To add a functional component to our pediatric precision oncology program, INFORM, we initiated a personalized drug sensitivity and resistance profiling platform based on metabolic activity measurements complementing next-generation sequencing (NGS)-based target identification. The platform consists of three principal steps (Fig. 1): (a) collection and shipment of viable tumor tissue, (b) dissociation of tumor tissue, followed by generation and drug treatment of patient-derived ex vivo fresh tissue spheroid cultures (FTCs) in 384-well drug plates, and (c) data collection and determination of individual drug sensitivity and resistance profiles to be reported to the INFORM molecular tumor board. In parallel, long-term cultures (LTCs) and PDX models are being established for expanded drug testing approaches and clinical trial development.

a Sample collection and shipment. b Generation of patient-derived ex vivo fresh tissue spheroid cultures (FTCs) and treatment with a drug library. Readout: metabolic activity. c Data collection, analysis and preparation of drug sensitivity reports.

Collection and shipment of viable tumor tissue (a)

As a multicenter, multinational real-world precision oncology program, INFORM receives tumor material from 13 countries and over 100 pediatric oncology centers, many of which are nonuniversity centers with limited laboratory support. Thus, to keep the shipment conditions for viable tumor tissue as simple as possible, we allowed the shipment of tissue at room temperature or cooled overnight either in physiological sodium chloride solution or in any serum-free cell culture medium (e.g., RPMI, Neurobasal or other) (see “shipment protocol” Supplementary Note 1). We did not observe significant differences in the viability of incoming tissue shipped in 0.9% NaCl solution versus cell culture medium, although only samples shipped in NaCl (6/92) showed viability values <60% (Fig. 2a). Furthermore, screening success was higher if the sample was shipped in medium versus 0.9% NaCl solution (Fig. 2b). A comparison of the viability between samples shipped at room temperature versus sampled shipped cooled similarly revealed no significant difference in the samples, and shipment temperature did not affect the screening success rate (Fig. 2c). Although the study protocol requires a minimum tissue size of a “pea-sized tissue fragment”, incoming tissue piece sizes varied significantly depending on the extent of surgical resection. The average tissue volume was 3070 mm³, and 80% of all samples with a size ≥250 mm³ could be screened, while this was only possible in ~50% of all samples smaller than 250 mm³. However, in individual cases, successful screens could be performed on smaller tissue samples derived from stereotactic biopsies with sizes as low as 12 mm³. Notably, a correlation between tissue volume and total derived viable cell numbers and, hence, successful drug screening was observed for rhabdomyosarcomas, Ewing sarcomas and high-grade gliomas but not for neuroblastomas and ependymomas (Fig. 2d).

a Effect of sample shipment in 0.9% NaCl versus medium on sample viability, measured with automated viable and dead cell counting (ViCell or Cellometer). The color code on the right reflects the different tumor diagnoses. The black dots indicate the mean, error bars reflect SD. Students t-test (two-sided). b Violin plot: Effect of sample shipment in 0.9% NaCl versus cell culture medium on sample viability. Colors indicate screen type (full, meaning all three plates versus partial, meaning 1–2 plates versus no screen/screen that failed QC). The black dots indicate the mean, error bars reflect SD. Students t-test (two-sided). Pie diagrams: percentage of samples per screen type for samples shipped in 0.9% NaCl or medium. The color code in the middle reflects the screen type. c Violin plot: Effect of sample shipment at room temperature versus cooled on sample viability, with colors indicating the screening type. The black dots indicate the mean, error bars reflect SD. Students t-test (two-sided). Pie diagrams: percentage of samples per screening type for samples shipped at room temperature or code. The color code in the middle reflects the screen type. d Correlation plots calculated after log10 transformation comparing the volume of tissue piece to the viable cell number and the type of screen (full, meaning all three plates versus partial, meaning 1–2 plates versus no screen/screen (full or partial) that failed QC). The gray area represents the confidence intervals. Not all samples are displayed as volume or cell number data were not always available. Statistical method: Pearson correlation with R. e Accumulated incoming sample number of the two-year pilot phase (n = 132). f Effect of transport duration on sample viability, measured with automated viable and dead cell counting (ViCell or Cellometer). The color code reflects the tumor diagnoses as in panel b). g Viability at DSP seeding. The color code reflects the type of screen (full, meaning all three plates versus partial, meaning 1–2 plates versus screen that failed QC). The black dots indicate the mean, error bars reflect SD. EPDN ependymoma, EWS Ewing sarcoma, NB neuroblastoma, HGG high-grade glioma, RMS rhabdomyosarcoma. n.s.: not significant.

