The cellular hierarchy of acute myeloid leukemia informs personalized treatment

Multiplexed imaging and deep-learning of AML cell heterogeneity

To assess the functional impact of AML cell maturation heterogeneity on response to first line therapy (3+7, cytarabine plus daunorubicin), we collected a total of 174 samples from 44 newly diagnosed AML patients (Figure 1A, Table 1 and Table S1). Specifically, we prospectively collected bone marrow (BM) and peripheral blood samples at three different time points per patient: Prior to first-line induction chemotherapy, after hematopoietic recovery of the first and second induction cycle (Figure 1A & B). All samples were processed directly on the same day of biopsy and stained after 24 hours of incubation with drugs and control compounds. The patient cohort spanned a diverse set of WHO subclassifications (Table 1) and underlying molecular entities (Table S2). The median age of the study population was 57 years and the sex distribution was 64% male vs. 36% female. According to ELN 2022 risk categories ²⁸, 20% of the patients were classified as favorable, 36% intermediate and 43% adverse. 77% of the patients achieved a complete remission after the second induction cycle of which 41% still had detectable minimal residual disease (Table 1 & Figure S1A).

• Download figure
• Open in new tab

Figure 1.

A, Circos plot overview of the newly diagnosed AML prospective patient cohort (174 samples from 44 patients). Concentric circles from outermost to innermost show the disease subtype, the ELN risk classification, the tissue type of each sample, patient sex, time point of sampling, clinical patient response after the first and second induction cycle, if a sample was sequenced and patient age at sampling. See Table 1 and Supplementary Table 1&2 for more cohort information. B, Workflow integrating ex vivo drug sensitivity screening (pharmacoscopy) with multiplexed single cell phenotyping, clinical data and direct molecular associations. Patient blood and bone marrow samples of newly diagnosed acute myeloid leukemia (AML) patients are taken over three time points and mononuclear cells are isolated and purified. Cells are seeded in 384-well plates, containing an indication-specific drug library. After 24h incubation, cells are fixed and stained with a multiplexed antibody panel and imaged by automated confocal microscopy. A deep learning pipeline consisting of several convolutional neural network (CNNs), classify each cell into debris, 4 cancer populations, T-cells or negative stained cells. T-cells get additionally evaluated for polarized or round morphologies indicating their activation status. Finally, ex vivo drug response prediction and a multi-omics analysis are calculated to identify optimal treatment options and clinically relevant subpopulations. C, Example image of multiplexed staining immunofluorescence on a AML sample at diagnosis. Green: CD3, CD117; Orange: CD34; Red: CD117, CD33; Blue: DAPI. Imaged at 20x resolution. D, Representative example images of curated cells for AML blast populations (CD34+ and CD33+ single positive cells (H0 and H3), combinations of CD33+/-CD117+ (H2) or CD33+CD34+ +/- CD117 (H1) marker positive cells, activated T-cells (T_P) and conventional T-cells (T_O) as well as cells negative for all stained markers (Negs). The size of crops is 48×48 pixels (14.4×14.4µm). E, Confusion matrix of CNN performance across 1000 test cells per class.

To characterize different functionally divergent AML subpopulations and associated non-malignant immune populations, we adapted our recently developed high-content multiplexed immunofluorescence approach ²⁹. This new workflow enables the simultaneous integration of high-throughput drug screening with multiplexed single-cell phenotyping and associated molecular measurements in a single round of staining and imaging. To identify differentiation stages of AML blasts, we stained for the surface markers CD34, CD117 (c-kit) and CD33 (Figure 1C). CD34 is known to be expressed on undifferentiated AML blasts, including leukemic stem cells. CD117 is expressed on less differentiated AML blasts at different stages of early lineage commitment, while CD33 is present on more mature cells. As myeloid differentiation occurs, CD33 expression levels typically increase, while CD34 and CD117 expression decreases ^30,31. In our multiplexed workflow, we additionally stained for CD3 to capture the T-cell population, furthermore allowing us to probe the inflammatory state via morphological T-cell profiling ^29,32 and to assess off-target drug effects (Figure 1C).

We employed deep learning to identify putative AML cells and classify their maturation-associated marker profile. Specifically, we used a convolutional neural network with an adapted ResNet architecture (Figure S2A), as has been previously used in this context with high accuracy ^19,29, to classify each cell into six mutually exclusive and marker-defined cell populations: Putative AML cell classes included CD34+ single-positive cells (referred to as AML class H0), CD34+ double-positive cells expressing CD33+ and/or CD117+ (AML class H1), CD34- and CD117+ cells optionally expressing CD33+ (AML class H2), and CD33+ single-positive cells (AML class H3). We furthermore classified CD3+ T-cells, and cells negative for all stained markers (Negs) (Figure 1D).

