Endothelial cell–leukemia interactions remodel drug responses, uncovering T-ALL vulnerabilities

T-cell acute lymphoblastic leukemia/lymphoma (T-ALL/LBL)¹ is an aggressive hematological malignancy with high treatment failure and poor overall survival (OS).^2-4 Developments in therapy and stem cell transplants have improved the event-free survival rates of young patients (∼80%). However, refractory/relapsed pediatric (∼20%) and adult patients with T-ALL (∼40%) often die of the disease.^5,6 

The original stratification of T-ALL (European Group for the Immunological Classification of Leukemia [EGIL] scoring system) highlights a class of early T-cell precursor acute lymphoblastic leukemia: ETP-ALL, representing ∼5% to 15% of T-ALL. Patients with ETP-ALL have significantly worse survival,^7,8 a course improved via intensive asparaginase-based regimens.⁵ Within T-ALL, the deregulated expression of multiple transcription factors and CDKN2A/2B cell-cycle regulators, and hyperactive NOTCH1 signaling, were initially described as key genomic aberrations. Furthermore, defects affecting Janus kinase/signal transducer and activator of transcription (JAK/STAT) signaling, protein translation, and epigenetic control provide novel and attractive targets for therapy.^9-13 Of note, the implementation of transcriptional signatures is providing actionable diagnostic criteria and improved clinical management.¹⁴ 

The cancer microenvironment plays a critical role^15-17 in modulating cell growth and therapeutic responses.^16-19 Within the T-ALL niche, collective protumorigenic signals arise from multiple cellular and noncellular components. Notably, tumor-associated endothelial cells (TECs) are crucial in the pathogenesis of solid and liquid malignancies,^20-22 ultimately contributing to multiple protumorigenic phenotypes. In T-ALL, TECs modulate leukemic extravasation²³ and tumor aggressiveness via multiple and synergistic mechanisms, including SDF1α/CXCR4, DLL4-JAG1-2/NOTCH, and IGFBP7/IGF1 prosurvival pathways.^18,24,25 However, it is unclear how these individual compartments contribute to the education of leukemic and microenvironment cells. Importantly, the microenvironment can affect drug responses^15,26 via dysregulated chemokine signals,²⁷ enhanced production of oxidative radicals and exosomes,²⁸ and hypoxia.²⁹ As a result, therapies targeting microenvironment cells are predicted to improve clinical responses.^30,31 However, the influence of endothelial-based cultures on perturbing drug responses has not been fully explored yet.

By taking advantage of endothelial cells (ECs) transduced with E4ORF1 (E4-ECs),^32,33 we developed an endothelial-leukemia platform to explore the subverted role of tumor vascular niches. This platform integrates T-ALL patient–derived tumor xenografts (PDXs) and ECs with drug-screening libraries, ultimately translating findings to in vivo preclinical trials. This approach led to the generation of proof-of-concept transcriptional signatures that are predictive of drug response, and the selection of effective compounds/combinations in T-ALL. More importantly, we show that T-ALL and ECs bidirectionally modulate their transcriptomes and phenotypes, and ECs provide signals counteracting the efficacy of selected drugs.

We established an EC-based platform to unveil new microenvironment-leukemia interactions and cancer vulnerabilities. We anticipate that this approach could foster the design of precision medicine–based approaches to improve the treatment of patients with T-ALL.

Pathological samples from bone marrow or peripheral blood (blasts, >80%) were collected at the Hematology Center of Sapienza University of Rome or Weill Cornell Medicine of New York. Diagnoses were assigned based on the World Health Organization classification. According to the Declaration of Helsinki, deidentified patient samples were obtained with informed consent. The institutional review boards of Sapienza University of Rome and Weill Cornell Medicine approved human studies.

PDX implantation and propagation were previously reported.³⁴ Briefly, fresh or cryopreserved patient samples were implanted into NOD.Cg-B2m^tm1Unc Prkdc^scid Il2rg^tm1Wjl/SzJ (NSG B2m) mice. Animals were euthanized when lymphoblastic cells represented at least 50% of the total circulating blood cells. Serial passages were achieved by IV injection into NOD.Cg-Prkdc^scidIl2rg^tm1Wjl/SzJ (NSG) mice. Animal studies were approved by the Weill Cornell Medicine Animal Care and Use Committee (2014-0024).

Flow cytometry was performed as described previously,³⁵ using dedicated instrumentations (BD LSR II, BD FACSCanto flow cytometers and BD FACSCelesta with plate reader for high-throughput screenings) and high-speed sorting (BD Aria).

