Characterization of chemoresistant cell populations improves risk stratification and therapy prediction in pediatric acute myeloid leukemia

Abstract

Most pediatric acute myeloid leukemia (pAML) patients achieve complete remission after chemotherapy, yet nearly 40% ultimately relapse and die from their disease. High risk cases are typically referred for stem cell transplantation (SCT) in first remission, but because SCT is associated with substantial toxicity and cures after relapse are rare, accurate risk prediction is critical. Here, we show that combining the detection of chemoresistant cell populations, inferred from single-cell transcriptomic profiles of paired diagnosis–relapse samples, and cytogenetic biomarkers at diagnosis with measurable residual disease after induction chemotherapy, substantially improves risk prediction. This approach identifies a patient subgroup comprising 20% of the patients who do not undergo SCT, with a 5-year event-free survival below 40%, and accounting for nearly half of all deaths in this population. Molecular characterization of these cell populations reveals therapeutic vulnerabilities and potential targeted treatment strategies for high-risk pAML, providing a framework for precision oncology diagnosis and intervention.

Early risk prediction is essential to decide the treatment procedure for pediatric acute myeloid leukemia. Here, the authors show that rare chemoresistant cell populations, detectable at diagnosis, improve risk stratification and reveal potential targets for more effective, personalized therapies.

Introduction

Over 40% of pediatric acute myeloid leukemia (pAML) patients relapse, including nearly 30% who are initially classified as low risk, and these relapsed patients have poor outcomes^1–3. Aggressive treatment options—such as high-dose chemotherapy, targeted therapies, and stem cell transplantation (SCT)—are available for high-risk patients and for those who relapse. Current methods for predicting relapse risk rely on identifying recurrent prognostic genetic variants at diagnosis and detecting measurable residual disease (MRD) after the first round of chemotherapy^1,4. SCT is frequently recommended for patients with high-risk variants or positive MRD^5,6. However, up to 30% of patients who test negative for MRD and do not carry high-risk variants relapse^3,7. Early identification of these patients—before relapse, when SCT is most effective—is therefore critical for improving outcomes⁸.

A critical barrier to relapse prediction is the high intratumoral heterogeneity of pAML, in which individual patients often harbor leukemia cells with diverse DNA alterations, gene expression signatures, protein markers, and treatment responses. While some cell populations are effectively eradicated by standard chemotherapy, others can persist at near undetectable levels after treatment. Consequently, comprehensive molecular and clinical characterization of each cellular subpopulation within each leukemia is necessary to predict therapeutic response and evaluate relapse risk^3,9.

Past efforts to characterize cellular subpopulations in rare pediatric and adult AML leveraged high-resolution RNA and protein profiling technologies^3,10. For example, single-cell RNA sequencing (scRNA-seq) profiles of adult AML patients revealed malignant cells with monocyte-like signatures that expressed T cell suppressing genes¹¹, and scRNA-seq analysis of chronic myeloid leukemia samples identified a response-predictive signature¹². Another single-cell profiling study of five paired AML samples collected at diagnosis and relapse identified common pathways associated with disease progression, including enhanced fatty acid oxidation and amino acid metabolism¹³. While these studies provided insight into the biology of pAML, they neither improved current pAML risk prediction nor proposed therapeutic strategies for high-risk pAML³.

Our prior work suggested that knowledge about the abundance of transcriptionally distinct pAML cell populations (pAML aggregates) and their response to therapies could help identify high-risk subpopulations¹⁰. Towards this goal, we analyzed scRNA-seq expression profiles of paired diagnosis–relapse pAML samples obtained from 33 pAML patients together with healthy donor bone marrow aspirates. Here we report our analysis, which identified pre-treatment pAML aggregates that were inferred to be intrinsically chemoresistant or to have acquired chemoresistance following therapy. We showed that their detection can both inform risk prediction in pAML clinical trials and predict the response of patient-derived xenograft models (PDXs) to chemotherapy. We also showed that the combination of chemoresistant pAML aggregate detection with cytogenetic risk models and MRD significantly improves risk prediction, revealing a high-risk patient population that may benefit from more aggressive therapies. Moreover, the detection of pAML aggregates that were inferred to acquire chemoresistance during therapy revealed potential drivers of chemoresistance and vulnerabilities in these cell populations. We showed that targeting these inferred vulnerabilities using biological perturbations and small-molecule inhibitors in cell lines and PDXs specifically altered and depleted the targeted pAML aggregates.

Results

Patient-specific effects predominated over longitudinal transcriptomic changes

Because of the heterogeneous response of pAML subpopulations to therapies, we used longitudinal scRNA-seq to identify cell populations that were present before and after chemotherapy, as well as pAML subpopulations with evidence for chemoresistance acquisition. We analyzed paired diagnostic and relapse bone marrow samples from 33 pAML patients and six samples from non-cancer donors (Supplementary Data 1). This included 22 Children’s Oncology Group (COG) AAML1031 patients from the Lambo cohort³, four non-cancer donors profiled by Bailur et al.¹⁴, and paired profiles of 13 pAML patients plus two non-cancer donors—the Baylor College of Medicine (BCM) cohort. The BCM cohort included six Texas Children’s Hospital pAML patients, five previously uncharacterized COG AAML1031 patients, and two re-profiled COG cases that were included in the Lambo cohort. These re-profiled patients allowed for cross-cohort comparisons and reproducibility assessments.

As expected, longitudinal samples from the same patient were more transcriptionally similar to each other than to samples from different patients (Fig. 1A, B). Surprisingly, even the BCM re-profiled diagnostic samples were more similar to their matched relapse samples than to the Lambo profiles of the same specimens (PAUVAT samples S3 and S53 and PAVBFN samples S11 and S55 in Fig. 1). In contrast, non-malignant hematopoietic cell populations, including T cells and B cells, showed relatively similar transcriptomes across patients (Fig. 1C). Using a variety of representations of samples and their cellular composition, we observed both differences and similarities across cohorts, patients, time points, and cell types (Fig. 1). These findings revealed that patient identity was the dominant driver of observed transcriptomic states, indicating that a patient-by-patient analytical strategy is appropriate (Supplementary Data 2, 3).

Fig. 1. Merged single-cell RNA sequencing (scRNA-seq) profiles in pediatric acute myeloid leukemia (pAML) patients.

A Uniform manifold approximation and projection (UMAP) visualizations of merged scRNA-seq expression profiles from paired diagnosis and relapse samples profiled by Lambo et al. (DX Lambo and RE Lambo) and this study (DX BCM and RE BCM), together with bone marrow profiles from healthy donors from this study (N BCM) and from Bailur et al. (N Bailur). B The clustering of scRNA-seq data by transcriptome similarity. C A comprehensive schematic summary of the identified cell clusters matching the UMAP in Fig. 1B and indicating their inferred cell types and differentiation stages (columns), and sample sources with sample ID written using the color palette in Fig. 1A (rows). Clusters were annotated using SingleR-inferred cell types. To aid visualization, relative cell abundance was recorded using both color and marker size. Samples from patients PAUVAT and PAVBFN were profiled in both the BCM and Lambo cohorts; these were circled (Fig. 1A) and bolded (Fig. 1C). GMP: Granulocyte-monocyte progenitor cells; CMP: Common myeloid progenitor cells; MEP: Megakaryocyte erythroid progenitor cells. Source data are provided as a Source Data file.

Chemoresistant cells were inferred in most pAML cases

Using our patient-by-patient approach, we sought to address a critical but unresolved question in the field surrounding the etiology of chemoresistance in relapsed pAML patients: whether it is intrinsic or acquired. To distinguish intrinsic from acquired resistance, we developed a progression model describing the temporal dynamics of non-malignant, chemosensitive, and chemoresistant cell populations before and after therapy (Supplementary Fig. S1). This model delineates the expected presentation of non-malignant cells, chemosensitive pAML aggregates, and chemoresistant pAML aggregates in the profiles of patient samples before and after treatment.