The Convolutional neural network (CNN) was trained and tested across 60,000 manually curated example images from across patients, samples, and timepoints, and achieved a classification accuracy of 96.7% in a ten-fold cross validation setting (Figure 1E & S2B, dataset available at: https://doi.org/10.3929/ethz-b-000680077). T-cells were additionally evaluated for distinct polarization and activation-associated phenotypes, as previously established (Figure S2C & 1D) ^29,32–34. We further applied a ‘clean-up’ CNN analyzing DAPI and brightfield phenotypes, which was used to filter out dead cells, cell debris, segmentation errors, and other artifacts (Figure S2D). After clean-up, remaining cells showed limited levels of apoptosis in validation experiments (Figure S2E). We also included non-multiplexed control stains on each plate, allowing us to evaluate false-positive class assignments. This revealed no signs of systematic misclassification (Figure S2F) and highlighted the excellent technical reproducibility within control wells and across drug conditions (average correlation between replicates: 0.99) (Figure S3 A & B). Lastly, we analyzed the robustness of the approach by screening samples from three independent patients with both a 24- and 48-hour drug incubation. The longer incubation only showed minor impacts on baseline population abundances and associated drug screening results (Figure S3C).

To validate our image-based deep learning workflow, we compared the results with sample-matched flow cytometry performed as part of the clinical routine. Population comparisons were possible for 27 samples at diagnosis with matched CD34, CD117, and CD33 staining. We employed two complementary flow cytometry gating strategies: one replicating the CNN-matched marker-based populations on all cells, and the other analyzing marker expression specifically within the commonly used CD45dim blast populations (Figure S4A). At diagnosis, the vast majority of CD34, CD117, or CD33 expressing cells were part of the CD45dim blast gate (90%, 95%, and 92%, respectively), and only 3% of CD45dim blasts were not captured by these markers (Figure S4B). Furthermore, we observed high correlations between the cell population abundances quantified by our image-based approach and flow cytometry across the 27 patient samples (average Pearson correlation:0.61; Figure S4C & D). Thus, our image-based approach accurately captured the majority of AML cells present in these samples at diagnosis.

In summary, through combining single-cell microscopy of AML patient samples with multiplexed immunofluorescence and cell-based CNNs, we have developed an efficient method able to robustly and rapidly profile cellular phenotypes in high-throughput imaging of AML biopsies.

Phenotypic profiling quantifies the AML cell hierarchy across patients

We next analyzed the heterogeneity in hierarchical profiles and clinical associations across the patient cohort. We observed a high level of heterogeneity in the AML cell-type composition of AML patient biopsies (Figure 2A): CD34 expressing populations (H0 + H1) showed an inverse correlation with CD33 (H3) (Figure S5A). CD34 expression (all CD34+) was associated with fewer myeloid marker-positive cells but more with marker-negative cells, whereas CD33 (H3) expression was strongly linked to the overall abundance of blasts (Figure S5A). To investigate the differential contributions of sampling time, tissue type, patient age and sex to population heterogeneity and marker expression variations, we conducted an analysis of variance (ANOVA) (Figure S5B). As expected, general subpopulation composition was mainly influenced by the sampling stage during first-line induction therapy, reflecting the response to treatment. Sampling bone marrow or peripheral blood showed no systematic influence on the cancer subpopulation composition, with only T-cell and marker-negative cells significantly differing between tissues. While patient sex did not influence population composition, age was associated significantly with tumor burden, with older patients having more AML cells relative to T-cells and marker negative cells, consistent with their known poorer prognosis (Figure S5B).

• Download figure
• Open in new tab

Figure 2

A, Deep learning-assisted phenotyping of all samples. Heatmap shows the mean population fractions across DMSO treated wells per sample. H0, H1, H2, H3, all CD34+ population fractions were normalized towards the absolute AML abundance. Samples are annotated with additional clinical annotations, as presented in the legend. B, Correlation of population abundance at diagnosis and AML cell type scores. AML cell types scores were calculated as defined by Van Galen et al. ³¹ and correlated to absolute population abundances for each sequenced sample. C, Principal component analysis of population fractions of all samples shown in A colored by clinical response to the first induction round. Shapes indicate response to the second induction round. D, Population abundance in relation to treatment outcome. Violin plots display the distributions of different population fractions measured before clinical treatment response (first or second round). CR=complete remission. MRD=minimal residual disease. NR=non responder. Samples with no response were excluded from analysis. Like in A, AML population fractions were normalized towards the absolute AML abundance. E, Blood CD34 single positive abundance (H0) at diagnosis-based stratification of progression-free (PFS) and overall survival (OS) (n=42 patients) plotted as Kaplan-Meier survival curves. Two patients were excluded because no sample at diagnosis was available. H0 high population is defined as an absolute fraction higher than 0.083. P Value asterisks are defined as following: *p<0.05, **p<0.01, ***p<10^-3.

Van Galen et al. recently identified gene expression patterns for six distinct cell types in AML along their maturation axis using single-cell sequencing. This included hematopoietic stem cell (HSC)-like, progenitor-like, granulocyte-monocyte progenitor (GMP)-like, promonocyte-like, monocyte-like and conventional dendritic cell (cDC)-like types ³¹. To transcriptionally characterize our image-based AML maturation stages, we performed bulk RNA sequencing on 103 matched samples (Figure S6A-D), and correlated cell type scores, derived from the gene sets previously published ³¹, for each sample with the image-based sample composition (Figure 2B). Globally, the bulk transcriptome at diagnosis reflected the abundance of different maturation states: as anticipated, the fraction of CD34-expressing populations (H0 and H1) demonstrated the strongest correlation with the HSC and progenitor-like cell signatures. With the loss of CD34 expression, the population loses its transcriptional HSC association, and with decreased CD117 expression, the progenitor stage as well. The CD33+ H3 population was primarily associated with GMP, pro-mono and monocyte-like cell type associations (Figure 2B). Furthermore, we correlated the image-based subpopulation abundances with a previously reported transcriptional AML stemness signature, the LSC17 score ¹¹. Again, CD34 single-positive cells (H0) exhibited a mild but significant positive correlation with the LSC17 stemness score, followed by the H1 population, while CD34-negative populations (H2 & H3) showed no significant association (Figure S6E). Thus, the image- and marker-based strategy classified AML maturation stages in line with previously reported transcriptional maturation signatures.