PDX-derived T-ALL cells were cultured in RPMI with 10% fetal calf serum (FCS) or StemSpan (SS; StemCell Technologies) supplemented with 10% knockout serum replacement (KSR; Life Technologies) and a cocktail of antibiotics and interleukins (see supplemental Data, available on the Blood website). Human umbilical vein endothelial cells transduced with a lentiviral gene transfer vector expressing the E4-ORF1 gene (VeraVec ECs; Angiocrine Bioscience), namely E4-ECs. Monolayers of E4-ECs were cultured in ex vivo medium supplemented with 10% FCS (Lonza) and maintained at 37°C in a humidified 5% CO₂ atmosphere. Compounds for the drug screenings were purchased from Selleckchem and tested for 72 hours (at least in duplicate) at a concentration of 1.0 μM or in dose-response dilutions where specified. The 433-drug screenings were executed in 384-well plates under either RPMI (plus 10% FCS) or SS (plus 10% KSR) conditions, and the viability was assessed using the CellTiter-Glo assay (Promega). The EC/T-ALL coculture screenings were performed in 96-well plates, and data were collected using the high-throughput screening flow cytometer. The gating strategy was based on appropriately compensated settings for multicolor flow cytometry (propidium iodide for cell death, CFSE for EC labeling, and cell tracer violet for T-ALL cells). To provide uniform measures of drug effects and to estimate EC rescue, we computed specific cell death by normalizing viability on drug exposure with the viability of matched DMSO controls [formula: (death under drug x) – (death of control)/(100 – death of control) × 100]. Drugs were defined as active when able to kill more than 50% of leukemic cells compared with controls.

Refer to the supplemental Methods for details.

Mice underwent IV injection (1 × 10⁶ T-ALL cells), and engraftment was assessed by flow cytometric analysis of peripheral blood cells. When human T cells reached 5% of the total, mice were randomized into untreated (vehicle) and treated arms. Numbers, age, and sex of the mice were equally distributed among arms. Compounds were administered as described in the figure legends and in the supplemental Methods. Tumor burden was monitored by flow cytometric analysis of peripheral blood or spleen size, using total body magnetic resonance imaging (MRI) or measurements postautopsy. Body weight was used as a surrogate for drug toxicity.

The appropriate statistical methods were applied as indicated in the related figure legends. Continuous variables were analyzed on the basis of normal distribution and homogeneity of variance. The Kaplan-Meier method and log-rank test were used to assess survival.

Tests were performed with Prism version 9.0 (GraphPad, San Diego, CA), MATLAB version 9.0 (MathWorks, Natick, MA), or R version 4.0 (released in 2020; R Foundation for Statistical Computing).

We first characterized the transcriptomic profile of 39 T-ALL patient samples (Figure 1A), from which we established a library of PDX models (n = 22) (supplemental Table 1; supplemental Figure 1A-C). These closely recapitulated the transcriptomic profile and shared the identical T-cell receptor with those of matched patients (Figure 1B-C; supplemental Figure 1D; supplemental Table 1).

We then implemented a multistep approach, including (1) optimizing T-ALL PDX cell culture conditions ex vivo, (2) screening a broad library of drugs, (3) testing drug candidates in an EC-based coculture system, and (4) confirming key candidates in vivo using PDX models.

First, we tested various culture conditions and media, demonstrating improved viability and increased cell number in supplemented SS medium relative to supplemented RPMI medium (supplemental Figure 1E).

Second, we challenged T-ALL PDX cells ex vivo with a drug library of 433 compounds (1 μM), targeting ∼634 proteins (Figure 1D,F; supplemental Table 2). Prior to choosing the 1 μM final concentration, we verified that the 50% inhibitory concentration (IC₅₀) would be reached in most of the compounds (400 of 433; 92%). In detail, 13 samples spanning different passages of 6 T-ALL PDX models were tested. We observed high reproducibility among biological and technical replicates (supplemental Figure 1F-I). SS-cultured PDX cells showed that most compounds (∼89%) had little to no effect on T-ALL viability. Meanwhile, only 11% of drugs reduced average cell viabilities across the PDX models to less than 50%, corresponding to a cluster of 39 drugs (Figure 1F). To further elaborate on the drug concentration (1 μM), we tested serial dilutions of 10 of 39 broadly active compounds in 4 additional PDX models, identifying unique IC₅₀ values in individual models (supplemental Figure 1P). In all cases, 1 μM was compatible with the efficacy range.