Across all paired diagnosis–relapse samples, we identified 83 and 178 pAML aggregates in the BCM and Lambo cohorts, respectively. To align with our progression model, we classified pAML aggregates detected at diagnosis as expanded, stable, or diminished pAMLs, based on their pre- to post-therapy abundance ratios. In addition, pAML aggregates detected only at relapse were termed relapse pAMLs. Expanded pAMLs were interpreted as intrinsically chemoresistant, while diminished pAMLs with evidence for transforming into relapse pAMLs were inferred to have acquired resistance.

Applying this model, we inferred intrinsically chemoresistant pAML aggregates in 14 of 33 patients (42%; Group A, Fig. 2A, B). The remaining patients were classified as Group B. Note that this does not exclude the presence of undetected intrinsically chemoresistant pAML cells at diagnosis. Our classification was corroborated by single-cell protein expression profiling by time-of-flight cytometry (CyTOF) in five paired cases (three Group A and two Group B; Supplementary Fig. S2) and remained consistent across the two re-profiled samples despite their technical variation. Notably, time to relapse did not significantly differ between the groups (Fig. 2C). In addition, we identified diminished pAMLs that may have transformed into relapse pAMLs using trajectory analysis. This analysis supports transcriptomic transformations between adjacent cell populations across longitudinal samples from each patient (Fig. 2D; S3-4). We identified diminished pAMLs with significant evidence for acquired chemoresistance according to our model in 19 of the 33 profiled pAML patients (58% of the pAML cases).

Fig. 2. Longitudinal single-cell profiles identify patients with expanded and transforming pAMLs.

Patients from the (A) BCM cohort and (B) Lambo cohort were classified based on the presence (Group A) or absence (Group B) of expanded pAMLs (intrinsically chemoresistant cell populations). Line plots map the estimated relative abundance of cell clusters identified at diagnosis and relapse; these were categorized into 4 cell types: Diminished, Stable, Expanded, and Other. C A comparison of the time to relapse for Group A versus Group B patients; n = 33, p = 0.87 by two-sided t test, mean ± s.e.m., quartiles shown. (D) Trajectory plots based on scRNA-seq profiles of representative pAML pairs from Group A and B patients. UMAP color gradients represent timepoint (top) and transcriptome-based clustering (bottom). Predicted trajectories are illustrated with solid lines. Source data are provided as a Source Data file.

In total, chemoresistant pAML aggregates were inferred in 32 of the 33 pAML cases. Of these, one case (PAUMTZ) was inferred to have both acquired and intrinsically chemoresistant pAML aggregates. Importantly, our analysis suggested that acquired chemoresistance may be common in pAML and that nearly half of the pAML diagnosis samples may also contain intrinsically chemoresistant cells. However, the abundance of these inferred intrinsically chemoresistant cells at diagnosis was relatively low, accounting for 5-6% of the pAML cells in each sample, on average.

Gene signatures are insufficient for predicting patient outcomes

Having inferred chemoresistant pAML aggregates in 32 of the 33 pAML cases, we asked whether their transcriptomic signatures could help predict patient outcomes. Towards this goal, we evaluated previously proposed outcomes-predictive pathways and gene sets^15–19 and identified genes that were recurrently dysregulated in chemoresistant pAML aggregates (Expanded Genes, Supplementary Data 4, 5). Multiple previously proposed gene signatures were upregulated in cell populations that were inferred to be intrinsically chemoresistant, including genes associated with FMS-related receptor tyrosine kinase 3 (FLT3) and cyclin-dependent kinase 6 (CDK6) activation²⁰, the stem cell signatures LSC47 and HSC^3,17, and the MYC and oxidative phosphorylation pathways. As proposed by Lambo et al., this analysis confirmed an association between pAML chemoresistance and early hematopoietic embryonic differentiation stages (Supplementary Fig. S5).

To evaluate the potential of these gene signatures as prognostic biomarkers, we tested their outcomes-predictive ability using clinical and molecular profiles of AAML1031 patients produced by Therapeutically Applicable Research to Generate Effective Treatments (TARGET). We elected to focus on AAML1031 patients with no detected FLT3-ITD mutations to allow for treatment-independent patient outcomes comparisons. We found that all selected gene signatures could distinguish between diagnostic and relapse AAML1031 samples (Supplementary Fig. S5), but gene set enrichment analysis (GSEA) failed to predict outcomes for most pAML patients (Supplementary Fig. S6). Consistent with prior work³, stem-cell signatures were predictive only in patients with KMT2A rearrangements. Given that patients with KMT2A rearrangements had significantly higher composition of inferred chemoresistant pAML aggregates (Supplementary Fig. S7), we reasoned that successful GSEA analysis may depend on high abundance for chemoresistant pAML aggregates in diagnostic samples. Because most intrinsically chemoresistant pAML aggregates represented less than 5% of the pAML cells in diagnostic samples, their signal may be too weak to be reliably detected by GSEA.

Inferred chemoresistant pAML aggregate abundance improves risk prediction

Our prior work suggested that the impact of low-abundance chemoresistant pAML aggregates could be detected in bulk RNA profiles using deconvolution⁹. We therefore tested whether the estimated abundance of our inferred chemoresistant pAML aggregates is predictive of AAML1031 patient outcomes. In total, we inferred five BCM- and eighteen Lambo-cohort chemoresistant pAML aggregates whose inferred abundance each significantly predicted both overall survival (OS) and event-free survival (EFS) of AAML1031 patients (adjusted p < 0.05; predictive aggregates). Of these, four and seven aggregates were inferred to acquire chemoresistance in the BCM and Lambo cohorts, respectively. To minimize the analytical uncertainties introduced by integrating these two cohorts, we elected to infer the abundance of predictive aggregates in parallel. As expected, the inference of their total abundance—calculated as the sum of the abundances of predictive aggregates in each dataset—was significantly more predictive of OS and EFS than that of any individual predictive aggregate (p < 1E-4 by F-test), and the combination of the two quotients significantly improved on individual predictive abilities.

Importantly, predictive aggregate abundance improved risk prediction in AAML1031 patients beyond cytogenetics and further refined MRD-based stratification. To evaluate the benefit of predictive aggregate abundance, we tested whether it could improve on the risk-prediction framework proposed in the ongoing COG AAML1831 trial (Fig. 3A and Supplementary Fig. S8A). Validation in two independent cohorts, AML08⁹ and AAML0531²¹ (Fig. 3B and Supplementary Fig. S8B), confirmed that predictive abundance enhanced outcome classification without additional model fitting (p = 1E-4 for both EFS and OS; Supplementary Data 6). While improvements were evident across nearly all risk categories, we focused on those with the greatest potential to affect treatment decisions in COG AAML1831.

Fig. 3. Overall survival analysis of predictive pAML aggregates.

A, B Estimated abundance of inferred chemoresistant pAML aggregates (resistant cells) significantly predicted patient outcomes in (A) AAML1031 and (B) AML08 + AAML0531 cohorts, as determined by Kaplan–Meier survival analysis. Patients were stratified by MRD (darker shading indicates positive) after one cycle of chemotherapy and stratified by cytogenetics according to the COG AAML1831 protocol to Favorable, Neutral, or Unfavorable (color-coded green, blue, and red, respectively). Across independent trials, the abundance of resistant cells improved OS prediction (p = 1E-4). Solid and dotted lines represent high- and low-resistant-cell-abundance patients, respectively. Arrows indicate patients within the same AAML1831 risk categories. C, D Inferred resistant cell abundance reclassified patients in (C) AAML1031 and (D) AML08 + AAML0531 who would be allocated identical therapies under the current AAML1831 protocol. This analysis revealed higher-risk patients (purple) who would not be prescribed SCT under the COG AAML1831 protocol, with OS < 0.6 and EFS < 0.4 (Supplementary Fig. S8), and lower-risk patients (orange) who would be prescribed SCT but would achieve OS > 0.6 and EFS > 0.4 without SCT across AAML1031 and AML08 + AAML0531; p-values by FET, n: patient count. Source data are provided as a Source Data file.