We next sought to investigate therapeutic resistance to drugs by investigating the known relationship between Tumor necrosis factor alpha (TNF-α) and the maintenance thus chemoresistance of LSCs ³⁵. We therefore treated cells from three AML patients at diagnosis with increasing concentrations of TNF-α in vitro, and analyzed results by image-based profiling (Figure S7A). Following 48 hours of treatment, we observed a significant increase in the CD34+ (H0) abundance (Figure S7A) and a decrease in CD117CD33 (H2) cells. While prolonged ex-vivo exposure to cytarabine resulted in a significant reduction of the CD34+ population, the presence of high levels of TNF-α substantially increased their chemoresistance (Figure S7B), supporting the notion that CD34+ cells represent a stem cell-like population with the TNF-α pathway as a potential target in AML management.

Our combined findings therefore confirm that the identified populations represent distinct stages of AML blast differentiation. The most immature, stem cell-like H0 population is marked by CD34 single-positive cells, followed by the H1 population. The more differentiated H2 cells precede the most mature H3 CD33+ single-positive cells, illustrating a clear progression in AML blast maturation in our population definition (Figure 2B).

AML cell maturation and sample composition predicts clinical outcome

LSCs hold significant prognostic importance due to their unique properties which render them resistant to standard treatments ³⁶. We therefore investigated whether our image-based AML maturation profiling could stratify the clinical outcomes across the cohort. Utilizing principal component analysis of patient sample composition, we observed that patients were clustered according to their response to first-line treatment (Figure 2C). In line with previous reports, AML blast maturation was a strong clinical predictor of treatment response, progression-free survival and overall survival, whereby high abundance of CD34 expressing cells prior to treatment strongly predicted poor treatment response (Figure 2D-E and Figure S8A-C) ^6,8–10,36. Furthermore, a high fraction of H0 stem cells in blood at diagnosis was a significant stratifier of progression free and overall survival, notably better than the ELN risk score within this cohort (Figure 2E and Figure S8D). For patients who achieved complete remission (CR) after the first round of treatment, we observed a significant relative reduction of cells with immature phenotypes. On the other hand, persistent immature cell phenotypes were observed in patients experiencing treatment failure to the first induction chemotherapy (Figure S9A). At the end of the second cycle, we noted an increase in H3 cells, in patients who achieved CR compared to patients which did not respond, likely indicating a recovery of the healthy hematopoietic environment (Figure S9B). Finally, we aimed to determine whether our findings could be replicated using standard diagnostic flow cytometry. While the correlation between the H0 population and our readout was modest (Figure S4C & D), the relative abundance of this population within all marker-positive groups prior to treatment initiation remained a significant predictor of first-line treatment response when assessed by flow cytometry (Figure S10).

Using a recently developed tool to profile T cell activation via morphological polarization ²⁹, we investigated whether additional AML clinical parameters could be predicted. An interesting side-observation was that T-cell morphology appeared predictive of several AML relevant clinical parameters. Onset of neutropenic enterocolitis (NE), a critical complication that is associated with higher mortality, was associated with the inflammatory morphological signature of T-cells. Although patients which developed NE showed an overall higher amount of T-cells prior to treatment initiation, we observed that those T-cells showed a significantly lower inflammatory morphological signature compared to patients who did not develop NE (Figure S11A). Furthermore, in patients with remaining minimal residual disease (MRD), we observed a significant increase in the inflammatory T-cell subset compared to pre-treatment levels (Figure S11B). This persisted even after hematopoietic recovery following the second induction cycle, distinguishing MRD+ patients from those who achieved a CR (Figure S11C).

In summary, our high-content, image-based profiling of the AML cellular hierarchy effectively stratifies patients based on clinical outcomes including treatment response, neutropenic enterocolitis, progression-free survival, and overall survival. Importantly, CD34 single positive H0 abundance could serve as an accessible diagnostic marker for chemoresistance to cytarabine and anthracycline-based first-line induction treatment.

AML drug response profiles are mainly determined by the degree of maturation

To gain deeper insights into intra-patient drug response heterogeneity in relation to AML maturation stages, we screened a set of 80 unique drugs and selected drug combinations (Figure 3A & Supplementary Table 3). This drug panel includes not only common first-line AML treatments but also a variety of FDA-approved blood cancer drugs, such as tyrosine kinase and immune checkpoint inhibitors. To quantify chemo sensitivities across the cellular maturation hierarchy, we calculated relative blast fractions (RBFs) as previously described and clinically validated ^18,19,23,26, for all four AML cell maturation classes within a sample. Here, positive scores indicate a higher chemosensitivity of this population compared to others, whilst negative scores denote relative ex vivo chemoresistance. Clustering the complete profiles revealed substantial drug response heterogeneity among cancer subpopulations and patient samples (Figure 3A), with the drug response of an AML blast population being strongly influenced by its maturation stage (Figure B). Notably, within-sample drug response differences following the degree of maturation strongly outweighed between-sample drug response differences (Figure 3). Furthermore, different drugs were strongly clustered by mode-of-action and class, demonstrating a high degree of technical robustness of this dataset and enabling mode-of-action investigation of less well characterized drugs (Figure 3A & B).