Using unsupervised hierarchical clustering and principal component analysis (PCA), we identified 2 clusters of T-ALL samples based on differential drug susceptibility (Figure 1G). Specimens belonging to earlier passages populated cluster 1 (C1), whereas later passages were more often skewed toward cluster 2 (C2), suggesting decreased therapeutic efficacy along serial passages (supplemental Figure 1I). We confirmed this reduced viability along serial passages using the 3053 T-ALL PDX (supplemental Figure 1J), in which the enrichment of the SMCHD1-JAK2 fusion overtime was observed (supplemental Figure 1K-M). To describe the response heterogeneity among the different T-ALL models, we combined the viability results of the 3119/3053 and R05/RO6 T-ALL PDXs based on the highest response similarity scores (supplemental Figure 1N) identifying differentially active drugs between the 2 groups (Figure 1H; supplemental Figure 1O). Last, we associated gene expression profiles with drug responses, to identify putative drug-related biomarkers. Despite the limited number of samples, we discovered several gene signatures linked with drug responses (Figure 1I; supplemental Figure 1Q-R).

Within the tumor microenvironment, endothelial cells (ECs) by supplying angiocrine factors sustain hematopoietic and neoplastic cells^36,37 and enhance their survival under toxic stress.^{17,18,25,38,39} We previously demonstrated that primary nontransformed E4ORF1-transduced ECs³² (herein referred to as E4-ECs) are an instructive serum-free/xenobiotic-free model to study tumor-host interactions.²⁵ Here, we first showed that E4-ECs sustained T-ALL cells under stress conditions such as starvation, lymphokine depletion, and oxygen deprivation (supplemental Figure 2A-E). We then developed a high-throughput flow cytometry–based endothelial-leukemia coculture system. Violet tracer–labeled T-ALL cells from 10 PDX models were cultured in SS medium supplemented with 10% KSR and challenged with the most effective drugs (n = 39) from the initial screening (n = 433) in the presence of CFSE-labeled E4-ECs (supplemental Figure 2F).

Sensitivity data showed a high concordance among T-ALL cells isolated from separate compartments (spleen and bone marrow) from the same PDX mouse (supplemental Figure 2G), from sibling mice implanted with the same tumor seeds (supplemental Figure 2H), and from mice belonging to 2 consecutive passages (supplemental Figure 2I). In addition, the percentage of viable E4-EC cells cultured alone and/or with T-ALL cells was similar across all compounds (supplemental Table 5). These findings suggested that T-ALL responses likely were not linked to suboptimal drug concentrations (supplemental Figure 2J).

Overall, E4-ECs significantly improved the T-ALL survival of individual PDX models against selected compounds after 72 hours of ex vivo treatment (Figure 2A-D; supplemental Figure 2K). A compound was defined as “rescued” when there was a decrease of at least 20% in specific cell death compared with the no-EC condition. By PCA, the vectorized EC rescue showed converging directions in all models, suggesting conserved rescue mechanisms (Figure 2A-B; supplemental Figure 2K). Of note, some compounds were reproducibly rescued across multiple PDX models (including oligomycin, JQ1, NSC319726, SN38, belinostat, CEP-18870, KPT-185, SC144, TSA, and YM155). Others were poorly effective in specific models, underlying potential “private” rescue mechanisms (Figure 2C-D; supplemental Figure 2L-M). Further, we tested 12 additional drugs on 7 of 10 PDX models screened, including conventional chemotherapeutics for T-ALL and novel compounds in clinical trials (supplemental Figure 2N). We identified commonly rescued drugs: asparaginase, dexamethasone, vincristine, ponatinib, dasatinib, and YKL-5-124 (CDK7 inhibitor) in 4 models. Overall, rescued compounds were active on molecules belonging to interconnected pathways, such as PI3K/AKT, p53, FoxO, JAK/STAT, NOD-like receptor, and HIF-1 (supplemental Figure 2M). Next, we computed a modified Shannon entropy for each PDX model (see “Methods”) (Figure 2E), showing an overall tendency toward increased response heterogeneity and plasticity in the presence of E4-ECs. We then focused on the different response profiles of the 10 PDX models challenged in the presence of E4-ECs. Among the most effective drugs across all models, some are of particular interest because of their potential clinical applicability: proteasome inhibitors (bortezomib, CEP18870, and MLN2238; average specific cell death [SCD]: 83%, 79%, and 78%, respectively), histone deacetylase (HDAC) inhibitors (panobinostat, belinostat, and CHIR124; average SCD: 75%, 69%, and 51%, respectively), heat shock protein (HSP)-90 inhibitor (ganetespib; average SCD, 49%), and BCL2 inhibitors (venetoclax and navitoclax; average SCD: 27% and 66%, respectively). Notably, other drugs displayed unique efficacy in selected models and against different T-ALL phenotypes, with ETP-ALL samples showing increased resistance to several compounds (Figure 2F; supplemental Figure 2O-R): YM155 (survivin inhibitor; Δ SCD [ETP – non-ETP] = 39.9%; P < .0001), SNS032 (CDK2/9 inhibitor; Δ SCD = 34.7%; P = .01), ouabain (ATP1A1 inhibitor; Δ SCD = 31%; P = .02), STF118804 (nicotinamide phosphoribosyltransferase [NAMPT] inhibitor; Δ SCD = 20.9%), and KPT-185 (exportin 1 [XPO1] inhibitor; Δ SCD = 17.2%; P < .0001).