During the first two years (June 2019 to the end of May 2021) of the INFORM DSP pilot program, we received 132 viable tumor tissue samples from 122 patients (Fig. 2e) from 35 pediatric oncology sites in seven European countries (Germany, n = 21; Switzerland, n = 5; Austria, n = 2; Belgium, n = 2; Finland, n = 2; Sweden, n = 2; Greece, n = 1) mostly within 48 h after surgery (mean shipment time 1.2 days). We did not observe substantial differences in the average shipment time and yield of tumor cells for drug screens in Germany versus cases from other countries. Notably, a few samples in transit for as long as four days still demonstrated >90% viability. However, screening success was seemingly higher with a shipment time of fewer than 3 days (Fig. 2f). The average cell viability at seeding for DSP was 85%, with no major difference in the average viability between screens that passed quality control (QC; full or partial, 87% and 82%) or failed QC (81%), indicating that high viability upon seeding does not guarantee screening success. However, with one exception, all samples (n = 4) with viability <65% failed QC (Fig. 2g). Of note, from 89 drug screens (full or partial), 69 (78%) passed internal QC (Supplementary Fig. 1). Despite the large sample heterogeneity in terms of tumor diagnoses, a successful screen was performed for all diagnoses except for the two non-Hodgkin-lymphoma samples (Supplementary Fig. 2). Shipment conditions, screening success rates and QC scores for all diagnoses are summarized in Supplementary Data Table 1.

Tissue dissociation and 3D tumor spheroid preculture (b)

As we initially experienced significant challenges in trying to dissociate and preculture all incoming tumor tissues in a “one-size-fits-all” protocol, we established tumor diagnosis-specific dissociation SOPs for optimal 3D tumor spheroid formation in 6-well preculture and 384-well U-bottom plates as described in the Methods section. The detailed protocols for processing vital tumor material are listed in Supplementary Note 2. These protocols mainly differ in the enzymes used and the incubation time at 37 °C. In principle, tissue dissociation consisted of the following steps: (1) extensive mechanical dissociation prior to (2) enzymatic digestion with papain (brain tumors and brain metastases), trypsin (neuroblastomas) or a mix of trypsin and collagenase II for osteosarcomas, soft tissue sarcomas and rare (non-brain) tumor entities; (3) stopping of the enzymatic reaction, digestion of DNA released from dying cells; (4) filtering of the cell suspension, followed by (5) (repeated) red blood cell lysis; and (6) determination of cell number and cell viability (Fig. 3a). We opted for preculturing the cells following tissue dissociation and prior to drug screening for two main reasons, (i) to avoid priming the cells for cell death due to shipment or dissociation-induced stress potentially resulting in false positive drug hits, and (ii) to increase the number of viable cells to allow drug screening, as in many cases the viable cell yield immediately after dissociation was not sufficient to perform a partial (at least one drug plate) or full drug screen (three plates). Depending on the amount of initial material and the speed of recovery or cell expansion, the preculture time varied, with a current median time of four days (mean 6.9 +/− 9.5 days) and with few (seven in total) samples requiring a preculture time exceeding 14 days (Fig. 3b). An outlier sample requiring 77 days of preculturing was not reported to the tumor board. The different diagnoses were distributed across all culture durations, indicating no correlation between the duration of preculture and tumor diagnosis (color code Fig. 3b). Successful pre- and long-term cultures typically resulted in 3D growing spheres with diverse morphology, as shown in Fig. 3c–e. At the drug screening start time, the precultures were dissociated into single-cell solutions (see “drug screening”, Supplementary Note 5) and, after cell counting, transferred onto 384-well preprinted round bottom drug plates to allow for heterogeneous 3D mini tumor spheroid formation (Fig. 3d; Supplementary Figs. 3–9). We aimed to seed 1000 cells per well to achieve comparable conditions. The cell number per well varied between 400/well and 1250/well, depending on the number of viable cells obtained. In 59% of the screens, the intended number of 1000 cells/well could be used (Fig. 3f).

