the data, and wrote the manuscript; B.L., D.W.A., D.E.S., and H.T. supervised the project; A.B. established and performed primary cell culture and image quality control analyses; J.Y. prototyped high-content screening and data analyses; J.Y., M.X.L., and A.B. performed imaged-based experiments and contributed to image analysis; J.Y. wrote custom software scripts for image analysis; S.U. performed and contributed to flow cytometry experiments and analysis; G.C.B. and M.X.L. performed data analysis; B.L., H.T., and D.E.S. provided clinical insight into experimental design and data analyses.

Declaration of interests

D.W.A. is a consultant for the company Saguaro Biosciences.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
CD45 KO	Beckman Coulter	Product: A96416; Clone: J33; RRID: AB_2888654
CD19 ECD	Beckman Coulter	Product: IM2708U; Clone:J3-119; RRID: AB_130854
CD3 PC5.5	Beckman Coulter	Product: A66327; Clone: UCHT1; RRID: AB_10638217
CD5 PB	Beckman Coulter	Product: A82790; Clone: BL1a
CD56 APC	Beckman Coulter	Product: IM2474U; Clone: N901; RRID: AB_2876784
CD11c PE	Beckman Coulter	Product: IM1760U; Clone: BU15; RRID: AB_10645163
CD38 PC5.5	Beckman Coulter	Product: A70205; Clone: LS198-4-3
Biological samples
Human blood samples	This Study	N/A
Chemicals, peptides, and recombinant proteins
Resiquimod (toll-like receptors-7/8 agonist)	Alexis Biochemicals (San Diego, CA, USA)	N/A
Interleukin-2 (IL-2)	Novartis Pharmaceuticals Canada Inc., Dorval, Quebec	N/A
Interleukin-4 (IL-4)	PeproTech, Inc.	N/A
ChromaLIVE™	Saguaro Biosciences, Quebec City	N/A
Hoechst 33342	Thermofisher	Cat# H1399
Deposited data
All of the data extracted from the images is available in tabular form	Borealis: https://doi.org/10.5683/SP3/ZOLWB3	N/A
Code used to generate the analyses	Github: https://github.com/Glauber234/Manuscript_code	N/A
Software and algorithms
R (version 4.4.2)	R Foundation for Statistical Computing	N/A
RStudio	R Foundation for Statistical Computing	N/A
Python 3.12.5	Python Software Foundation	N/A
Harmony High-Content Imaging and Analysis Software	Revvity	N/A
SankeyMATIC	https://sankeymatic.com/build/	N/A

Experimental model and study participant details

Study design and patients included in the study

Criteria of patient inclusion in our study were as follows: all patients had either not previously been treated or were assayed after a several month period of cessation of treatment, thereby minimizing potential confounding effects of clinical interventions on assay outcomes. Race and/or ethnicity data for the study participants are not reported because this information was not collected at the time of enrollment according to Health Canada guidance. Heparinized blood was obtained from consenting patients with CLL (Table 1). Protocols were approved by the Sunnybrook Ethics Review Board, (PIN 222–2014) and informed consent was obtained in compliance with the Declaration of Helsinki. For stimulation of CLL cells, resiquimod, interleukin-2, and interleukin-4 were added at 1 μg/mL, 500 U/mL and 20 ng/mL final, respectively. These stimulated cells are referred to as “3S” cells. Cell number and viability were checked using Countess automated cell counters (Thermofisher) prior to assay plate seeding. At the start of the assay, cell viability was measured by trypan blue staining. A total of 133 patient samples acquired from 89 individual patients passed the quality check and were analyzed in this study shortly after collection. No frozen samples were used.