Finally, we used our dual T-ALL-EC platform to test targets based on the genetic background of individual T-ALL PDXs. We treated the 3119 PDX cells (carrying a NUP14/ABL translocation) with dasatinib, and RO2 PDX cells (mutant JAK1-JAK3-STAT5B) and RO6 PDX cells (mutant STAT5) (supplemental Table 1) with Jak inhibitors (ruxolitinib and tofacitinib, alone or in combination), in the presence or absence of E4-ECs (supplemental Figure 2S). E4-ECs partially rescued T-ALL cells from single-agent treatments. However, in combination, ruxolitinib and tofacitinib were significantly more effective in both RO2 and RO6 PDXs.

These results highlighted the utility of the T-ALL-EC platform to assess (1) drugs that are effectively rescued by ECs, (2) pathways collectively or privately involved in EC rescue, (3) differential T-ALL response profiles based on intrinsic tumor features, and (4) unique targetable T-ALL vulnerabilities.

Aiming to elucidate the relationships between ECs and T-ALL cells, we investigated the transcriptomic profiles of cocultured E4-ECs and T-ALL cells from 3119 and RO2 PDX models, using single-cell RNA sequencing (scRNA-Seq). After quality control filtering (supplemental Figure 3A), we clustered all cells on the basis of the most variable genes (Figure 3A; supplemental Figure 3B-D), and analyzed the 3119 and RO2 T-ALL PDX models separately, given their divergent genetic profiles (supplemental Table 1). Notably, proliferating T-ALL cells were most frequently included in clusters enriched in EC-cocultured T-ALL cells in both models (Figure 3B; supplemental Figure 3E). We found that T-ALL cocultured with E4-ECs underwent profound transcriptional changes (Figure 3C; supplemental Table 6). Furthermore, the differentially expressed genes defined an “EC-mediated T-ALL education signature” for each model, with significant overlap between the 2 models and a high concordance with bulk RNA-Seq data from cells cultured under the same conditions (Figure 3D; supplemental Figure 3F-G). In addition, we detected enhanced single-cell entropy in T-ALL cocultured with E4-ECs (supplemental Figure 3H), accounting for increased plasticity.^40,41 

Then, we compared differential pathway activation between the 2 conditions. Both EC-cocultured RO2 and 3119 T-ALL cells upregulated JAK/STAT, mitogen-activated protein kinase (MAPK), transforming growth factor-β (TGFB), and epidermal growth factor receptor (EGFR) pathways, with downregulation of the p53 signaling pathway (Figure 3E). Gene set enrichment analysis (GSEA) showed enrichment of the c-Myc pathway in cocultured elements (supplemental Figure 3I), along with increased expression of genes belonging to the NOTCH pathway (NOTCH3, HES1, HES4)^42,43 (Figure 3F). Notably, insulin-like growth factor 1 receptor (IGF1R) was also commonly upregulated in cocultured T-ALL cells in both models (Figure 3F; supplemental Figure 3J). Other commonly upregulated genes included ribosomal genes RPS6 (highly expressed in several solid cancers⁴⁴ and in immature leukemic blasts),^45,RPL11 (whose activating mutations have been described in T-ALL),⁴⁶ and HMGB2, a known poor prognostic factor in several cancers.^47,48 In both models, MALAT1 (known to control the crosstalk between ECs and tumor cells⁴⁹) was identified as one of the most downregulated genes in E4-EC-cocultured T-ALL. Similarly, RHOH⁵⁰ (Figure 3C,F), normally expressed in hematopoietic cells where it functions as a negative regulator of cell growth and survival, was also downregulated. Importantly, these differences were maintained at single-cell resolution, allowing us to define model-specific clusters with enrichment of educated cells (RO2, cluster 5; 3119, clusters 0 and 4) (Figure 3F). Finally, to investigate whether E4-ECs could contribute to preserving a more immature/stemlike phenotype, we evaluated the differential expression of genes linked to T-cell commitment and maturation.^51,52 In both models, EC-cocultured T-ALL cells expressed higher levels of genes related to early double-negative (DN) or double-positive proliferating (DP-P) thymocytes. In contrast, T-ALL cells cultured alone were shifted more toward DP-quiescent (Q) or mature T-cell phenotypes (Figure 3G).