• Download figure
• Open in new tab

Figure 3

A, Overview of the pan-maturation drug screen. Individual data points represent relative blast fraction scores (1-RBF) of a sample specific population per single drug. Low abundant populations (fraction below 0.05) were excluded. Values were standardized (z-scored) within each treatment (n=80). Columns of the heatmap were grouped into five individual clusters (C1-C5) and the donut plots show the proportion of populations within each group. B, t-Distributed Stochastic Neighbor Embedding (t-SNE) visualization of all drug responses shown in A per individual tested treatment. C, Principal component analysis of AML population drug responses of all samples colored either by population ID or drug response to cytarabine, daunorubicin or plerixafor.

To estimate a drug’s general effectiveness against different blast populations, we calculated the fraction of patient samples displaying a significant on-target effect across all samples for each population (Figure S12A&B). While the more mature H1, H2 and H3 populations exhibited high sensitivities to common first-line treatments across the majority of samples, such as anthracyclines, anti-metabolites, or selected protein kinase inhibitors, stem cell-like H0 cells generally displayed ex-vivo chemoresistance, including to first-line treatments cytarabine and daunorubicin (Figure 3C & S12A&B). We thus probed the dataset for drugs specifically targeting the CD34 LSC-like cells. The only compounds with limited general activity against the CD34 population were Plerixafor (relative chemosensitive in 21% of H0 populations) and dexamethasone (relative chemosensitive in 24% of H0 populations) ³⁷. Dexamethasone (Dex) is a synthetic corticosteroid medication with potent immunosuppressive properties, and is often used in supportive care during AML treatment. Intriguingly, both treatments have been shown to exhibit activity against CXCR4 ^37–39, a chemokine critical for the maintenance of leukemia cells within the bone marrow. Although we did not identify a universally effective treatment against the H0 stem cell compartment, we found at least one drug per patient that the H0 population was sensitive to ex vivo (defined by an absolute reduction of the population) for the majority (65%) of patients.

We further investigated the differential impact of patient age, sampling time point, and sampling location (tissue) on drug response for each population by ANOVA (Figure S13A). Elderly patients, who are known to have a lower tolerability to first-line induction chemotherapy, showed a significantly reduced ex vivo response to nucleoside analogs cytarabine, cladribine, clofarabine, and fludarabine (Figure S13B&C). In contrast, the immunomodulator pomalidomide demonstrated improved on-target scores in elderly patients, potentially providing a better-tolerated treatment regimen for this demographic (Figure S13B). However, both pomalidomide and lenalidomide exhibited significant compartment-dependent differences, with weakened responses against cells from bone marrow compared to blood (Figure S13D &E). In contrast, gilteritinib showed stronger on-target responses in bone marrow samples, possibly indicating a higher sensitivity to FLT3 inhibition in this tissue context (Figure S13D).

Lastly, we evaluated to what degree drugs activated T-cells ex vivo. Venetoclax was recently shown to directly activate T-cells and enhance T-cell-mediated antileukemic activity ⁴⁰. Consistently, in our screen, venetoclax also emerged as one of the top activators of T-cell activity but surprisingly midostaurin and kinase inhibitors alvocidib and crenolanib emerged as very potent activators of T-cell activity, revealing a potential new anticancer mechanism of these drugs (Figure S13F).

In summary, we systematically analyzed the drug response heterogeneity and sensitivity profiles of various AML maturation stages, revealing maturation to be an extremely strong determinant of drug sensitivity. This high dimensional dataset, featuring differential population, time and tissue resolution will be made accessible to the community.

Predictive power of pharmacoscopy for clinical response in newly diagnosed AML and identification of personalized treatments

We next evaluated the power of pharmacoscopy-based drug testing to predict clinical response of newly diagnosed AML patients to their first line treatment. To account for the vast heterogeneity observed across donors, we developed a ‘mean maturation score’ (MMS), which is calculated as the average response of each AML subpopulation as well as the all marker positive class and the total cell number readout (Figure 4A & S14A). Drug responses were included into the MMS only for AML subpopulations that were present in at least 0.5% of all nucleated cells in a sample in control conditions. The MMS score also considers the unwanted reduction of the T-cell population, thereby penalizing off-target cytotoxicity.