a Workflow of incoming tissue processing until the generation of patient-derived 3D multicellular fresh tissue culture (FTC). b Violin dot depicting the preculture time of screened samples in days; the y-axis is square-root transformed (sqrt) to better illustrate the distribution of all data points. The color code reflects the different tumor diagnoses (same as in d). The black dots indicate the mean, error bars reflect SD. c Bright field images (×10 magnification, cropped) for patient-derived 3D FTC precultures (d3, d5, d8) at passage 0. d Bright-field image (×10 magnification, cropped) for a patient-derived 3D culture from the same ependymoma (EPDN) sample at p0 (d3), at drug screen (384-well) and p1. e Bright-field images (×10 magnification, cropped) for patient-derived 3D long-term cultures (>p6, LTC). f Violin dot plot displaying the seeding cell number per well of the screened samples. The color code on the right reflects the different tumor diagnoses. The black dots indicate the mean, error bars reflect SD. g Violin dot plot illustrating the tumor cell content (in percent) of fresh frozen material accompanying the fresh tumor specimen submitted for DSP. The color code reflects the screening type (full, meaning all three plates versus partial, meaning 1–2 plates versus no screen/screen that failed quality control (QC)). The black dots indicate the mean, error bars reflect SD. h Immune cell type deconvolution results from the same medulloblastoma (MB) sample from FF (fresh frozen; original tumor), at p. 0 (directly after dissociation) and p. 1 (seeding time-point) with the most commonly used bulk RNA-seq deconvolution tools: CIBERSORT, QuantiSeq, and EPIC. TC tumor cell.

For 66/68 QC-passed samples, the tumor cell content was estimated from the histopathology department for the corresponding fresh frozen tumor. The mean tumor cell content was 78%, and for 48/66 (73%), the tumor cell content was 80% or higher (Fig. 3g). As the tumor cell percentage was lower than 80% in n = 18 samples, we analyzed for a subset of samples the composition of the cell population at the time of sample dissociation (passage 0) and time of drug screening (passage 1) with deconvolution assays, such as 850k methylation arrays (Supplementary Data Table 2) and RNA-seq (Fig. 3h, Supplementary Fig. 10a, b). Immune cell type deconvolution and immune/stromal score determination based on RNA-seq revealed a multicellular composition of the original tumor (FF, fresh frozen), the dissociated samples (passage 0) and of the FTC at the seeding time point for the drug screen (passage 1) with a trend for tumor cell enrichment (Fig. 3h, Supplementary Fig. 10a, b). Overall, the fraction of tumor cells remained largely stable or was further enriched during preculture, with ~50–90% tumor cells after dissociation and ~77–87% tumor cell cells at the time of drug screening (Supplementary Data Table 2).