We then focused on the transcriptomes of E4-ECs cultured alone or with T-ALL cells of the 3119 and RO2 PDX models. Overall, we distinguished different clusters corresponding to T-ALL-cocultured ECs (educated E4-ECs) or E4-ECs cultured alone (Figure 4A; supplemental Figure 4A). These populations were transcriptionally distinct, leading to the construction of T-ALL-mediated “EC education signatures” (Figure 4A; supplemental Table 7). Globally, cocultured E4-ECs upregulated interferon-γ and -α response hallmark pathways (supplemental Figure 4B) and increased the activity of multiple pathways, including JAK/STAT, nuclear factor-κB (NF-κB), tumor necrosis factor-α (TNF-α), and vascular endothelial growth factor A (VEGF-A) (Figure 4B), with a concurrent increase in entropy reflecting in higher differentiation potency and plasticity (supplemental Figure 4C).^40,41 At the gene level, cocultured E4-ECs displayed enrichment in Notch signaling, angiogenesis (DLL1,^53,JAG1,⁵⁴ and CD34), cell migration, and survival (ETS1, ETS2, and IGFBP4) genes (Figure 4C). Conversely, IGFBP7, a known IGF1R decoy with tumor suppression features in different cancer types;⁵⁵ CD63/TIMP-1 (involved in endothelial migration); CD9; and S100A10 (regulating cell-cell interaction) were downregulated in cocultured E4-ECs (Figure 4C).

We then performed a ligand-receptor analysis in the 3119 and RO2 PDX models to investigate the bidirectional crosstalk between T-ALL and E4-ECs (Figure 4D). This approach led to the identification of heterogeneous interactions, mainly coshared by the 2 models (Figure 4D; supplemental Figure 4D). Among them, the IGF1-IGF1R axis was significantly enriched (outlined in red in Figure 4D). Because cocultured T-ALL upregulated IGF1R, and conversely E4-ECs downregulated IGFBP7 (see above), we attempted to functionally verify whether the selective abrogation of the IGF1 pathway could impact the EC-mediated drug rescue. We tested the 39 pan-effective drugs (Figure 2C) with or without the recombinant IGFBP7 (500 ng/mL) in the 3119 and 3053 PDXs. Of note, IGFBP7 abrogated the E4-EC-mediated rescue of enzastaurin (PKC-β inhibitor) and SC144 (GP130 inhibitor) in T-ALL PDX 3119 and 3053 models, and CHIR124 (Chk1 inhibitor) and YM155 (survivin inhibitor) in T-ALL PDX 3119 (supplemental Figure 4E). As these drugs regulate the activity of IGF1R or its downstream effectors^56-58 (supplemental Figure 4F), we postulate that the EC-mediated rescue is in part mediated by the engagement of IGF1R of T-ALL cells.

To further explore the changes in T-ALL-cocultured E4-ECs, we correlated their transcriptomics profiles with tumor ECs (TECs—directly sorted from tumor masses) and naive normal/E4-ECs (supplemental Table 8). These analyses demonstrated that in vitro educated E4-ECs and TECs share a set of differentially expressed genes, relative to normal E4-ECs (Figure 4E-F; supplemental Figure 4G-H). These genes have been implicated in orchestrating invasion and metastasis (ie, GSK3B, BRAF, STAT3, MET), genome instability (ie, PTEN, BRAF, KEF1, ATM), angiogenesis (ie, FOXO3, CREBBP, HIF1A), and enabling immortality (ie, CDK6, SPEN, TNKS).

Last, we looked for genes coshared by both RO2 and 3119 PDX models whose expression contributed to the education signatures in both compartments (EC and T-ALL) (Figure 4G; supplemental Figure 4I; supplemental Table 9). As expected, genes linked to the EC and T-ALL compartments were distinct and displayed different functional modules (Figure 4G). Therefore, the T-ALL-EC model is an informative and useful tool to probe EC-tumor interactions.