Under the AAML1831 protocol, patients with Neutral cytogenetics and negative MRD and those with Favorable cytogenetics and positive MRD (LR2 patients) receive five rounds of chemotherapy, whereas patients with Unfavorable cytogenetics and negative MRD (HR patients) are allocated to SCT. We retrospectively applied the AAML1831 risk group definitions to the AAML1031, AAML0531, and AML08 cohorts, recognizing that as the definition of High Risk evolved, so too did the characteristics of patients who did and did not undergo SCT at first remission. Our analysis showed that LR2 patients with a high predictive aggregate abundance (LR2-high) had significantly worse OS and EFS than LR2-low patients (Fig. 3C, D, Supplementary Fig. S7C, D and Supplementary Data 6).

The 5-year EFS for the LR2-high patient category in AAML1031 and AML08 + AAML0531 were 38% and 27%, respectively. Strikingly, LR2-high patients represented ~ 20% of each cohort but accounted for ~ 50% of deaths among patients not prescribed SCT. Moreover, while few of the AML08, AAML0531, and AAML1031 patients were transplanted, restricting the analysis to patients known to have not received SCT in first remission confirmed that LR2-high patients had had a survival rate similar to that of HR-low patients (59% vs. 61%). In total, only eight LR2-high patients received SCT, and 88% of these patients lived (Supplementary Data 6). Moreover, the 5-year EFS was 16% for the 115 LR2-high AAML1031 and AML08 + AAML0531 patients that are known not to receive SCT. We also note that although fewer patients were classified as HR with low predictive aggregate abundance (HR-low), their EFS (>40% across trials) and OS (> 60% across trials) were dramatically better than those of other SCT-allocated patients (EFS < 30% and OS < 40%; p = 5E-4 for AAML1031). Decision-curve analysis (Supplementary Fig. S9) and multivariate analysis (Supplementary Fig. S10) confirmed the predictive value of pAML aggregate abundance.

Predictive aggregates stratify outcomes across cytogenetic classes

Given that the detection of predictive aggregates significantly improved risk prediction after consideration of both cytogenetic risk and MRD, we next explored whether their predictive ability was linked to their cytogenetic subtype. Namely, we evaluated the relationship between predictive aggregates, the cytogenetic classification of the patients from which these predictive aggregates were derived, and the cytogenetic classification of the patients whose outcomes they predicted (Supplementary Fig. S11). Because KMT2A rearrangements were frequent in the BCM and Lambo cohorts, and due to the complexity of the AAML1831 protocol, we elected to investigate the traditional cytogenetic categories employed by Lambo et al. (2023); a mapping between these cytogenetic categories and the three risk categories used in AAML1831 across trials is given in Supplementary Fig. S11C.

The cytogenetic categories that were most strongly associated with poor pAML patient outcomes were KMT2A Rearrangement and Other. Our analysis revealed that most predictive aggregates were identified in samples from these high-risk pAML cytogenetic categories. Only three (13%) were derived from lower-risk cases, including PAVGJG (inv(16)), and PAWMKP and PAWNPG (t(8;21))—all Lambo cohort cases. We note, however, that patients with the highest risk cytogenetics also accounted for most of the BCM and Lambo cohorts, 64% and 73%, respectively.

Notably, predictive aggregates that were identified in samples from high-risk pAML cytogenetic categories effectively classified the outcomes of patients with low-risk cytogenetics (Fig. 3). For example, while no predictive aggregates were sourced from patients in the Normal cytogenetics category, predictive aggregates significantly improved outcomes prediction for both MRD-positive and MRD-negative patients with Normal cytogenetics (Supplementary Fig. S11). However, predictive aggregates were significantly more abundant in patients whose cytogenetic category matched the source patients, e.g., predictive aggregates characterized in samples from KMT2A patients were enriched in other KMT2A patients (Supplementary Fig. S11). This suggests that while predictive aggregates were best at predicting outcomes for patients with similar cytogenetic backgrounds, aggregates derived from samples with high-risk cytogenetic biomarkers can also be predictive of the outcomes of patients identified as low risk by cytogenetics.

Finally, to determine whether predictive aggregates were enriched for genetic alterations, we combined analysis of whole-genome sequencing (WGS) of four patients, single-cell ATAC sequencing (scATAC-seq), and scRNA-seq^3,22. WGS identified no copy number changes in pre-treatment samples, but revealed recurrent gains in chromosomes 2, 8, 17, and 21 in post-treatment samples. Notably, inferCNV analysis of scRNA-seq data did infer enrichments for gains in chromosomes 2, 8, and 17 in pre-treatment cell populations predicted to acquire resistance (Supplementary Fig. S4). Moreover, the corresponding pAML cell populations in post-treatment samples recurrently had significantly expanded enrichment for these gains (chromosomes 2, 8, and 17 following treatment). The presence of these copy number alterations were confirmed by scATAC-seq analysis with Copy-scAT²³. We argue that focusing on a subset of cells at diagnosis revealed alterations that could not be identified in bulk DNA profiles.

Predictive aggregate abundance is predictive of clonal PDX response to chemotherapy

Having demonstrated that predictive aggregates selected based on their fit with our model for chemoresistance were predictive of clinical outcomes in retrospective analysis of multiple clinical trials, we directly tested whether they were also predictive of chemoresistance using pAML PDXs. Towards this goal, we evaluated tumor responses to cytarabine in eight models from the Pediatric Acute Leukemia Xenograft (PALeX) resource²⁴. We compared the proportion of human pAML cells in PDX bone marrow and spleen on day eight, four days after treating with cytarabine (50 mg/kg/day) or saline for four days (Fig. 4A and Supplementary Fig. S12, and Supplementary Data 7). This regimen is intended to enrich for resistant cells persisting at the nadir of disease burden, as confirmed by our pilot studies^25,26. We identified two PDXs with significant sensitivity to cytarabine (AML006 and AML005). Of the remaining six PDXs, two (AML001 and AML010) showed some response to cytarabine, but this response was not statistically significant. The remaining four, including AML903 and AML905, showed no indication of sensitivity to cytarabine (Fig. 4B).

Fig. 4. Outcome-predictive pAML aggregates predict PDX chemotherapy response.

A PDX models with established tumor burden were treated with 4-day courses of cytarabine (AraC) or saline and then euthanized for disease evaluation, eight PDX models were represented by a total of 51 animals, see Supplementary Data 7; adapted from Flaticon⁶³. B Bone marrow blast proportions revealed significant AraC sensitivity in AML006 (n = 7) and AML005 (n = 10), mean ± s.e.m.; box plots show quartiles, p values by two-sided t test. C scRNA-seq profiles of saline- and AraC-treated AML006, AML005, and four less chemosensitive PDXs revealed. D PDX-specific, but not treatment-specific, transcriptomes, with 2–3 distinct clusters per model. E Cluster-level analysis revealed AraC-sensitive clusters (C3, C5) and AraC-resistant clusters (C8, C14) in AML006 and AML005, respectively, while other PDXs showed no treatment-related cluster-abundance differences. F, G Deconvolution mapped PDX clusters to predictive pAML aggregates from the (F) Lambo and (G) BCM cohorts. AraC-resistant clusters C8 and C14 matched inferred chemoresistant pAML aggregates in patients, whereas the AraC-sensitive clusters C3 and C5 did not. Source data are provided as a Source Data file.