• Download figure
• Open in new tab

Figure 4

A, Cytarabine response at diagnosis. Top: Heatmap shows cytarabine response (1-RBF) for all samples at diagnosis split by population. Responses for low abundant populations (fraction below 0.05) are plotted in gray and crossed out. Columns of heatmap are sorted by the average response calculated across all populations, referred to as the mean maturation score (MMS), for more details see methods. Bottom: Sample MMS scores plotted as matching bar graphs. B, MMS all screened drugs for two selected patients, with corresponding patient numbers, clinical outcome, timepoint and screened tissue. Bar graphs show drugs ranked by the MMS. C, Violin plot shows the integrated MMS at diagnosis stratified by treatment response (n=34). Integrated MMS score was calculated by summing all MMS scores for every individual chemotherapy a patient was treated with. If multiple tissues (blood and bone marrow) were measured per patient, the average score across both samples was calculated per patient. Measurements were taken before a patient was treated. Patients without T1 samples (n=2), clinical response annotation (n=5) or patients which received drugs not in the screening panel (n=3) were excluded from this analysis. CR=complete remission. MRD=minimal residual disease. NR=non responder. D, Receiver operating characteristic (ROC) curve plot based on the integrated AUC values from responding and non-responding patients shown in A. The ROC curve shows the trade-off between sensitivity and specificity for the optimal MMS cutoff stratifying responders from non-responders. The area under the curve (AUC) is 0.91. E, Comparison of AUC rankings for different drug response readouts. AUC values were calculated based on ROC curves generated from different integrated drug response measurements of responding and non-responding patients. The AUC values ranged from 0.5 to 0.92, with the highest AUC value indicating the most accurate drug response readout in distinguishing between responding and non-responding patients.

As an example, Patient AML008 demonstrated significant on-target ex vivo effects for a broad panel of antineoplastic agents, including first-line treatments cytarabine and daunorubicin (Figure 4B left). Subsequent treatment with the standard ‘7+3’ regimen led to a complete remission for this patient. In contrast, Patient AML015 showed strong off-target effects for both agents ex vivo, and subsequently no clinical response to the chemotherapy combination of cytarabine and daunorubicin (Figure 4B, right). However, ex vivo drug testing identified sensitivity to both the protein translation inhibitor omacetaxine and immune checkpoint inhibition, suggesting an alternative personalized treatment regimen might have yielded better clinical outcomes for this patient. To evaluate the predictive power of pharmacoscopy in our whole AML cohort, we integrated the MMS drug responses for each individual chemotherapy treatment a patient received (on average 3.8 drugs per patient) (Figure 4C). We next related the response to the first induction round with the average integrated MMS response across both patient bone marrow and blood samples at diagnosis. The integrated drug sensitivity score successfully distinguished non-responders from responders with an area under the ROC curve of 0.91 (Figure 4C&D, S13A), highlighting the predictive power of our ex vivo functional drug testing platform. Furthermore, systematic comparison of the AUCs for different drug response readouts showed that considering the cellular AML hierarchy is the most accurate pharmacoscopy-based response prediction for this cohort, while bulk cell death performed barely above random (Figure 4E). Interestingly, we observed that the predictive accuracy of our readout improved when we also considered the drug responses of the marker-negative population as a target, indicating the presence of potential malignant cells or other cancer-associated cells within that class (Figure 4E).

In light of the accurate predictions of first-line treatment response, we were interested in determining if we could identify alternative treatments that might have been effective for non-responding patients (n=13; NR to 1st cycle). While cytarabine had a minimal MMS for these patients, we could additionally identify on-target treatments for every non-responding patient. Here, the most effective individual agents identified in our screen demonstrated significantly higher on-target MMS values than cytarabine (Figure S14B). These individual compounds varied considerably between patients, and included omacetaxine, panobinostat, ipilimumab, a combination of cytarabine and fludarabine, and alvocidib (Figure S14C). Furthermore, we ranked the drugs based on their effectiveness per patient at diagnosis to identify treatments that consistently ranked highly across the entire cohort (Figure S14D). In this analysis, alvocidib emerged as a top-performing drug with favorable ex vivo drug scores. Alvocidib is a CDK9 kinase inhibitor ⁴¹ and currently in clinical trials with initial promising results and could provide an additional complementary candidate for more effective first-line treatments ⁴². Also Omacetaxine, cladribine, clofarabine, crenolanib, venetoclax, and azacytidine demonstrated promising ex vivo drug rankings across the donor cohort (Figure S14D). Interestingly, Venetoclax is already in use in combination with other agents, such as the hypomethylating agent azacitidine, which led to improved response rates compared to previous regimens ^43,44. In conclusion, our study showcases the potential of pharmacoscopy in predicting treatment response in AML patients, allowing for more personalized approaches to therapy. By taking into account the targeted reduction of the entire AML cellular hierarchy, our method accurately distinguishes between responders and non-responders in a prospective observational setting. The identification of alternative treatments that might have been effective for non-responding patients highlights the potential for optimizing treatment strategies for AML patients, ultimately improving clinical outcomes. With agents such as alvocidib and venetoclax demonstrating promising results, there is potential for developing more effective first-line treatment options for AML patients in the future.

Improving response prediction and treatment selection in resistant/relapsed AML

Lastly, we aimed to validate our findings in a relapsed/refractory (r/r) AML setting. We recently established that pharmacoscopy can be employed for therapy selection in rr-AML patients with no registered treatment options and demonstrated clinical benefit in heavily pretreated and frail patients ²³. We now evaluated whether we could retrospectively improve response predictions for this rr-AML cohort in a post hoc analysis. Most patient samples were analyzed by pharmacoscopy utilizing a CD33, CD34 & CD3 antibody staining panel and we therefore attempted to define relevant AML hierarchy populations based on combinations of the different blast markers used. Similar to the drug response calculation in the newly diagnosed AML cohort, we then calculated the average response scores across all present populations.

