Abstract
Contemporary chemotherapy for childhood acute lymphoblastic leukemia (ALL) is risk-adapted based on clinical features, leukemia genomics and minimal residual disease (MRD); however, the pharmacological basis of these prognostic variables remains unclear. Analyzing samples from 805 children with newly diagnosed ALL from three consecutive clinical trials, we determined the ex vivo sensitivity of primary leukemia cells to 18 therapeutic agents across 23 molecular subtypes defined by leukemia genomics. There was wide variability in drug response, with favorable ALL subtypes exhibiting the greatest sensitivity to L-asparaginase and glucocorticoids. Leukemia sensitivity to these two agents was highly associated with MRD although with distinct patterns and only in B cell ALL. We identified six patient clusters based on ALL pharmacotypes, which were associated with event-free survival, even after adjusting for MRD. Pharmacotyping identified a T cell ALL subset with a poor prognosis that was sensitive to targeted agents, pointing to alternative therapeutic strategies. Our study comprehensively described the pharmacological heterogeneity of ALL, highlighting opportunities for further individualizing therapy for this most common childhood cancer.
Main
Acute lymphoblastic leukemia (ALL), the most common cancer in childhood, comprises a constellation of clinically heterogenous molecular subtypes1. The current paradigm of ALL risk stratification integrates clinical features, leukemia somatic genomic aberrations and early treatment response as measured by minimal residual disease (MRD)2,3,4,5. However, the pharmacological basis of inter-patient variability in MRD is poorly understood and the relationships between somatic genomics and drug resistance phenotypes are unclear. Thus, ALL treatment regimens worldwide almost uniformly use an identical repertoire of conventional chemotherapy drugs with little variation in timing and intensity, and are not influenced by the patient’s sensitivity to specific antileukemic agents. The limited exceptions include the addition of tyrosine kinase inhibitors for ALL with the BCR–ABL1 or ABL class fusions6. In the current era of personalized medicine, it is imperative to determine whether biologically and pharmacologically informed selection of chemotherapy can further improve ALL treatment outcomes.
MRD reflects the in vivo response to combination chemotherapeutic agents utilized in the ‘induction’ phase in clearing disease to submicroscopic levels, and is widely regarded as the most powerful prognostic factor in ALL treatment3,7. However, MRD after combination chemotherapy does not indicate which of the multiple antileukemic agents a patient is responding to, and which are producing little or no therapeutic benefit. De novo and acquired resistance of ALL cells to chemotherapy is known to be a major cause of treatment failure8. Directly assessing leukemia cell sensitivity to cytotoxic chemotherapeutic agents ex vivo (prednisolone, vincristine, daunorubicin and L-asparaginase (PVDL)), we and others described the correlation of drug resistance phenotype with treatment response9,10,11 and association with clinical characteristics and somatic genomic features (for example, BCR–ABL1, KMT2A rearrangements)12,13,14,15,16. In fact, prospective assessment of these drug sensitivity profiles was shown to have utility in predicting patients at high risk of poor treatment response. However, these studies have several limitations in the context of contemporary ALL treatment because (1) only a narrow selection of conventional chemotherapeutic agents were screened and (2) molecular subtyping was restricted to 4–5 common fusions and aneuploidy9,17,18,19,20,21,22,23,24,25. Therefore, the drug response profiles of most ALL subtypes are unknown.
The current prevailing paradigm of precision oncology is largely driven by cancer genomics, that is, therapy is selected for and matched with a given genomic abnormality26. Although this is highly informative for drugs with a clear target (for example, imatinib and BCR–ABL1 leukemias), the mechanisms of drug response are highly complex for most agents. Therefore, assessment of genetic drivers alone may not adequately translate into accurate prediction of therapy effectiveness. With few exceptions, recent trials attempting to select therapy based on molecular target identification have not consistently yielded benefit for patients27,28,29,30,31. For this reason, there is a growing interest in functional precision medicine approaches in cancer treatment individualization, such as augmenting genomic testing with direct drug sensitivity profiling using primary tumor specimens25,32,33. For example, the recent EXALT trial34 used drug profiling to match therapy in patients with aggressive hematological cancers and has shown feasibility for clinical implementation, with promising improvement in patient outcome compared to traditional therapy.