To evaluate clonal response to chemotherapy, we profiled representative AML006, AML005, AML001, AML010, AML903, and AML905 PDXs (two from each response type) by scRNA-seq. Consistent with our observations in human tumors, scRNA-seq profiles of cytarabine-treated and saline-treated PDXs revealed PDX-specific molecular signatures (Fig. 4C), with two to three distinct transcriptionally-defined cell clusters identified per model (Fig. 4D and Supplementary Data 8). Our analysis revealed expanding and diminishing pAML cell populations in the chemosensitive AML006 and AML005 but not in other PDXs, in which the abundance ratios after treatment by cytarabine and saline were unchanged (Fig. 4E). Namely, cells from clusters C3 and C5 accounted for 77% and 87% of the saline-treated pAML cells in AML006 and AML005, respectively, but these clusters showed significantly lower abundance in cytarabine-treated mice (p = 1E-3 by student’s t test, Supplementary Data 8). In contrast, cells from clusters C8 (AML006) and C14 (AML005) had a 3-fold increase and a significantly higher abundance in cytarabine-treated models (p = 1E-4, Student’s t test). Thus, we concluded that the abundance of C3 and C5 cells decreased after treatment with cytarabine due to sensitivity to chemotherapy, while C8 and C14 cells increased due to chemoresistance. Analogously, this analysis suggested that pAML cells that were characterized in other PDXs were chemoresistant.

The inference of chemosensitive and chemoresistant PDX pAML cell populations allowed us to evaluate the correspondence between predictive pAML aggregate abundance and cytarabine response²⁷. Towards this goal, we evaluated how well each PDX pAML cell cluster transcriptomically matched predictive aggregates identified in the Lambo (Fig. 4F) and BCM (Fig. 4G) cohorts. Our analysis revealed that patient-derived predictive aggregates—from both the BCM and Lambo cohorts—exhibited a high degree of transcriptional similarity to pAML cells in all clusters inferred as chemoresistant (Supplementary Data 9). However, this similarity was not observed in any cells from the chemosensitive clusters C3 (AML006) and C5 (AML005). This result directly supported our assertion that our predictive aggregates represent bona fide chemoresistant cell populations.

Pathways linked to inferred chemoresistance may represent therapeutic vulnerabilities

The identification of patient pAML aggregates that were inferred to acquire chemoresistance presented an opportunity to infer molecular mechanisms that may drive chemoresistance and to identify vulnerabilities in these cells. First, we identified dysregulated gene sets and regulatory modules with differential RNA expression and chromatin accessibility across pAML aggregates from the same diagnostic sample. Then, we identified which of these were enriched in diminished pAML aggregates that were inferred to acquire chemoresistance (Supplementary Fig. S13). These included pAML signatures that were tested for risk prediction improvement for AAML1031 patients (Supplementary Fig. S6) and regulatory network modules downstream from transcription factors, as identified in our pAML-specific reverse-engineered regulatory network^28–30. This analysis identified recurrent dysregulation in FLT3- and CDK6-associated gene sets, as well as pathways downstream from or associated with RXRA and RARA, MYC, G2–M checkpoint, p53, TNFα, KRAS, and ALK.

To test the therapeutic relevance of targeting these gene sets and pathways, we curated publicly available data from assays that targeted inferred vulnerabilities in pAML cells and tumors. These data included assays using biological and small molecule inhibitors to target IRF8, CHAF1B, KLF4, DNMT1, BCL6, CDK4/6, and ATF3 (Fig. 5A–F and S14A-F)^31–38. We also used small molecule inhibitors to target FLT3, CDK4/6, and RARA in patient-derived pAML cells in vitro and pAML PDXs in vivo (Fig. 5G–K and S14G–K). For both the curated data and our own experiments, we compared the observed changes—comparing targeting agents and controls—in the abundance of targeted pAML aggregates versus non-targeted control pAML aggregates.

Fig. 5. Predicted vulnerabilities in inferred chemoresistant pAML aggregates.

A–K Predicted vulnerabilities were validated using public perturbation datasets (A–F) and de novo assays (G–K), showing significant declines in the abundance of targeted pAML aggregates after treatment with biologicals or small-molecule inhibitors. The pathway or gene being tested is listed above each panel, and the control condition is shown on the left. L Aggregates not predicted to be vulnerable showed no significant changes; see Supplementary Fig. S14 for details. M Predicted vulnerability to the retinoic acid receptor agonist tamibarotene correlated with tumor cell viability in four pAML patient samples; note that samples from patients p198 and p401 were used to generate PDXs AML006 and AML001 (mean ± s.e.m.). N, O Predicted vulnerability to the FLT3 inhibitor quizartinib and CDK4/6 inhibitor palbociclib was confirmed by (N) reduced cell viability and (O) CD14 protein expression relative to DMSO controls; mean ± s.e.m. and quartiles shown, p-values by two-sided t test, n: number of tested clusters. DMSO: dimethyl sulfoxide control; dtAG: degradation tag; KO: knockout; WT: wildtype; NT: non-targeting; fl/fl: floxed/floxed; dt/dt: degradation tag; OE: CHAF1B overexpression; UF146inh: UF146 inhibitor; Quiz: quizartinib; Palb: palbociclib; Tami: tamibarotene. Source data are provided as a Source Data file.

Across all perturbations, targeted pAML aggregates showed consistent depletion upon targeted inhibition, while pAML aggregates that were detected in the same samples but were not predicted to respond to the specific agent were unaffected by the treatment (Fig. 5A–L, S14). Moreover, 60% of AAML1031, AML08, and AAML0531 patients who would be allocated SCT in COG AAML1831 had a high abundance of at least one targeted predictive aggregate (> 5% for Lambo and BCM cohort aggregates, on average) with significant response to one of the five tested inhibitors (Supplementary Fig. S15 and Supplementary Data 12). This analysis supports the therapeutic potential of these inhibitors for high-risk pAML patients, including FLT3 inhibitors like quizartinib, which also inhibits the wild-type FLT3, even for FLT3-ITD-negative patients³⁹.

Next, we sought to confirm the impact of targeting these gene sets and pathways on tumor viability by testing whether the abundances of pAML aggregates that were inferred to be vulnerable to (1) retinoic acid receptor activation and (2) FLT3 and CDK4/6 inhibition are predictive of PDX responses to these treatments. Our RNA profiles of PDX models targeted with the second-generation retinoic acid receptor agonist tamibarotene by Perez et al.⁴⁰, revealed that the abundance of pAML aggregates with suppressed RXRA- and RARA-activity predicted response to tamibarotene. Specifically, the inferred total abundance of these pAML aggregates was significantly correlated with the observed change in tumor viability upon treatment with tamibarotene in vitro (R² = 0.8, p = 0.02, Fig. 5M)⁴¹. Patient sample P401 and its corresponding PDX model AML001 had a high abundance of cells with suppressed retinoic acid activity and were highly sensitive to tamibarotene in vitro and in vivo. In contrast, patient sample P198 (PDX model AML006) and P411, lacking these pAML aggregates, did not show benefits (Fig. 5M).

Analysis of scRNA-seq data of our PDX panel (Fig. 4C, D) suggested that neither of the three clusters of the chemosensitive AML006 contained pAML aggregates that are enriched for the upregulation of FLT3- and CDK4/6-associated gene sets. In contrast, one of the clusters of the chemoresistant AML903 and both clusters of the chemoresistant AML001 contained such aggregates. This suggested that AML903 and AML001, but not AML006, may be vulnerable to the FLT3 inhibitor quizartinib and the CDK4/6 inhibitor palbociclib. Treatment of these PDX cells in vitro revealed that tumor viability and protein expression of CD14—a treatment-induced differentiation marker that is upregulated in the targeted predictive aggregates—significantly correlated with the abundance of pAML aggregates that were predicted to be vulnerable to FLT3 and CDK4/6 inhibition (Fig. 5N, O). Treatment of AML903 in vivo with the combination of quizartinib and palbociclib increased median survival by 210%, from 29 to 61 days. Note that AML903 contained predictive pAML aggregates that responded to both quizartinib and palbociclib (Fig. 5H, I and Supplementary Fig. S14H, I). Taken together, these findings demonstrated that predictive aggregates not only stratify patient risk but also nominate actionable vulnerabilities that could be exploited to target chemoresistant pAML cell populations.