Method details

Assay consumables and compounds dispensing

Selected testing compounds (Table S1) covering different mechanisms of action were dissolved in DMSO and stored in a NuncⓇ polypropylene clear 384-well microplate (Thermofisher, Catalog number: 264574). For each assay plate, compounds were dispensed (25 nL) into the wells of a 1536-well SCREENSTAR imaging microplate (Greiner Bio-One, Catalog number: 789866) using a MosquitoⓇ HTS nanoliter liquid handler followed immediately with 5 μL cell sample dispensing by MANTISⓇ microfluidic liquid handler. 1 μL combinations of venetoclax and/or ibrutinib (stocks prepared in DMSO, then diluted in UNS or 3S media, 0 · 1% v/v DMSO final) or DMSO control (UNS or 3S media with 0 · 1% v/v DMSO final) were delivered using a MANTISⓇ microfluidic liquid handler. The final volume in each well was 6 μL with DMSO content adjusted to 0 · 43% v/v. Drugs used in the assays were obtained from Selleck Chemicals.

In vitro cytokine stimulation and high-content imaging

For high-content imaging, compounds were added to assay wells of 1536-well plates imaging plates (Table S1) and then suspensions of 5 μL (1 × 10⁶ cells per mL) unstimulated or 3S CLL cells were added to the wells via a MANTIS dispenser. After the standard 72 h of compound treatment, sufficient to ensure even epigenetic changes were manifest, the cells were stained with the non-toxic ChromaLIVE dye (1: 1000 dilution final, Saguaro Biosciences, Quebec City) and Hoechst 33342 (5 μg/mL final, Thermofisher) for 30 min before being imaged by automated live-cell high-content confocal fluorescence microscopy with a 20X water immersion lens (Opera Phenix high-content screening system; PerkinElmer). For each sample well, 2 out of 4 total fields of view were captured.

Visualization of the spatial relationships of the cell images from selected control sample wells

In brief, for micrographs of cells from the seven control compound treatments (e.g., Figure 1) for 40 randomly selected patients, 110 features were measured and Z score normalized before being projected onto 3-dimensional space using the t-Distributed Stochastic Neighbor Embedding (t-SNE) algorithm.³⁸ We set the perplexity value to 10 and used 1,000 iterations to generate the 3-dimensional representation of the 110-dimensional data. Due to the limitations of the algorithm regarding memory and computation cost, we did not use all the data for visualization. For each control treatment only the 200 cell images nearest to the centroid were used for visualization (Figure 2A).³⁹

The number of dead cells was highly variable between patients in the DMSO controls. Dead cells were present in freshly isolated CLL samples but generally increased to 2–30% in either unstimulated or 3S media during the 72-h incubation. However, in a few samples there was marked increase in dead cells in the DMSO controls after 72 h incubation in 3S media. In micrographs, cells with shrunken nuclei and reduced mitochondrial transmembrane potential were classified as “Dead-in-Media” (DIM). The nuclear area and ChromaLIVE_561 intensity thresholds used to identify DIM cells were established by comparing the DMSO controls with images of cells treated with 100 nM staurosporine plus 10 nM venetoclax (STS+VEN). The latter were representative of treatment-induced apoptotic cell death (“Dead-in-Drug”, DID). Although in both cases the cells were dead, DIM and DID cells remained distinguishable by measured features, as evidenced by the proximal but well separated clusters observed in the t-SNE space that are distant from other control phenotypic states (Figure 2A). Cells treated with 0.1 μM tunicamycin (TN), 100 ng/mL phorbol 12, 13-dibutyrate (PDB) and 3 mM 2 deoxy-D-Glucose (2DG) were distant from DMSO, but not completely separated among themselves, mirroring the expected phenotypic overlap between these cell states. To identify a cell state control closer to DMSO that could be used in our analyses, some cells were treated with 1 μM lenalidomide (LEN), a drug that alters the specificity of the proteasome. As anticipated the LEN phenotype was intermediate between DMSO, PDB and 2DG (Figure 2A).