To validate the data generated using the in vitro coculture platform, we performed scRNA-Seq on T-ALL cells from the 3119 PDX spleen (ex vivo splenic T-ALL cells). We identified 4 clusters representative of the intratumor heterogeneity based on the differential expression of genes related to: cell cycle, MYC, PI3K/AKT signaling, and epithelial-mesenchymal transition (supplemental Figure 5A-C). Then, we enriched our analysis by including cells from the same source cultured for 5 days ex vivo in SS medium either with E4-ECs (in vitro EC-educated T-ALL) or without E4-ECs (T-ALL alone) (Figure 5A). Overall, 7 T-ALL cell clusters were identified (Figure 5B-C). Clusters 0 and 4 were enriched with T-ALL elements cultured without ECs; clusters 1, 5, and 6 with T-ALL cultured with ECs; and clusters 2 and 3 with T-ALL cultured from spleen. Of note, clusters 2, 3, 5, and 6 showed a great abundance of E4-EC-educated T-ALL or spleen T-ALL cells (Figure 5D), pointing to the closer similarity of these 2 compartments compared with cultured alone T-ALL cells. Notably, clusters 2 and 3 were highly enriched for proliferating cells (supplemental Figure 5D). Then, we applied the T-ALL education signature defined in the previous in vitro experiment within each cluster (Figure 5E). Remarkably, in vivo splenic T-ALL were globally more similar to in vitro EC-educated T-ALL than T-ALL alone. This was more evident in clusters 0 and 4, containing enough cells belonging to all different entities (Figure 5F-G; supplemental Figure 5E). Ultimately, splenic T-ALL cells recapitulated the different expressions of genes contributing to the education signatures of both 3119 and RO2 PDX models (Figure 5F-G; supplemental Figure 5E). These data demonstrated the reliability of the EC/T-ALL coculture platform in recapitulating in vivo conditions.

To validate our stepwise drug screening platform in vivo, we selected 5 compounds shown to be effective on T-ALL cells in the presence of E4-ECs (Figure 2A-C; supplemental Figure 2O-P), effective in preclinical settings, and/or effective in targeting underlying PDX T-ALL vulnerabilities. Panobinostat (HDAC inhibitor), ruxolitinib (JAK1/2 inhibitor), tofacitinib (JAK1/2/3, TYK2 inhibitor), bortezomib (20S proteasome inhibitor), and irinotecan (topoisomerase I inhibitor) were tested in 11 PDX models, accounting for an overall number of 111 mice (Figure 6A-B). Overall, treated mice had a better prognosis than mice receiving vehicles (Figure 6B). First, R02 and R06 T-ALL PDX mice were challenged with ruxolitinib alone or in combination with bortezomib (RO2) or tofacitinib (RO6) (supplemental Figure 6A). Treatments were well tolerated, with no signs of toxicity (data not shown). In comparing treated and control mice, spleen size and percentage of leukemic circulating cells were significantly reduced (Figure 6C-D; supplemental Figure 6B-C). Among the 3 single agents tested, ruxolitinib was least effective in reducing circulating disease in both models, whereas both combinations allowed for better disease control, more pronounced reduction in spleen volumes, and improved overall survival (Figure 6C-E; supplemental Figure 6B-E). Based on the ex vivo coculture drug screening, we identified SN38 (the active metabolite of irinotecan) as active in all models, despite an insufficient partial rescue by E4-ECs (death rate of T-ALL cells, >50% compared with controls in all cases [mean SCD, 68%; range, 54%-83%]; Figures 1F and 2C). Accordingly, we randomized the 3053 and 3119 T-ALL PDX mice to receive either vehicle or irinotecan (20 mg/kg per day for 2 weeks) (supplemental Figure 6F). Total body MRI documented a normalized spleen size by day 15 in treated mice (Figure 6F). We also detected a reduced blood leukemia burden in treated mice up to 40 days and improved overall survival (Figure 6G-H) (median OS, 103 vs 52 days; P = .0002). No signs of overt toxicity were documented (supplemental Figure 6G).