Discussion

Guided by a model of pAML progression (Supplementary Fig. S1), we analyzed longitudinal scRNA-seq profiles of paired pAML samples collected before and after chemotherapy to infer and characterize pAML aggregates that are intrinsically chemoresistant or that acquire chemoresistance following treatment. Our results indicated that cell populations that were inferred to be chemoresistant were already present in diagnostic samples from at least half of the patients who eventually relapsed, although they represented only 5-6% of the total pAML population at diagnosis (Fig. 2). Furthermore, analysis of cell populations that were inferred to acquire chemoresistance suggested enrichment of copy number changes that are selected for by therapy (Supplementary Fig. S4). This suggests that either very low abundance intrinsically chemoresistant pAML cells or cells that are in intermediate states of progression towards therapeutic resistance are selected by therapy. This finding underscores the need for technologies capable of detecting rare, low-abundance chemoresistant cells prior to relapse.

To this end, we applied the deconvolution method SQUID¹⁰ to RNA-seq profiles of patients enrolled in the AAML1031, AML08, and AAML0531 clinical trials. Our analysis of patient data from these clinical trials suggested that the presence of scRNA-seq-characterized pAML aggregates that were inferred to be chemoresistant is an effective classifier for both MRD-negative patients with high-risk cytogenetics and MRD-positive patients with low-risk cytogenetics (Figs. 3, 4). Adoption of this improved risk-predictive approach has the potential to enable earlier identification of MRD-negative and additional MRD-positive patients who would benefit from aggressive treatment, including SCT and targeted therapies, before relapse, when outcomes from these treatments are most favorable. Such an adaptation of treatment protocols has the potential to improve survival for the 40% of pAML patients who currently experience poor outcomes.

In addition, we showed that molecular characterization of outcome- and chemoresistance-predictive aggregates can reveal their potential therapeutic vulnerabilities. Functional assays targeting IRF8, CHAF1B, KLF4, DNMT1, BCL6, FLT3, CDK4/6, RXRA, and ATF3 in cell lines and animal models showed that predictive pAML aggregates predicted to be sensitive to a given agent were specifically and significantly altered, while other pAML aggregates were not affected (Fig. 5 and Supplementary Fig. S14). Collectively, these findings suggest that identifying and molecularly profiling pAML populations inferred to be chemoresistant can both improve relapse prediction and inform the development and selection of personalized targeted therapies for high-risk patients. However, we emphasize that ours is a proof-of-concept study, and we do not intend to pursue clinical testing for any individual proposed targeting agents at this time. Moreover, as with any targeted therapy, targeted cells may evolve secondary resistance strategies.

The identification of predictive pAML aggregates confirmed the relevance of key pAML pathways. Yet, as previously observed³, the enrichment of these gene sets in bulk profiles of diagnostic samples failed to predict outcomes for most patients. Namely, pathways associated with chemoresistance and hematopoiesis stemness, including MYC¹, FLT3, CDK6, the oxidative phosphorylation pathway²⁵, and the stem cell signatures LSC47 and HSC^3,17, were significantly upregulated in predictive aggregates and relapse samples, but the corresponding gene sets were not significantly enriched in the diagnostic RNA-seq profiles of high-risk patients. Our analysis suggested that relapse may be driven by rare chemoresistant cell populations and that technologies to predict relapse need to detect signals produced by as few as 5% of the pAML cells. Methods akin to gene-set enrichment that rely on the dysregulation of predictive biomarkers in bulk profiles of pAML samples are ill-equipped to detect such rare populations. To overcome this obstacle, we leveraged computational technology developed in prior work, where we showed that RNA-seq deconvolution is sensitive enough to detect and estimate the abundance of rare cells¹⁰.

By applying this technology to predict outcomes for patients in the AAML1031, AML08, and AAML0531 trials, as well as to predict PDX responses to chemotherapy, we were able to infer pAML cell populations that drive relapse and patient outcomes. Our results rely on the integration of multiple independent lines of evidence. First, we developed a progression model to define the expected dynamics and distributions of pAML and non-malignant cells before and after treatment (Supplementary Fig. S1), enabling us to nominate candidate chemoresistant populations. We then tested these candidates for their ability to predict relapse and patient outcomes. To minimize overfitting, we applied only limited machine learning when training our predictive algorithm on molecular and clinical profiles of AAML1031 patients, and used no additional learning when assessing risk in the AML08 and AAML0531 cohorts. These analyses confirmed that our inferred chemoresistant pAML aggregates are predictive of patient outcomes. Finally, we validated their inferred chemoresistant nature by showing that they are molecular matches to chemoresistant PDX populations, including rare resistant subpopulations within PDXs that are otherwise predominantly composed of chemosensitive pAML cells.

Current risk-prediction strategies combine cytogenetic analysis at diagnosis with MRD detection during treatment^9,42, and our analysis confirmed the strength of these predictors. In diagnostic samples from FLT3-ITD–negative AAML1031 patients, we observed that tumors with core binding factor fusions (CBFB::MYH11 and RUNX1::RUNX1T1) or normal karyotypes were associated with more favorable outcomes than those with KMT2A rearrangements or other cytogenetic abnormalities. While most of our predictive aggregates originated from these high-risk patients, we found that they also predicted outcomes among lower-risk patients, including patients with Normal and t(8;21) cytogenetics, albeit at lower prevalence (Supplementary Fig. S11). This suggests that these aggregates encapsulate high-risk features that are independent of current cytogenetic biomarkers and risk stratification methods. Notably, integrating all three predictive approaches yielded significantly higher accuracy than any two in combination, underscoring their complementary strengths.

Our strategy for predicting patient outcomes, chemotherapy resistance, and response to targeted therapies centered on estimating the abundance of chemoresistant pAML aggregates. This builds naturally on existing approaches that assess relapse risk through MRD detection. However, instead of quantifying residual pAML cells after treatment, we estimated the abundance of chemoresistant pAML cells prior to treatment, as well as pAML populations with specific vulnerabilities prior to their therapeutic targeting. Importantly, we demonstrated the feasibility of this approach using both computational deconvolution of bulk profiles and more direct single-cell–resolution profiling. Notably, our analyses also suggested that the vulnerabilities we identified, along with potential drivers of pAML progression, are epigenetically regulated. In particular, we found that regulatory networks differentially activated in chemoresistant pAML cells are frequently controlled through chromatin accessibility (Supplementary Fig. S13).

Our current approach has several noteworthy limitations. First, our model does not account for the possibility that pAML cells may arise de novo from non-malignant cells following chemotherapy, or that convergent evolution may occur when chemoresistant cells adopt transcriptomic profiles resembling chemosensitive cells. Second, the model assumes that relapse pAML cells lacking pre-treatment chemosensitive counterparts are more likely to be chemoresistant. These assumptions may not always hold; however, the model serves only as a screening framework to select candidates, who are subsequently tested for their ability to predict patient outcomes before being established as predictive pAML aggregates. Third, although we demonstrated that pAML aggregates with predicted vulnerabilities are altered by therapies targeting those vulnerabilities (Fig. 5A–L), we did not directly establish that these therapies reduce the viability of the corresponding cells. Such treatments may instead induce transient transcriptomic state changes without compromising survival. To address this limitation, we evaluated PDX models treated with a RARA agonist and FLT3 and CDK4/6 inhibitors. In these experiments, PDXs enriched for predicted vulnerable populations showed significantly reduced viability (Fig. 5M–O), supporting the conclusion that pAML aggregates with predicted vulnerabilities can indeed be selectively targeted. Finally, some molecular categories were underrepresented in our discovery cohort, and the detection of pAML aggregates that are specific to these molecular categories will benefit from including additional patients that represent them. We also note that prospective validation of our prognostic model can be achieved with bulk RNA-seq profiling of newly diagnosed patients who are followed for treatment response and outcomes.