Clinical flow cytometry

Out of 89 individual patients in this study, peripheral blood clinical flow cytometry data was available for 35 patients. Where patients had multiple clinical flow cytometry timepoints, flow cytometry data temporally closest to the date of high content analysis was selected. Flow cytometry was performed on a Navios flow cytometer (Beckman Coulter) and data extracted from Kaluza Analysis (Beckman Coulter, V2.1). Fluorescence intensities from cells of interest were obtained from standardized (fixed) clinical gates. The flow cytometry panels comprised the following markers: Panel 1: Kappa FITC, Lambda PE, CD19 ECD, CD38 PC5.5, CD20 PC7, CD34 APC, CD23 A700, CD10 A750, CD5 PB, CD45 KO; Panel 2: CD57 FITC, CD11c PE, CD8 ECD, CD3 PC5.5, CD2 PC7, CD56 APC, CD7 A700, CD4 A750, CD5 PB, CD45 KO. All antibodies were purchased from Beckman Coulter (Miami, FL, USA). The per-cell fluorescent intensity for CD38, CD5 and CD11c were extracted from sequentially gated CLL populations defined as either: Panel 1: time, singlets, viability (forward scatter vs. side scatter), CD45⁺, lymphocytes (CD45 vs. side scatter), CD19⁺. Panel 2: time, singlets, viability (forward scatter vs. side scatter), CD45⁺, lymphocytes (CD45 vs. side scatter), CD3⁻CD56⁻ (i.e., inferred non-T cells and non-NK cells).

Clinical flow cytometry data was analyzed using pairwise Kolmogorov-Smirnov tests (KS tests). Intensity distributions were computed for patients assigned to individual clusters and KS statistics calculated between pairs of patients to assess the within cluster similarity of CD5, CD38, and CD11c intensity distributions. These values were employed to hierarchically cluster patients in groups of distribution similarity (Figure S4). To compare distributions of intensity between clusters of patient samples, the fluorescent intensities for all cells for all of the patients were pooled per marker and used to calculate deciles. Then for each cluster the number of cells in each decile was calculated and expressed as a percentage of the total number of cells in that cluster.

Quantification and statistical analysis

Image analysis and machine learning based data reduction

The experimental approach for image-based cell profiling involved a multi-step process of feature extraction and data aggregation for preliminary image analysis. Each patient sample was analyzed on a single 1536-well imaging plate by acquiring a collection of 3000 micrographs, collectively representing roughly 3 million cells. Micrographs of individual cells were identified by segmentation and measurement of morphological and texture-based fluorescence intensity features that were then analyzed using Harmony 4 · 9 (PerkinElmer) (Table S2). The 110 image features included 47 morphological measurements, 24 fluorescence intensity measurements, and 39 fluorescence texture features. From these features, a reduced set of important features (Table S3) were selected iteratively based on the relative importance scores from random forest classifiers designed to separate the different control treatments including: negative control cells (DMSO treated), dead cells (DIM), cells undergoing apoptosis after being treated with the pan-kinase inhibitor staurosporine (STS) plus venetoclax (DID) described above and the other controls including: ER stress (Tunicamycin, TN); quiescence (2-deoxyglucose, 2DG); activation of protein kinase C (phorbol 12, 13-dibutyrate, PDB); alteration of proteasome activity (Lenalidomide, LEN) (Table S4). In addition, a change in personnel and preparation of new dye stocks resulted in a batch effect in the data that was eliminated by training a classifier to distinguish cell images from the two ‘batches’. Top features from this classifier that were also being considered for the cell phenotype classifier were removed to generate the final list of 38 features.

Characterizing nuanced cell states induced by each control compound posed challenges. Establishing a definitive ground truth for the underlying biology was not possible due to the small number of patient samples compared to the phenotypic diversity observed in response to various cellular stresses in heterogeneous patient cell populations. This inherent variability among patient responses to control drugs and the off-target effects of all drugs, precluded identifying a single concentration universally eliciting a uniform response. Therefore, rational concentrations of these control compounds were determined through titration experiments, ensuring that the selected control compounds induced phenotypic states discernible from one another using the 38 phenotypic features measured.

Subsequently, every individual CLL cell image from untreated and treated patient samples was classified into one of seven phenotypic states (or bins), thereby characterizing each patient sample (DMSO) or sample response to treatment (other compounds) as a distribution of cells among the seven phenotypic bins (Figure 2B). This approach not only preserved cell-level heterogeneity but enabled making use of this heterogeneity in subsequent clustering experiments. Linking image features to seven high-level phenotypic states drastically reduced data dimensionality as each treatment response amongst the 133 patient samples was reduced from thousands of individual cells, each with 110 image features, to a distribution in 7 high-level feature bins. In addition, by training on responses to compounds with well-characterized mechanisms of actions we also connect the image data with biologically interpretable context, as is done in other cell painting experiments.^12,18 While each high-level phenotypic bin describes the dominant cellular responses, it should not be simply equated to the specific physiological response in the patient samples. Nevertheless, the descriptions used here, while approximate, provide useful uncomplicated descriptors. In this way the high-level phenotypic bins provide insights of chemo-responses beyond live or dead and effectively enable clustering of patient samples.