The metabolic activity readout was performed after 72 h of drug exposure. To test for genomic stability and identity of the spheroid culture, we have exemplarily compared the copy number profile based on WES and lcWGS data for two samples of FTCs (p. 0/p. 1) with the original tumor (FF) (Supplementary Fig. 11), and the 850k DNA methylation profiles and CNA patterns of passaged long-term cultures (LTCs) with the original tumor (FF) in twelve cases. For the two FTC samples, the RNA-seq analysis confirmed that gene fusions present in the original tumor were also present in the FTCs (Supplementary Figs. 12, 13). Methylation profiling-based t-distributed stochastic neighbor embedding (t-SNE) analysis clustered all LTCs closely with the original tumors. Moreover, all FF/LTC pairs showed high similarities to the corresponding reference methylation classes (diagnoses), as evidenced by grouping with these classes in the t-SNE (Fig. 4a). The maintenance of relevant molecular alterations was confirmed by comparison of the CNA plots for the fresh frozen material from the original tumor (FF) and the corresponding LTC, calculated from the same DNA methylation array dataset (Fig. 4b). Overall, all analyses reveal high concordance between the original tumor sample and corresponding FTC or LTC, indicating the ability of our approach to establish cell culture models closely reflecting the respective patient tumor. This result is especially evident in the maintenance of relevant molecular driver aberrations. All assays reveal an enrichment of tumor cells over passaging time, and the enrichment of tumor cells in the LTC can explain most of the CNA-plot differences. One sample, HGG_DMG_K27, showed a slight separation in the t-SNE but still clustered closely to the reference cohort. In this case, we cannot entirely exclude that we may have established a cell culture reflecting an expansion of a subclone of the original tumor.

a t-SNE analysis of DNA methylation profiles for comparison of the original tumors and the corresponding tumor-derived long-term culture (LTC) models with already existing well-characterized reference tumors (malignant rhabdoid tumors, FN-RMS tumors and high-risk (HR) MYCNamp neuroblastomas). b Pairwise comparison of the copy-number profiles of nine tumors (upper panel) and their corresponding LTCs (lower panel) reveal similar genome-wide methylation patterns and maintenance of relevant driver events. FF fresh frozen material of the original tumor, LTC long-term culture, EWS Ewing sarcoma, FN-RMS fusion-negative rhabdomyosarcoma, HGG, DMG_K27M high-grade glioma, subtype K27M mutant diffuse midline glioma HGG, other another subtype of high-grade glioma, MRT malignant rhabdoid tumor, NB, HR, MYCNamp high-risk neuroblastoma with MYCN amplification, osteosarcoma (HG) high-grade osteosarcoma, sarcoma undiff undifferentiated sarcoma. Sample abbreviations: r relapse, p progression.

Moreover, we used Bland-Altman plots to compare the drug sensitivities of FTCs and LTCs obtained from the same original tumor for two exemplary cases. The models showed similar responsiveness to individual drugs, with only a few outliers above or below the 95% limits of agreement (LoA). The LoA is the average difference ± 1.96 standard deviations of the difference and, hence, is a judgment of how well the measurements agree. The smaller the range is, the better the agreement accuracy will be (Supplementary Fig. 10). Hence, our analyses show that the LTCs closely reflect the original tumor specimens by preserving the molecular diagnoses, genetic driver events, and drug sensitivity patterns.

Overall, we have set up an efficient workflow for handling fresh pediatric solid and brain tumor specimens and established a robust tumor spheroid culture system suitable for drug response profiling.

Drug sensitivity profiling and tumor board presentation (c)

Due to the limited availability of tissue under real-world clinical conditions, we used a clinically focused drug library (n = 75–78 drugs) (Supplementary Data Table 3), established within the COMPASS (Clinical implementation Of Multidimensional PhenotypicAl drug SenSitivities in pediatric precision oncology) cooperation covering most of the standard chemotherapeutic drugs used in pediatric oncology treatment protocols and representative small molecular kinase inhibitors, epigenetic modifiers, apoptotic modulators, metabolic modifiers and others (Fig. 5a). Most of the drugs are either European Medicines Agency (EMA) or U.S. Food and Drug Administration (FDA) approved or in late-stage clinical development. All drugs were administered batchwise on 384-well U-bottom plates and stored in an oxygen- and moisture-free environment at room temperature to allow on-demand availability before the seeding of cells.