In this analysis a relative reduction of marker-positive cancer cells across all subsequently received treatments showed promising but not statistically significant trends. On the other hand, the integrated mean maturation score (MMS) significantly stratified patients achieving complete remission from those with progressive disease (Figure S15A). In contrary to newly diagnosed AML, we realized that in this r/r AML cohort, very small populations were less meaningful, and the most optimal separation was achieved by taking into account only more abundant AML subpopulations as well as the all marker-positive cell drug responses and the total cell number drug responses (Figure S15B). These findings highlight the importance of considering functionally different populations also in an r/r AML setting, as it has a significant impact on response predictions. In light of our findings, we highlight the case of patient P31 who received a pharmacoscopy-guided treatment with Cladribine, Cytarabine, and Venetoclax, and achieved a complete remission. Although cytarabine and cladribine induced a reduction in the overall cancer marker-positive population, including the CD33-expressing population, they did not effectively target the CD34 population. In contrast, venetoclax demonstrated strong ex vivo results against this specific population, emphasizing the importance of subpopulation resolution in personalized treatment selection (Figure S15C).

Study design and participants

For this prospective, non-interventional, single-center study, samples and clinical data were collected from patients with newly diagnosed acute myeloid leukemias undergoing intensive induction chemotherapy. The samples were taken at three timepoints: prior to induction, after regeneration of the first chemotherapy cycle and after hematopoietic recovery following the two cycles. Patients were eligible for inclusion if newly diagnosed or with relapsed AML who have not been given systemic chemotherapy within the past 12 months prior to inclusion; the patient was over 18 years of age; the patient was receiving intensive induction chemotherapy; the patient was able and willing to provide written informed consent and to comply with the study protocol procedures. The recruitment of patients occurred through the project leader or the team of treating physicians of the Department of Medical Oncology and Hematology of the University Hospital Zurich (USZ) during daily clinical practice, when patients were hospitalized for diagnostic leukemia work-up and initiation of induction chemotherapy. Inclusion of patients strictly occurred before the start of induction chemotherapy. Clinical responses were measured by the remission state in the bone marrow sampled after regeneration of the hematopoietic system (about 28 days after start of chemotherapy). Here, blast persistence resulting from bone marrow cytology and histology were used as indication of treatment response. If patients displayed persistent blasts in the bone marrow at the prior aplasia control (a bone marrow puncture about 14-16 days after start of chemotherapy), a second induction cycle was induced without waiting for hematological recovery, resulting in no response data. Furthermore, the lack of response data in other incidences is a result of patients dropping out of or discontinuing in this study. The research project was carried out in accordance with the research plan and with principles enunciated in the current version of the Declaration of Helsinki (DoH), the Principles of Good Clinical Practice (GCP), the Swiss Law and Swiss regulatory authority’s requirements as applicable. Ethical approval was granted by the Ethics Committee of the Kanton Zurich (CEC Zurich, BASEC-Nr: 2018-01547).

Pharmacoscopy

Bone marrow aspirate or peripheral blood tubes were provided by the team of treating physicians of the Department of Medical Oncology and Hematology of the University Hospital Zurich. To isolate MNCs, the samples were diluted in PBS (Gibco) + 2 mM EDTA (Sigma-Aldrich) and were purified using a Lymphoprep density gradient (STEMCELL Technologies) according to the manufacturer’s instructions. The resulting MNCs at the interface were collected, washed once with PBS+EDTA and resuspended in RPMI 1640 + GlutaMax medium (Gibco) supplemented with 10 % human serum (Chemie Brunschwig). The single-cell suspension of immune cells were seeded (2*10⁴ cell/well with 50 μl/well) in CellCarrier 384 Ultra, clear-bottom, tissue-culture-treated plates (PerkinElmer) containing antineoplastic agents (see Supplementary Table 3) and incubated overnight (24 h at 37 °C and 5 % CO₂). Cell number and viability was determined by use of a Countess II Cell Counter (Thermo Fisher).

For all the patients, the drug screen library included 80 single drugs and drug combinations with two to six technical replicates in total (see Supplementary Table 3). The assay was stopped by fixing and permeabilizing the cells with 20 μl/well of a solution containing 0.5 % (w/v) Formaldehyde (Sigma-Aldrich), 0.05 % (v/v) Triton X-100 (Sigma-Aldrich), 10 mM Sodium(meta)periodate (Sigma-Aldrich) and 75 mM L-Lysine monohydrochloride (Sigma-Aldrich). After 20 minutes incubation at room temperature (RT), the fixative-containing media was aspirated by use of a HydroSpeed plate washer (Tecan). The cells were then blocked (50 μl/well) with PBS supplemented with 5 % fetal bovine serum (FBS, Gibco) and photobleached for 4 h to 24 h (at 4 °C) to reduce background fluorescence by illuminating the fixed cells with conventional white light LED panels.