To determine the extent to which the control compounds altered the percentages of the cell phenotypes used as features, we plotted the classification of patient samples after treatment with each of the control compounds. Training was performed with a random set of 40 out of the 133 patient samples. Therefore, when tested against the control compounds >70% of the patient samples had not been used in training. As expected, treatment with the control compounds increased the percentage of cells assigned to the corresponding phenotype (Figure S2A). For example, for patient samples from cluster 0, the DMSO control resulted in ∼11% of cells classified as most resembling 2DG while treatment of the same patient samples with 2DG resulted in ∼26% of cells being assigned as 2DG (Figure S2A).

To assess the use of image features to predict clinical risk status, the 110 original features were normalized and the top 40 features associated with risk status (high and low) using Wasserstein distance were selected. These features were used to calculate the mean Wasserstein distance between patients, producing a patient similarity matrix that was used to hierarchically cluster patients or as an input for different machine learning models in Python: Logistic Regression, Decision Tree, Support Vector Machine, K-nearest Neighbors, AdaBoost, Bagging Classifier, Random Forest, Gaussian Naive Bayes, and Extra Tree Classifier. All these models showed low performance, which was evaluated according to different measures: accuracy, precision, recall, and F1 score.

Additionally, we performed hierarchical clustering to evaluate cluster stability by removing one patient and reclustering the rest. Thus, we calculated the frequency in which a patient was classified as high risk. A patient showing a consistent percentage of high-risk classification (∼0% or ∼100%) was considered as a stably clustered patient. Although the patients tended to be hierarchically clustered in the same clusters, we observed low overall performance in terms of prediction, evaluated as a measure of correlation between percentage of high-risk classification observed across patient dropouts and the true labels.

Unsupervised clustering of image data

After assigning each cell in micrographs of stimulated (3S CLL) negative control (DMSO treated) patient samples to one of the 7 phenotypes by Random Forest classification, the resulting phenotypic distributions were then clustered using K-means⁴⁰ and affinity propagation⁴¹ methods to determine whether there were common yet differentiated responses to incubation in 3S media. The identification of clusters of control patient samples were hypothesized to reflect different cohorts of patients within our dataset.

K-means clustering

K-means is a centroid-based clustering algorithm that partitions data into K distinct clusters by minimizing the Within-Cluster-Sum of Squared Errors (WSS) between data points and their assigned cluster centroids. Based on the total number of patient samples and initial results in which the WSS for different values of K were calculated from 1,000 iterations with random starting centroids, we determined the optimal range of K values was between 4 and 8. If the clustering was due to inherent heterogeneity of the patient samples, then the clusters should be relatively stable. To test this, ‘consensus patient clusters’ were identified by running K-means 100 times for different values of K and determining the frequencies with which patient pairs remained in the same clusters.

Cluster stability measured by cluster similarity

To identify robust and stable patient clusters, K-means derived consensus patient clusters for values of K between 4 and 7 were tested for their stability over an additional 99 iterations in which 10% of the patient samples were omitted at random (drop-out test) (Figure S2B). The more stable a cluster is, the higher the frequency with which patient samples remain in similar clusters. To identify similar clusters from each drop-out test, the results of each clustering were compared with the original clustering of all of the data using Jaccard similarity with a cut-off of 40%. Comparing the cluster stability results indicated that K = 5 was the optimal number of clusters for our dataset.