Methods

All presented research complies with all relevant ethical regulations of Baylor College of Medicine under IRB protocol H-24170 and animal assurance protocol AN-5764. All presented analyses using human samples were retrospective, and investigators had access to patient data a priori.

Patient samples

Bone marrow or blood samples were collected at diagnosis and relapse for pAML patients as young as 4 months and as old as 17 years, and whose families consented to banking tissue for research in accordance with the guidelines of the Declaration of Helsinki. The study is limited by the number of patients with paired samples, before and after treatment, and sex and gender were not considered in the study design. Sample sizes were dictated by tissue availability. Six diagnosis–relapse pairs were obtained from patients diagnosed and treated at Texas Children’s Hospital (TCH). Seven diagnosis–relapse pairs were obtained from the COG Biopathology Center. All patients received initial chemotherapy while enrolled in or according to the protocol used in the Phase 3 COG clinical trial AAML1031⁷. Two normal bone marrow samples were obtained from healthy donors at the TCH Pediatric Transplant Center by rinsing the discarded collection filter. All COG samples were enriched for mononuclear cells by density centrifugation and cryopreserved; these samples were subjected to fluorescence-activated cell sorting (FACS) to enrich for AML cells (side scatter [SSC] low, CD45 dim) prior to downstream analyses (Supplementary Fig. S16); antibodies used are described in Supplementary Data 11. Viably frozen samples from the TCH cases were sent for single-cell sequencing without prior sorting. To control the effects of sorting, one normal bone marrow aspirate was split into two, and only half the cells were FACS sorted to enrich for immature myeloid cells in the same gate as used for the COG cases. All cases passed quality control (QC) for inclusion in our analysis. All paired diagnosis-relapse samples with clinical annotations from Lambo et al. were included in the study. When calculating EFS, events included induction failure, relapse, and death.

Cytogenetic classification and treatment in the current COG trial

We assigned risk based on MRD and cytogenetic biomarkers according to the current COG AAML1831 trial protocol. The biomarkers used to classify patients into Favorable, Neutral, and Unfavorable cytogenetics categories are provided for readability in Supplementary Data 6. We relied on published trial data to determine MRD, which was evaluated independently in each trial, with positive MRD identified at 0.1%. Treatment allocation in AAML1831 is as follows. Patients with Favorable cytogenetics and negative MRD are allocated 4 cycles of chemotherapy; patients with Favorable cytogenetics and positive MRD and patients with Neutral cytogenetics and negative MRD are allocated 5 cycles of chemotherapy; the remaining patients, including patients with Neutral cytogenetics and positive MRD and patients with Unfavorable cytogenetics, are allocated 3 cycles of chemotherapy followed by SCT.

scRNA-seq profiling and analysis

Profiling

Both pAML patients and healthy donors provided informed consent for genomic profiling under Institutional Review Board Protocol H-24170. Healthy donor samples were obtained from discarded collection bags. Samples were flow-sorted, gating for CD45-dim and SSC low cells; 10K–15 K cells were then labeled using the 10x Genomics (Pleasanton, CA) Chromium Next GEM Single Cell 3’ Kit v3.1 and sequenced at 200 M reads per sample using an Illumina NovaSeq 6000. The resulting scRNA-seq data were analyzed by an anchor-based approach with the Seurat package v4.30. Specifically, cells were analyzed using three distinct methods. First, cell counts were log-normalized and clustered using the Louvain algorithm. Cells across all profiled and collected samples were then merged (Fig. 1A) using standard Seurat v4.30 packages⁴³. Second, cell profiles from paired samples and normal samples were normalized and integrated one patient at a time to classify pAML and non-tumor cells. Finally, patient-specific cell clusters were combined to merge similar cells within and across patients in an effort to reduce the complexity of the transcriptomic space.

Cell classification

To analyze patient-specific cell populations, we first log-normalized the count data from each patient pair containing diagnosis and relapse samples together with data from three normal bone marrow samples from two individuals (TCHnorm1 and TCHnorm2). Normal bone marrow samples were included to help distinguish malignant from non-malignant hematopoietic cells and to accommodate for the smaller number of cells associated with patient-by-patient analysis. Normalized data were then integrated by their sample types (tumor, normal) using the IntegrateData function, and cell clustering was performed on the integrated data using the Louvain algorithm at a resolution of 2.5. Clusters with fewer than 5% of the cells from each normal bone marrow sample were considered pAML candidates; inferred cell types were verified at a 100% rate as potential cancer cells. We identified expanded pAMLs as integrated cell clusters comprising cells from both the diagnosis and relapse samples. Expanded pAMLs had to include > 10% of the cells from the relapse sample. Similarly, diminished pAMLs were identified as clusters including > 10% of the cells from the diagnosis sample, whereas stable pAMLs were classified as those for which > 10% of cells from both paired samples passed QC. Clusters that included > 10% of the cells from the relapse sample only were identified as relapse pAMLs. The remaining clusters, including small clusters or clusters that were mostly composed of non-pAML cells, were identified as other cells. When combining clusters across patients, we merged clusters with similar (pseudobulk) expression profiles and argued, based on parameters set by SQUID¹⁰, that these clusters are derived from the same pAML aggregates.

Estimating cell population abundance based on bulk RNA-seq profiles

We used the SQUID framework to predict the abundance of scRNA-seq clusters and pAML aggregates in bulk RNA-seq profiled samples¹⁰. The first step of SQUID’s framework is the combination of cell clusters with similar expression patterns. This is accomplished by first estimating patient-specific cell population similarity by performing pairwise Pearson correlations of the mean log-normalized expression profile of each cluster, then, high-similarity groups that included at least two clusters with r > 0.986 were combined to produce an aggregate cluster. Since the number of effective expression profiles defining cluster centroids under this formulation is many magnitudes greater than the number of cell types in the dataset, merging cell populations with highly similar profiles allows molecular signatures from scRNA-seq profiles to indicate cell type identities. This merging operation and the removal of clusters with fewer than 50 cells reduced the cell-type space. Merged clusters were used to deconvolve RNA-seq profiles of patient samples, PDXs, and cell lines.

Cell type inference

We used the Celldex package v.2.0.0 in SingleR⁴⁴ to determine the cell type of each population. To enable higher-resolution cell classification, including the distinction between multiple myeloid cell differentiation stages, we combined reference databases, including the Human Primary Cell Atlas⁴⁵, Novershtern’s Hematopoietic Data⁴⁶, Blueprint and ENCODE⁴⁷, the Database of Immune Cell Expression⁴⁸, and Monaco’s Sorted Human Immune Cells Database⁴⁹.

Trajectory Inference Using Monocle 3

Trajectory analysis was conducted using the Monocle 3 package⁵⁰ in R to infer the temporal progression of cellular states from single-cell RNA sequencing data. We first constructed a cell data set (CDS) object and applied the preprocess_cds() function, performing normalization using size factor normalization (norm_method = “size_only”), which accounts for differences in sequencing depth across cells. Dimensionality reduction was then performed via Principal Component Analysis (PCA) using the top 100 principal components (num_dim = 100) to capture major sources of transcriptional variance. The reduce_dimension() function was used to project the data into a two-dimensional manifold using Uniform Manifold Approximation and Projection (UMAP), preserving both local and global transcriptomic relationships. Unsupervised clustering of cells was performed using cluster_cells(), which constructs a k-nearest neighbor (kNN) graph and applies Louvain or Leiden community detection algorithms to identify transcriptionally distinct cell clusters. In addition to fine-grained clustering, Monocle 3 computes partitions—superclusters of communities—via pruning of the kNN graph, facilitating coarse grouping of cells across potentially broad biological states. These cluster and partition assignments were stored within the CDS object and later used in trajectory inference. Next, the learn_graph() function was used to infer a principal graph over the UMAP-reduced space using Reversed Graph Embedding (RGE). This method learns a smooth, branched graph structure that approximates the manifold of cell-state transitions. We set use_partition = FALSE to learn a unified graph across all cells, allowing for global trajectory inference rather than isolated trajectories within partitions. The graph learning process includes loop closure, which connects distal tips in the trajectory graph when their Euclidean and geodesic distances satisfy specified thresholds (defaults: euclidean_distance_ratio = 1, geodesic_distance_ratio = 1/3). Insignificant branches were pruned based on a minimal branch length threshold (default: 10 units), and tip nodes could optionally undergo orthogonal projection to refine placement (default: orthogonal_proj_tip = FALSE). The RGE algorithm constructs the graph by identifying latent paths that minimize distortion while preserving the topological structure of the reduced space, resulting in a biologically meaningful representation of lineage progression.