Multiplexed immunostaining and imaging

All fluorescent primary antibodies used to identify the target blast population were used at a dilution of 1:300 in PBS and included CD3, CD117, CD33, CD34 (see Supplementary Table 4). For nuclear detection, all antibody cocktails contained 2 µg/mL DAPI (4’,6-Diamidino-2-Phenylindole, Biolegend). Before the immunofluorescence staining, the blocking solution was removed, 20 µl/well of the antibody cocktail added and the imaging plate was incubated for 1 h in the dark (at RT). Thereafter, the staining solution was removed and PBS was added (70 µL/well). Each plate contained both multiplexed and staining control wells, the latter of which, located at each of the corners of the plate, had a reduced number of antibodies to serve the evaluation of antibody functionality and generation of the CNN-training data (see below). The 384-well plates were imaged by use of a PerkinElmer Opera Phenix High-Content automated spinning-disk confocal microscope. At 20x magnification with 5×5 non-overlapping images, covering the whole well surface, each well of the plate was imaged. The images were taken respectively from the brightfield (650-760 nm), DAPI/Nuclear signal (435-480 nm), GFP/FITC/Green signal (500-550 nm), PE/Orange signal (570-630 nm) and APC/Red signal (650-760 nm) channels. For further analysis, the raw .tiff images from the microscope were used.

Casp3 experiment

To examine cell viability on the grounds of apoptosis, additionally isolated peripheral blood mononuclear cells (PBMCs) from three samples were used (see Figure S2E). For a full description on the workflow for this experiment see paragraph “Pharmacoscopy” above. Immunostaining and imaging were performed as described above (paragraph “Multiplexed immunostaining and imaging”). Here, the staining panel for two of the samples (AML046 blood T1 and AML051 blood T1) included CD117, CD33, Cleaved Caspase 3 and CD34, while the other (AML044 blood T1) was stained with CD38, CD34 and Cleaved Caspase 3 (see Supplementary Table 3). Cells were identified by maximum correlation thresholding on the DAPI channel as described below in more detail.

TNF-α experiment

A treatment with TNF-α (Peprotech) and cytarabine (Sigma-Aldrich) was used for the induction and suppression of the CD34 single positive phenotype (H0). This experiment was performed on previously isolated (for a full description on the workflow see paragraph “Pharmacoscopy” above) PBMCs of three AML patients at diagnosis that were frozen as cell suspension with 10 % dimethyl sulfoxide (DMSO) upon isolation and stored at -80 °C. After thawing the samples, approximately 10*10⁶ cells were recovered, washed with RPMI 1640 + GlutaMax medium (Gibco) supplemented with 10 % human serum (Chemie Brunschwig), seeded (2*10⁴ cells/well with 50 μl/well) in duplicate in CellCarrier 384 Ultra, clear-bottom, tissue-culture-treated plates (PerkinElmer) and incubated for 48 h at 37 °C and 5 % CO₂. Cell number and viability was determined by use of a Countess II Cell Counter (Thermo Fisher). The 384-well plates contained an increasing concentration of TNF-α (1 ng, 10 ng, 100 ng), cytarabine in 10 µM concentration and a combination (TNF-α in 100 ng; cytarabine in 10 µM) of such, each in two technical replicates (resulting in four technical replicates per sample per condition). As control for TNF-α treatment, PBS was used, while DMSO acted as control for the cytarabine treatment. Following incubation, the plates were fixed, blocked, stained and imaged as described in detail above (see paragraphs “Pharmacoscopy” and “Multiplexed immunostaining and imaging”).

RNA sequencing

For RNA sequencing, approximately 2*10⁷ cells were washed with PBS+EDTA prior to resuspension and a 10 minute incubation (at RT) in 1 x RBC lysis buffer (Biolegend). The resulting RBC-free immune cells were then washed once in PBS+EDTA prior to being split in two technical replicates. The cells of the two replicates were each resuspended in 350 µl TRIzol reagent (Invitrogen) and then frozen (at -80 °C) until further processing. RNA extraction was performed with a Quick-RNA MiniPrep Kit by Zymo according to the manufacturer’s instructions. RNA sequencing was performed by the Functional Genomics Center Zurich (FGCZ). In brief, cDNA libraries were obtained as described here: ⁴⁵. Illumina library was obtained via tagmentation using Illumina Nextera Kit. All samples were sequenced in a single run on a NovaSeq6000 (single read, 100bp, depth 20 Mio reads per sample). Illumina adapters, sequences of poor quality as well as polyA and polyT sequences were removed from the raw reads using TrimGalore v.0.6.0 with cutadapt v.2.0 prior to alignment. Reads were then aligned to the human reference genome GRCh38, v93 (Ensembl) using STAR v. 2.5.3a. Read per genes were counted with featureCounts v1.5.0. Gene counts below a threshold of 20 raw counts were filtered and raw counts were normalized (DESeq2 ⁴⁶). Only transcripts annotated as ‘protein coding’ or ‘long non-coding RNA’ were considered in the subsequent analysis. In total 108 samples were sequenced (of which three samples are batch controls). Samples with more than 5% of counts assigned to hemoglobin and no detectable blasts were excluded (n=6).

Cell detection and single cell feature extraction

CellProfiler v2 ⁴⁷ was used for single-cell image analysis. Single cell detection and nuclear segmentation was performed by maximum correlation thresholding on the DAPI channel. To extract cytoplasmic measurement, cellular outlines were estimated by a circular expansion of 8 pixels around the nucleus. To measure the local intensity background around each single cell, a second larger expansion of 30 pixels was performed. Standard CellProfiler raw fluorescent intensities were extracted, log₁₀ transformed and normalized towards the local cellular background as described in Vladimer et. al., ⁴⁸.