Therefore, we sought to comprehensively characterize the relationship between ex vivo drug sensitivity profiles and in vivo early treatment response across the updated taxonomy of molecular subtypes of ALL, to inform the design of optimal or even new combinatorial therapeutic strategies. To this end, we performed ex vivo pharmacotyping of 18 contemporary chemotherapeutic drugs on primary ALL cells from 805 patients with ALL comprising 23 molecular subtypes and evaluated the impact of leukemia drug sensitivity on initial treatment response as measured by MRD in the context of contemporary ALL treatment regimens.
Results
Association of clinical features with ALL drug sensitivity
The sensitivity of leukemia cells from 805 patients was measured for 18 drugs (Fig. 1and Table 1), with a total of 5,447 50% lethal concentration (LC50) measurements (Supplementary Tables 1 and 2). In parallel, we performed RNA sequencing (RNA-seq) on diagnostic ALL cells to assign each patient to 1 of 23 unique subtypes. Across all drugs, we observed wide interindividual variability in drug sensitivity, with an average coefficient of variation of 55.3% (range = 8.7–105.4%; Extended Data Fig. 1).
We first examined the LC50 distribution of each drug by patient characteristics (Supplementary Table 3). Because of the number of drugs evaluated, we also corrected for multiple testing using the Benjamini–Hochberg procedure. We found that patients ≥10 years old had a higher median normalized LC50 than those 1 to <10 years of age for L-asparaginase (0.62 versus 0.40, P < 0.001) and prednisolone (0.26 versus 0.18, P = 0.003). Patients with ≥50 × 109/L leukocyte count at diagnosis had a lower median LC50 than those <50 × 109/L for mercaptopurine (0.42 versus 0.50, P = 0.004) and dasatinib (0.86 versus 1.0, P < 0.001). Assessing by the National Cancer Institute (NCI) risk group for ALL, NCI high-risk patients (age ≥ 10 years or leukocyte count ≥50 × 109/L) had a higher median LC50 than NCI standard-risk patients (age < 10 years and leukocyte count <50 × 109/L) for only L-asparaginase (0.56 versus 0.37, P < 0.001). There were no differences in drug LC50 by sex or ancestry (Supplementary Table 3).
Drug sensitivity differs across ALL subtypes
We next sought to characterize drug sensitivities across the somatic genomic landscape of ALL. Fourteen of 18 drugs exhibited nominally significant inter-subtype variability (P < 0.05), with the exceptions being panobinostat, ruxolitinib, bortezomib and daunorubicin (Fig. 2a, Extended Data Fig. 2 and Supplementary Table 4).
ETV6–RUNX1 and hyperdiploid ALL were highly sensitive to the four drugs commonly used in remission induction (PVDL), recapitulating the known chemotherapy-responsive nature of these subtypes9,10. For these two subtypes with known favorable prognosis, the median PVDL LC50 was 0.18 for ETV6–RUNX1 and 0.26 for hyperdiploid, both significantly lower than 0.41 for the remaining subtypes (P < 0.001 and P = 0.004, respectively). By contrast, subtypes known to have poorer prognosis, BCR–ABL1, BCR–ABL1-like and KMT2A, exhibited higher PVDL LC50 (0.46, P = 0.001, 0.66, P < 0.001 and 0.54, P < 0.001, respectively compared to ETV6–RUNX1). DUX4, a recently discovered subtype with favorable outcomes, had an intermediate PVDL LC50 of 0.33 (P < 0.001 compared to ETV6–RUNX1), with a significantly higher L-asparaginase LC50 (0.73 versus 0 in ETV6–RUNX1, P < 0.001; Fig. 2b) consistent with its known slow MRD clearance during early remission induction featuring only these 4 drugs4.