Epigenetic targeting using the HDAC inhibitor panobinostat has been suggested as a potential therapy against T-ALL,⁵⁹ and this strategy was supported by our in vitro drug screening (Figures 1F and 2C). Thus, we executed a single-drug preclinical trial with panobinostat, in 8 different T-ALL PDX models (Figure 6I; supplemental Figure 6H). This treatment significantly improved OS (median OS, 117 vs 68; P = .03) (Figure 6I) and decreased circulating T-ALL blasts (Figure 6J). Last, based on the finding that the IGF1-IGF1R axis is upregulated in EC-cocultured T-ALL cells (Figures 3F and 4D; supplemental Figure 3J) and that its inhibition downregulated EC-mediated rescue (supplemental Figure 4E), we evaluated a selective IGF1R inhibitor (linsitinib) in vivo in 4 different PDX models (Figure 6K-L; supplemental Figure 6I). As a debulking agent, we used daunorubicin, an identified pan-active compound in our drug screening in vitro (Figures 1F and 2C). The combination of daunorubicin plus linsitinib controlled disease more effectively and improved OS (median, 84 days) compared with single agents (median: daunorubicin, 68 days; linsitinib, 63 days) or controls (median, 42 days) (P < .0001) (Figure 6K-L), suggesting that the IGF1-IGF1R inhibition counteracted leukemia progression in combination with standard chemotherapy.

These data demonstrated the predictive utility of our screening platform as a drug discovery process to identify active compounds and T-ALL dependencies.

Our study presents a novel customized vascular niche/cancer platform to define targetable vulnerabilities and identify effective drugs against T-ALL cells, potentially applicable to other cancers. Although the treatment of patients with T-ALL has considerably improved, relapsed T-ALL often remains incurable.^5,6 Preclinical studies have provided encouraging approaches;^60-62 however, therapeutic response does not always depend on pathogenetic mutations,^10,63 and preclinical trials are costly and require specialized units. We envisioned a multistep approach that integrates readouts from large drug screenings on PDX cells ex vivo, followed by “ad hoc” in vivo cross-validation (Figure 6A). We first identified a group of broadly active compounds across all models tested and observed that increased drug resistance might arise along serial PDX propagations. This should be further exploited to improve our understanding of therapeutic failures. Of interest, we could find a correlation between responses and a priori transcriptional profiles for selected agents, a strategy that can provide new avenues to discover biomarkers predictive of responses.

It is well known that the tumor microenvironment can modulate drug responses in acute leukemia, a scenario epitomized by the contribution of mesenchymal and endothelial cells.^15-19 Specifically, ECs have been seldom used in broad drug discovery programs.⁶⁴ Nevertheless, their role in the maintenance/regulation of HSCs^36,37 is established, fostering ALL survival via a multitude of signaling pathways (ie, Notch,⁶⁵ SDF-1,⁶⁶ and CXCL12⁶⁷). Accordingly, we generated a coculture system using E4ORF1-transduced ECs, which have been proven to sustain T-ALL cells in vitro and can be propagated for at least 5 to 10 passages in vitro. This system demonstrated that the T-ALL drug susceptibility was often downmodulated by ECs. Single-cell analysis revealed that a multitude of ligand-receptor interactions and downstream pathways are involved in this process (eg, IGF1/IGF1R, NOTCH-DLL, and JAK/STAT), suggesting that the EC-mediated prosurvival effect may require a multitargeted approach to be fully counteracted. Toward this end, adding recombinant IGFBP7 in vitro partially abrogated EC rescue for selected compounds, suggesting reliance on the IGF1-IGF1R pathway in specific cases.

On the other hand, several compounds were effective in the presence of ECs across multiple models. These included drugs with clinical interest (proteasome, HDAC, HSP-90, and BCL2 inhibitors). We also observed a degree of heterogeneity which may be partially linked to unique T-ALL subgroups, with ETP-ALL being overall more resistant, specifically on exposure to drugs targeting PI3K/AKT and JAK/STAT pathways. A selection of 5 compounds was ultimately shown to be effective in vivo either alone or in combination. This was exemplified by irinotecan, a drug currently approved for treating many solid tumors, suggesting its potential repositioning in the context of T-ALL, and panobinostat, an HDAC inhibitor whose effect in T-ALL has not been fully unraveled. Similarly, the efficacy of JAK/STAT inhibitors has been documented in models carrying activating mutations.^10,13,68 Moreover, the data generated in vitro with recombinant IGFBP7 and the improved survival of daunorubicin-treated T-ALL PDX in combination with IGF1R inhibitor linsinitib (although limited on 4 models) further support the role of IGF1-IGF1R signaling in the leukemia-EC crosstalk.