CNN dataset generation and normalization

The curated dataset used for the 6 class cancer CNN was generated as described in ²⁹. In brief, 48×48 pixel wide single cell subimages were generated across all five measured channels. Single cell images were then manually curated for their respective class using custom Matlab scripts and included cells from all patients, both tissues and three timepoints. In total a dataset of 62318 hand selected images were generated. All intensity values as input for the 6-class Resnet were log₁₀ transformed.

The T-cell morphology CNN dataset was generated in the same manner, creating a total of 3545 (20078 all) T_ACT cells and 10596 (65918 all) T_CON cells.

Convolutional neural network (CNN) architecture, training and data augmentation

A 6-class 71-layer deep convolutional neural network with an adapted ResNet architecture (Figure S2A) ⁴⁹ was implemented using MATLAB’s Neural Network Toolbox of Version R2021a. 48×48 pixel and 5 channel input images were used. The complete training and test dataset contained a total of 60000 cells with 10000 cells per class. To evaluate CNN performance the full dataset was randomly split into a training set and a test set (Test data with 10% of data; 1000 cells per class). The 6-class CNN was trained using randomly initialized weights and biases and the adaptive learning rate optimization ‘ADAM’. The network was trained for 20 epochs with an initial learning rate of 0.001 which was dropped every 5 epochs with a factor of 0.1. Furthermore, a mini batch size of 512 images and L2 regularization with 0.001 was applied. To further strengthen generalization, input images were augmented in each iteration. Here, images were randomly rotated in 45-degree steps with an additional possibility to be also flipped vertically or horizontally.

T-cell polarization

The T-cell morphology CNN was trained to classify fixed and permeabilized human T-cells as T_P or T_O cells via DAPI, Brightfield and CD3 48×48 pixel (14.4µm x 14.4µm) single cell crops. The training set comprised 3301 (16777 all) T_P and 9805 (56113 all) T_O cells, derived from primary human samples. Testing was performed on 244 (1363 all) T_P and 791 (4638 all) T_O cells (approximately 10% of the training set). The network employed a 39-layer CNN based on an adapted ResNet architecture. Training was performed using the same parameters as described for the 6-class network. The network was trained five times to assess training robustness, and the most accurate network was selected for downstream analysis. These CNNs were used to classify individual T-cell crops into their output classes.

The fraction of T_P cells was normalized towards the absolute fraction of T-cells to represent the relative polarization of T-cells.

Quality control and cellular cleanup

For quality control, we applied a custom small ‘Cleanup CNN’, which evaluated each individual cell for correct segmentation, contaminants or artifacts. Also here an adapted ResNet architecture was used (Figure S2A). The cleanup CNN was trained for 30 epochs, with a minibatch size of 512 images utilizing adaptive learning rate optimization ‘ADAM’. The training was started with an initial learning rate of 0.001 which was dropped every 5 epochs with a factor of 0.1. L2 regularization of 0.001 was employed. The hand-curated training and validation dataset (10% of the data) consisted of 60000 individual, 48×48 pixel and 2 channel (DAPI and brightfield) images. During training images were randomly rotated by 45-degrees and mirrored vertically or horizontally per iteration. In all CNN classifications, 48×18 pixel sub-images around each nuclei-center were generated. Cells closer than 25 pixels to the border of an image were excluded from all classifications.

Drug response analysis & response prediction

Relative blast fractions (RBFs) were calculated as the fraction of cells of population X (identified by the CNN) after drug treatment divided by the average fraction of population X cells measured in control wells. Antibody-based treatments were normalized towards their respective isotype control whereas all other drugs were normalized towards DMSO. For calculation of ex vivo drug responses, all RBF values per sample were averaged over technical replicates, and subsequently zero-centered (1–RBF). Thus, a positive score represents a relative reduction of that population (on-target effect). If a drug eradicates all cells of this population without reducing other populations, the RBF score will be equal to 1. A RBF negative score indicates relative ex vivo chemoresistance of that population and killing all non-target cells, the score goes to negative infinity however values will be capped at -1. If a drug destroys both target and non-target cell populations at an equal proportion or has no effect on either population, the score becomes 0.

Mean maturation scores were calculated by averaging the RBF values of all present target cell populations. A cell population was defined as being present in a sample if it had a fraction higher than 0.005. Specifically, RBF values of the H0, H1, H2, H3, Negs, All marker+, T-cell and the total cell number reduction were taken into account. To punish the reduction of the healthy T-cell population, T-cell RBF were sign changed before averaging (hence reducing the T-cell population will negatively affect the score). The rationale behind this score is to screen for compounds which specifically kill all the malignant target cell populations, irrespective of their relative abundances while largely sparing healthy non-target cells, because even small malignant populations can negatively affect treatment outcome. A positive score favors drugs which relatively reduce all target populations whereas a negative score indicates incomplete target populations reduction.

Statistical analysis and visualizations

Statistical analysis and visualization of data was performed using RStudio (v4.1.0) and Matlab R2021a. If not stated otherwise, all p-values were calculated based on wilcoxon rank sum test. The distribution of overall and event-free survival in the different subgroups was estimated using the Kaplan-Meier method and p-values for the survival analysis were calculated with the Gehan-Breslow-Wilcoxon Test.