We also observed highly subtype-dependent patterns of sensitivities to targeted agents. Hyperdiploidy was significantly more sensitive (P < 0.001) to venetoclax than non-hyperdiploidy (Fig. 2b). High sensitivities to dasatinib were seen for T cell ALL (LC50 0.20, P < 0.001 compared to non-T cell ALL), TCF3–PBX1 (LC50 0.28, P = 0.013 compared to non-TCF3–PBX1; Fig. 2b) and BCR–ABL1 (LC50 0, P < 0.001 compared to non-BCR–ABL1), consistent with previous findings by us and others35,36. BCR–ABL1 was also sensitive to ibrutinib (LC50 0.09, P = 0.002 compared to non-BCR–ABL1; Extended Data Fig. 3a). NUTM1 had low LC50 for several antimetabolites: LC50 0.09 for cytarabine (P = 0.007 compared to non-NUTM1 subtypes), LC50 0.16 for mercaptopurine (P = 0.002; Fig. 2b) and LC50 0 for thioguanine (P = 0.0003).
Although characterized by nearly identical global gene expression profiles, ETV6–RUNX1-like and ETV6–RUNX1 ALL showed notable differences in drug sensitivity: ETV6–RUNX1-like ALL was more resistant to L-asparaginase (LC50 0.55 compared to 0 in ETV6–RUNX1, P = 0.013; Fig. 2b) but had increased sensitivity to trametinib (LC50 0.20 versus 1.0, P = 0.002; Extended Data Fig. 3b). BCR–ABL1, BCR–ABL1-like and CRLF2 are also related to each other based on global transcriptional profile and somatic genomic features, yet these 3 subtypes had varying sensitivities to dasatinib (LC50 0 versus 1.0 versus 0.89, respectively, P < 0.001), mercaptopurine (LC50 0.54 versus 0.97 versus 0.48, P = 0.016) and prednisolone (LC50 0.54 versus 1.0 versus 0.27, P = 0.047) but with similar response to other cytotoxic and targeted drugs (Extended Data Fig. 4). Comparing with other T cell ALL (Extended Data Fig. 3c–h), early T cell precursor (ETP) ALL was more resistant to antimetabolite and non-antimetabolite cytotoxic drugs: vincristine (LC50 0.81 versus 0.33, P < 0.001), prednisolone (LC50 1.0 versus 0.36, P = 0.009; Fig. 2b), thiopurines (mercaptopurine LC50 0.50 versus 0.29; P = 0.017 and thioguanine LC50 0.42 versus 0.12; P = 0.006), cytarabine (LC50 0.76 versus 0.45, P = 0.023) and daunorubicin (LC50 0.65 versus 0.50, P = 0.001). ETP showed a trend for higher venetoclax sensitivity compared to T cell ALL, although this was not statistically significant in this relatively small cohort (LC50 0.46 versus 0.81). Taken together, our data point to new subtype-specific therapeutic opportunities in ALL. To explore the effects of multiple testing, we also performed Benjamini–Hochberg correction for these subtype-related comparisons (Supplementary Table 5) and the majority of the associations remained significant.
Association of ALL ex vivo drug sensitivity with MRD
We evaluated the association of drug LC50 with in vivo response, as reflected by MRD during remission induction therapy. For each patient, MRD was measured at day 15 and day 42. In this study, we evaluated the coefficients of linear regression (β) for each drug, where each unit change in coefficient represents a unit change of MRD (log10-transformed),that is, a coefficient of 0.30 (or log10(2)) indicates that the MRD is twice as high in resistant patients compared to sensitive patients; a coefficient of −0.30 (or log10(0.5)) indicates that the MRD is half as high. A coefficient of zero indicates that there is no difference in MRD between resistant and sensitive patients…