PDX-profile resolution

When analyzing bulk RNA-seq and scRNA-seq profiles of PDX samples, we classified the xenograft-derived sequence read data using Xenome⁵¹, which effectively handles a mix of reads from both the host and the graft, for further analyses.

CyTOF

Five COG diagnosis–relapse pairs profiled by scRNA-seq had a sufficient number of cells to perform CyTOF. After sorting, cells allocated for CyTOF were divided into two fractions; one was stimulated for 15 min with conditioned medium from HS-5 bone marrow stromal cells to activate intracellular signaling⁵², and the other was left unstimulated. Cells were fixed in 1.6% paraformaldehyde and barcoded (Cell-ID 20-Plex Pd Barcoding Kit, Fluidigm) for multiplexing, which enables the combination of cells for CyTOF analysis to reduce batch effects. Samples were then stained with 38 surface marker antibodies, permeabilized, and stained again for eight intracellular antigens; see Supplementary Data 11 for the Standard BioTools antibodies used. Normalization beads were included with every run. Doublets and non-viable (cleaved caspase 7 + ) cells were excluded. CyTOF assays were performed by the Baylor Cytometry and Cell Sorting Core. The data were compiled, cleaned, gated, and analyzed in Cytobank. Dimensionality reduction with UMAP was performed for each diagnosis–relapse pair using all surface and intracellular markers, except for cleaved caspase 7, which was used for the exclusion of non-viable cells.

GSEA to identify Expanded–Enriched Gene Sets and drivers of pAML transformation

Expanded Genes were identified by differential expression analysis comparing each expanded, relapsed, and stable pAMLs to all diminished pAMLs in each patient separately in the BCM cohort using Wilcoxon Rank–Sum tests. Genes expressed in at least 25% of the cells with an adjusted p < 0.05 were considered significant for this comparison. Expanded Genes were defined as those significantly upregulated in at least 21 of our 52 comparisons. This cutoff was selected because it split the bimodal distribution of tallies across all candidate genes. We then used enrichr⁵³ or hypeR⁵⁴ with adjusted p < 0.05 to identify gene sets and pathways that were enriched among Expanded Genes. We reverse-engineered a pAML-specific regulatory network based on TARGET pAML gene expression profiles as previously described^55–57. This network was used to identify pAML-specific transcription factor (TF) targets that were used as gene sets for GSEA analysis. This, in turn, was used to estimate TF activity when comparing expression profiles across pAML aggregates.

Evaluating the predictive ability of models and gene sets

We used pAML clinical characterizations and RNA-seq profiles collected by TARGET and St. Jude Research Hospital—including clinical trials AAML1031, AAML0531, and AML08 to evaluate the prognostic predictive ability of gene sets and chemoresistant-cell models. The only treatment variability within this patient population was the combination of standard chemotherapy and bortezomib, but the results suggested that bortezomib did not affect overall survival⁷. On the other hand, patients with high allelic ratio FLT3-ITD mutations were treated with sorafenib, which significantly affected patient outcomes. In total, TARGET profiles of AAML1031 included 1330 diagnoses, 43 relapses, and 62 healthy donor bone marrow samples; TARGET profiles of AAML0531 patients included 144 samples with molecular and clinical annotations, whereas the St. Jude collection included 141 AML08 samples with molecular and clinical annotations. RNA-seq count data were log-normalized, and we performed GSVA⁵⁸ analysis to evaluate whether gene sets were dysregulated in diagnosis and relapse samples (Supplementary Figs. S5 and S6).

Comparisons of GSVA scores between samples from healthy donors, pAML diagnosis, and pAML relapse samples were evaluated by Wilcoxon Rank–Sum tests. GSVA scores were also used to evaluate the predictive ability of gene sets for patient outcomes. We next used SQUID¹⁰ to evaluate the abundance of pAML aggregates and non-pAML cell populations. Abundance estimates were evaluated for predictive ability for both OS and EFS by Cox regression with Bonferroni corrections for multiple testing. Outcome-predictive aggregates were composed of diagnosis sample cells and were predictive of both OS and EFS at an adjusted p < 0.01. Tests of other cell populations yielded none that were outcome predictive. Because FLT3-ITD-positive patients were treated differently on the study’s protocol, we focused our analysis on FLT3–ITD–negative patients with profiled diagnosis samples.

In survival analyses, to limit machine learning, patients were equally partitioned based on the median inferred pAML aggregate abundance. Cumulative predictive aggregate abundance in the BCM and Lambo cohorts was combined using linear regression using AAML1031 patient profiles. The resulting parameters were used to analyze AAML0531 and AML08 data without additional learning. Indeed, the detection of inferred pAML aggregate was sufficient to classify patient samples as having either low or high abundance. Survival analysis based on MRD after one or two rounds of chemotherapy split patients into MRD-positive vs. MRD-negative groups, and analysis based on cytogenetics assigned patients to one of three risk categories. Feature combinations partitioned the patient populations into discrete groups to predict total outcomes, including predictive aggregate abundance, MRD, and cytogenetics. In addition, we performed analyses using MRD after two rounds of chemotherapy and the AAML1831 risk-prediction algorithm. However, these variations did not alter the conclusions of our presented analysis based on cytogenetics biomarker aggregations and MRD alone; note that including MRD after two rounds of chemotherapy did not improve predictive accuracy, and that the AAML1831 risk-prediction algorithm is a refinement of the classic cytogenetics classification. Multivariate Cox proportional hazard models assessing pAML aggregate abundance, cytogenetics, MRD, and these factors combined were constructed using the coxph function in R.

When evaluating the dysregulation of transcription factor activity or regulatory networks immediately downstream from a transcription factor, we relied on the transcription factor’s inferred set of targets. Namely, we used the cumulative expression dysregulation of the targets of a transcription factor as a surrogate for the dysregulation of transcription factor activity (Supplementary Fig. S13 and Supplementary Data 10)²³. We also used the cumulative differential DNA accessibility of pAML signatures and transcription factor target sets based on Lambo et al. ATAC-seq profiles to evaluate regulatory modules that are epigenetically altered in pAML aggregates that acquire resistance.

PDX and cell line experiments and analysis

The sources of curated data from publicly available assays, including the targeting of PDXs and cell lines, as given in Fig. 5 and Supplementary Fig. S14, are provided in Supplementary Data 12. When evaluating predictive aggregates, we required that every aggregate evaluated be present in at least 1% estimated abundance in control cells or animals. When evaluating the response of aggregates not predicted to be targeted, random samples of aggregates with sufficient estimated abundance were chosen. The size of random samples was set to match the number of predicted target aggregates, and sampling was repeated 10 times, with the geometric mean p-value reported in Supplementary Fig. S14. The median p-value random sample was selected for presentation in Fig. 5 and Supplementary Fig. S14.