Mechanistically, we show that ECs and T-ALL cells undergo bilateral transcriptional changes modulating critical signaling pathways at single-cell resolution. On the T-ALL side, these genes were mostly linked to stemness, T-cell commitment, and maturation, emphasizing the heterogeneity and plasticity of leukemic elements. Remarkably, T-ALL cocultured with E4-ECs were more similar to ex vivo splenic T-ALL cells freshly harvested from PDX mice compared with T-ALL cells cultured alone, proving the reliability of our coculture system. On the EC side, a significant set of differentially expressed genes was expressed in common with in vivo TECs. These bidirectional changes are reflected in increased transcriptomic and pharmacological entropy, further pointing toward the dynamicity of this dual-culture system compared with conventional static single cell suspension cultures.

Hence, the EC/T-ALL coculture system partially recapitulates in vivo conditions, representing a robust platform to study leukemia-host interactions and drug sensitivities. We envision that this integrated stepwise approach might overcome the liabilities of current drug screenings and propel the entry of new agents and combinations in clinical trials.

The authors thank J. Casano and A. Miyaguchi for technical support, and A. Arkur and S. B. Shah for administrative assistance. The authors are grateful to the Weill Cornell Medical College Epigenomics Core Facility for next-generation sequencing (NGS) library preparation and sequencing for this study. The authors also thank the members of the Weill Cornell Flow Cytometry Core Facility and Sandra and Edward Meyer Cancer Center PDX shared resource.

This work was supported in part by the Italian Association for Cancer Research (AIRC, Milan, Italy), Metastases 5×1000 Special Program, and grant 21198 (R.F.); National Institutes of Health (NIH), National Cancer Institute (NCI) grants CA229086, CA229100, Leukemia and Lymphoma Society grant LLS 7011-16, departmental funds, and the Sandra and Edward Meyer Cancer Fund (G.I.); and the Rita-Levi Montalcini grant from the Italian Ministry of University and Research (MIUR) (D.F.). J.M.P. is supported by NIH NCI grant R01CA18710903S1. L.V.C. is supported by an American Italian Cancer Foundation (AICF) postdoctoral research fellowship. S.R. is funded by the Hartman Institute for Therapeutic Organ Regeneration; the Ansary Stem Cell Institute; NIH, National Heart, Lung, and Blood Institute grant R35 HL150809, NIH, National Institute of Diabetes and Digestive and Kidney Diseases grant RC2DK114777 and NIH, National Institute of Allergy and Infectious Diseases grant U01AI138329; the Empire State Stem Cell Board (C030160); the Daedalus Fund for Innovation; the Selma and Lawrence Ruben Science to Industry Bridge Fund from Weill Cornell Medicine; and the Starr Foundation stem cell core project and initiatives (TRI-SCI 2019-029).

Contribution: L.V.C., D.F., J.M.P., L.Y., F.D.G., L. Cerchietti, S.R., and W.T. were responsible for conceptualization, formal analysis, methodology, writing-original draft, writing-review and editing; Y.H. and C.K. were responsible for formal analysis, methodology, and editing; W.C. was responsible for formal analysis and methodology; M.G., L. Consolino, and N.Z. were responsible for conceptualization, formal analysis, methodology, and editing; J.G.B.D. was responsible for conceptualization, methodology, and editing; F.B. was responsible for conceptualization, methodology, and editing; A.M.M., S.C., A.G., and R.F. were responsible for conceptualization, writing-review and editing; O.E. was responsible for conceptualization, formal analysis, writing-review and editing; and G.I. was responsible for conceptualization, data curation, formal analysis, supervision, methodology, writing-original draft, project administration, writing-review and editing.

Conflict-of-interest disclosure: M.G. is currently employed at Angiocrine Bioscience (San Diego, CA). S.R. is a cofounder and nonpaid consultant to Angiocrine Bioscience (San Diego, CA). The remaining authors declare no competing financial interests.

Correspondence: Giorgio Inghirami, Department of Pathology and Laboratory Medicine, Weill Cornell Medicine, 1300 York Ave, New York, NY 10065; e-mail: ggi9001@med.cornell.edu; Shahin Rafii, Hartman Institute for Therapeutic Organ Regeneration, Ansary Stem Cell Institute, Division of Regenerative Medicine, Department of Medicine, Weill Cornell Medicine, 1300 York Ave, New York, NY 10065; e-mail: srafii@med.cornell.edu; and Leandro Cerchietti, Division of Hematology and Medical Oncology, Weill Cornell Medicine, 1300 York Ave, New York, NY 10065; e-mail: lec2010@med.cornell.edu.