Cluster stability measured by patient entropy

As an alternative approach to determining the optimal value for K from this data, each patient was represented by a vector with the cluster memberships from each run. The patient entropy H(p) was calculated from 100 separate clusterings using the following equation.

$$H\left(p\right)=-\sum _i{p}_i{\log }_2\left(p_i\right)$$

Where pi is the probability of cluster membership i. We calculated the probability of a cluster membership as the fraction of times a patient sample was assigned to a class. The smallest possible value for patient entropy is 0, which occurs when all cluster memberships in the patient vector are the same, and the cluster membership for that patient is said to be stable. Entropies were calculated for different numbers of clusters (K = 5 to 8) of patient samples. K = 5 was selected as optimal because it resulted in the lowest total entropy.

Affinity Propagation

As an independent method, Affinity Propagation (AP) was used to cluster the data. AP is a clustering algorithm that identifies centroids (representative data points) and assigns other points to these centroids, forming clusters. Unlike other centroid-based methods (e.g., K-means clustering), AP does not require pre-specifying the number of clusters. Instead, it determines clusters through message passing between data points. Thus, hyperparameter tuning of the preference value resulted in the cluster number increasing exponentially with values beyond −5000, a sign of meaningless sub-clustering. Similar to K-means this preference value search resulted in an optimal cluster number of 5.⁴² Consistent with this result, the AP derived patient clusters shared a high degree of similarity compared to the clusters obtained by K-means as shown using a Sankey plot to track patient flow between the two clustering results (Figure S2B). In 2 independent clustering experiments around 97% of the patient samples stayed within the same clusters using the two techniques.

Phenotypic enrichment analysis

To determine the statistical significance of the enrichment for individual phenotypes within the clusters, the 7 phenotypes were subdivided into deciles bins (0, 10, 20, …, 90) and the number of patient samples within each cluster that fell into each phenotypic bin was determined. To define the enrichment of the 7 observed phenotype patterns, the expected cell distribution across the sample clusters in each phenotypic bin was determined by Monte-Carlo based simulation. That is, we evaluated the distribution of cell micrographs in each phenotype bin after shuffling the sample cluster labels 1,000 times. We then compared the observed cell distribution among phenotypic bins to the expected cell distribution based on the shuffled data. The enrichment Z score was calculated by comparing the observed results (Obs) to the mean and standard deviation (SD) of the expected distribution (Figure S2D) according to the following:

$$\mathrm{Z}-\text{score}=\frac{(\text{Obs}-\text{mean})}{\text{SD}}$$

Phenotypic similarity between drugs

To evaluate similarity between drugs, a query drug (columns) and 3 other testing drugs were randomly selected as a training set, while keeping a balanced number of query and testing sample wells (∼50% from each category, that is, 50% from query drug and 50% from testing drugs) representing patient samples. A hundred XGBoost models were trained via random sampling patient sample wells from the above categories, with replacement, training the model to predict if a sample well was treated with the query drug or not, by using the 7 phenotypic states as input features. Each XGBoost model functions by integrating multiple decision trees to construct a strong learner. XGBoost builds decision trees sequentially, where each tree aims to correct the errors of the previous ones.⁴³ Thus, several decision trees are trained on subsets of the data, and the predictions from each tree are pooled to form a final prediction.

In parallel, a new testing set (rows) was selected among drugs not used for training the model. Then, for the 100 XGBoost models, we evaluated for each new testing drug the percentage of sample wells predicted as the query drug as a proxy of phenotypic similarity between drugs (see Figure S6). As expected, this process produced a similarity matrix with a diagonal of 100% similarity between identical drugs. Following, the similarity matrix was used to hierarchically cluster drugs in an unsupervised way. To further validate our model and to exclude potential overfitting issues, we applied a hierarchical clustering to both rows and columns using phenotype values, which are the percentages of cells assigned to each phenotype. That additional hierarchical clustering (not shown) confirmed the neighbors structure found in the XGBoost model.

Statistical analysis

Continuous variables were summarized as means and standard deviations or medians with range and compared using unpaired t test or Kruskal-Wallis test as a non-parametric alternative to a one-way ANOVA, and the p-values adjusted using the Bonferroni method.

A p-value <0.05 was considered statistically significant. All statistical analyses were performed using R (version 4.4.2, the R Foundation for Statistical Computing, Vienna, Austria) and R Studio (version 2024.09.1 + 394) or Python (version 3.12.5).

Published: November 25, 2025