Established serially transplanting pediatric AML PDX models from the PALeX resource were used for the experiments reported in this study (https://pdxportal.research.bcm.edu/pdxportal)²⁴. Characteristics of patients from whom the models were derived and their viability after treatment with cytarabine and saline are listed in Supplementary Data 7. We used immunodeficient, human cytokine–producing mice (NSGS, MISTRG, or MISTRG6). NSGS are (NOD.Cg-Prkdc^scid Il2rg^tm1Wjl Tg(CMV-IL3,CSF2,KITLG)1Eav/MloySzJ), while MISTRG and MISTRG6 mouse lines were developed on a mixed BABL/c;129 genetic background. They are severely immunocompromised through the deletion of Rag2 and Il2rg genes, and are humanized by the knock-in of human cytokines and growth factors M-CSF, GM-CSF, IL-3, and TPO. MISTRG6 mice also have a knock-in of human IL-6. For each experiment, 6–10 mice were injected by tail vein with 2 × 10⁵ viable AML cells that were harvested from a prior passage and cryopreserved. Once mice demonstrated at least 2% human AML (hAML; human CD45⁺/CD3^‒) cells in peripheral blood, they were assigned to receive cytarabine or saline (Supplementary Fig. S12). Treatment groups were balanced for peripheral blood disease burden. Cytarabine (50 mg/kg) was administered via intraperitoneal injection once daily for four days; control animals received an equal volume of saline.

All mice were humanely euthanized four days following the last dose of cytarabine or saline control (experiment Day 8). This timeline was selected to evaluate the effects of a partial response to treatment^25,26. Mice were also humanely euthanized when they became symptomatic or demonstrated > 40% pAML in peripheral blood. One mouse from each treatment group, for 6 of the 8 models that underwent cytarabine response evaluation, was profiled by scRNA-seq. Bone marrow and spleen were harvested, and the pAML proportion was determined by flow cytometry. Bone marrow from one cytarabine-treated and one saline-treated mouse was cryopreserved and profiled. We used Dampened weighted least squares estimation (DWLS)⁵⁹ to estimate the number of cells matching outcome-predictive aggregates in each scRNA-seq cluster. We used GSVA to evaluate the enrichment of gene sets in PDX cluster pseudobulk profiles.

The effects of 1 µM treatment with all-trans retinoic acid (ATRA) in U937 and MOLM13 at 72 hours were estimated by Meier et al. with gene expression profiled by RNA-seq⁶⁰. The effects of patient sample treatment by tamibarotene (100 nM) at 24 h were estimated by Perez et al., with gene expression profiled by RNA-seq⁴⁰. We used SQUID to estimate the abundance of predictive aggregates in these samples, deconvolving RNA-seq profiles using all pAML and non-cancer aggregates as input. We note that the PDXs AML001 and AML006 were derived from patient samples p401 and p198, respectively. The correlation between predictive aggregate abundance and tumor viability was 0.95.

To test the effects of FLT3 and CDK4/6 inhibitors on chemoresistant pAML aggregates that were predicted to respond to FLT3 and CDK4/6 inhibition, we cultured cells from PDXs AML001, AML006, and AML903 in HS-5 conditioned medium and treated ex vivo with palbociclib, PF-07220060, or quizartinib (all 1 μM) or a combination of quizartinib and a CDK4/6 inhibitor for 3 days. Cells were harvested for flow cytometry determination of viability using Annexin V staining and differentiation using CD14 staining. Cells were also analyzed by bulk RNA-seq. We used DWLS to estimate the abundance of predictive aggregates in scRNA-seq profiles of saline-treated samples from these PDXs. The results suggested low, intermediate, and high cumulative abundance of predictive aggregates in AML006 (6%), AML903 (14%), and AML001 (24%). These estimates were significantly predictive of pAML tumor viability and CD14 expression in these PDXs. In particular, the correlations between predictive aggregate abundance and (1) tumor viability and (2) CD14 expression—a biomarker of myeloid differentiation—were 0.87 and 0.99, respectively; antibodies used are listed in Supplementary Data 11.

To validate the effects of palbociclib and quizartinib in vivo, we treated NSGS mice (n = 6) engrafted with AML903 with a combination of quizartinib (10 mg/kg) and palbociclib (100 mg/kg) or vehicle only (n = 5) by daily oral gavage for 14 days, and followed the mice for the length of survival. Bone marrow was collected for bulk RNA-seq. When analyzing mouse cell lines, we used the biomaRt package⁶¹ in R to retrieve human orthologs of mouse genes for making predictions about gene sets and to predict pAML aggregate abundance. Median survival of mice treated with palbociclib and quizartinib was longer (61 days) than that of mice treated with vehicle (29 days).

Evaluating potential patient responses to inhibitors

When evaluating patients who may benefit from treatment with a given inhibitor (Supplementary Fig. S15), we first identified all predictive aggregates that exhibited a significant response to the inhibitor. Patients with high inferred abundance of these aggregates were then flagged as potential candidates for therapy with this inhibitor. To determine whether an aggregate’s response was significant, we generated a null distribution of responses using aggregates not predicted to be targeted. Predictive aggregates were considered significant if their response exceeded the null average by at least two standard deviations.

Statistics & reproducibility

Sample sizes were chosen based on availability while trying to ensure sufficient power. No data were excluded from the analyses, and the Investigators were not blinded to allocation during experiments and outcome assessment.

Reporting summary

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

Author contributions

M.J.N., T.K.M., M.S.R., and P.S. designed the research and wrote the paper. A.M. Lyons-Warren managed, edited, and contributed to the writing of the manuscript. A.M.S., M.J.K., M.R., S.S., L.K., and J.S.Y. performed experiments. H-S.C., J.E., B.Z., M.R.M., M.R., M.J.N., S.U., B.B., M.S.R., and P.S. analyzed data.

Peer review

Peer review information

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

Funding

This project was partially funded by Wipe Out Kids Cancer, the Helis Medical Research Foundation’s Helis Research Program Grant, CPRIT awards RP240432, RP180672, RP180674, RP200504, RP170691, RP220646, and RP230120; European Union’s Horizon 2020 research and innovation program under grant agreement 826121; and the NIH awards R01CA258866, U24CA196173, U24CA114766, U10CA180899, U10CA180886, U24CA196173, R21CA223140, R21CA286257, CA125123, RR024574, OD036336, and OD038251. Baylor College of Medicine Advanced Cores are supported by NIH grants P01CA261669, S10OD018033, S10OD023469, S10OD025240, and P30EY002520. This project was supported by the Cytometry and Cell Sorting Core at Baylor College of Medicine, and the assistance of Joel M. Sederstrom. The COG Biospecimen Bank and PDX development were also supported by funding from the St. Baldrick’s Foundation U24CA196173, Texas Children’s Hospital Pediatric Pilot Research Fund, CURE Childhood Cancer, Leukemia and Lymphoma Society/Blood Cancer United, Target Pediatric AML/the Children’s Oncology Group Foundation, and gifts of funding from the Turn it Gold Fund. Support from the Cytometry and Cell Sorting Core includes NIH P30 AI036211, P30 CA125123, and S10 RR024574.

Data availability

The TCH pAML scRNA-seq and bulk RNA-seq datasets, as well as the COG pAML scRNA-seq dataset generated in this study, have been deposited in the GEO database under accession code Super Series GSE271137. This data is available without restriction. The processed data are available at GEO. Bulk RNA-seq data for the FLT3 inhibitor quizartinib and CDK4/6 inhibitor palbociclib treatments are also included in the same GEO accession. The COG pAML CyTOF profile data reported in this paper are publicly available in the GitHub repository mjnajafpanah/Paired-diagnosis-relapse-pAML-data-analysis [https://github.com/mjnajafpanah/Paired-diagnosis-relapse-pAML-data-analysis]⁶². Analyses of the data generated in this study are provided in the Supplementary Information/Source Data files. Note that the expression profiles of 65 protein markers were included for 12 of the TCH pAML samples, but these were not used in this study. Source data for all figures are provided as a Source Data file. Source data are provided in this paper.

Code availability

All scripts used in the generation of this manuscript were deposited in the GitHub repository mjnajafpanah/Paired-diagnosis-relapse-pAML-data-analysis [https://github.com/mjnajafpanah/Paired-diagnosis-relapse-pAML-data-analysis]⁶².

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-78283-5.