a Composition of the core drug library, consisting of 75–78 drugs. b Cohort overview. Left: Tumor diagnoses of fresh frozen material used for molecular analysis through NGS (n = 1642). Right: Tumor diagnosis distribution in the present cohort of INFORM samples with vital tissue submission. The pie diagrams represent the distribution of the indicated diagnoses within the whole cohort. The outer circles represent broad tumor categories: sarcomas (magenta), brain tumors (green), hematological malignancies (hemat. malig., red), neuroblastomas (orange) and others (brown). The inner circle represents the more detailed tumor diagnoses within each category, as explained in the color code below the pie charts. c Timeline from surgery to drug report for QC-passed full screens and where timeline data were available (n = 49). ALL acute lymphocytic leukemia, AML acute myeloid leukemia, AT/RT atypical teratoid rhabdoid tumor, DSRCT desmoplastic small-round-cell tumor, EPDN ependymoma, EWS Ewing sarcoma, HGG high-grade glioma, IMFT inflammatory myofibroblastic tumor, NB neuroblastoma, RMS rhabdomyosarcoma, QC quality control.

During the first 24 months of the INFORM DSP pilot program, in 89/132 (67%) of cases, a sufficient volume of viable tissue was received, allowing a partial or full library screen to be performed. During this period, INFORM received 998 fresh frozen tissue samples for NGS. Thus, 13% (132/998) of all INFORM cases submitted for genomic profiling had accompanying viable tissue submitted for drug sensitivity profiling (DSP). The pie diagrams in Fig. 5b display the distribution of broad categories (outer circle) and more detailed tumor diagnoses within these categories (inner circle) for DSP. The most frequent categories were sarcomas (44% soft tissue sarcomas and Ewing sarcomas and 11% osteosarcomas), followed by brain tumors (29%; high-grade gliomas, ependymomas, medulloblastomas and others), neuroblastomas (9%) and a mixed group of rare tumor entities (7%). This distribution of viable tissue samples for DSP reflects the overall distribution of categories and tumor diagnosis in the total INFORM cohort (n = 1642 at the time of data cutoff for this study). The submission of fresh viable tumor specimens is optional when including patients in INFORM, whereas sending fresh frozen material for NGS is mandatory. Thus, some samples, in particular rare tumor diagnoses (i.e., inflammatory myofibroblastic tumors, synovial sarcomas or atypical teratoid rhabdoid tumors) or samples with generally smaller specimen sizes due to difficult surgical approaches, may be underrepresented in the DSP cohort. DSP results from the QC passed drug screens were reported together with the respective NGS data at the weekly INFORM molecular tumor board meeting. The timeline in Fig. 5c illustrates the individual steps of the DSP workflow and corresponding median durations during the pilot phase, with drug treatment fixed at 72 h. The overall median turnaround time of the DSP from surgery to data analysis was 20 days (mean: ~24 days), with a median data analysis time of 12 days (mean: 12.7 +/− 8.5 days). The ‘data analysis’ time also includes the interpretation of the data.

Drug hits were identified after accounting for the following: (i) The drug sensitivity score (DSS_asym)^13,14, adjusted for effects on a set of healthy bone marrow control samples as published in Pemovska et al.^9,15 plus in-house controls (nonmalignant astrocytes and fibroblasts); (ii) the maximal effect of the drug, which should reach 75% inhibition or more; (iii) the absolute IC50, which should be lower than the in vivo C_max concentration; and (iv) the goodness of fit (R²) for the calculated growth curve, which should be 0.8 or higher. In addition, the cohort dot plot detected above-average DSS for individual samples, and the 75th percentile was considered to indicate an above-average response within the cohort. We only reported drugs already approved or in clinical studies to the tumor board. The effects of investigational drugs were considered as confirmation of in-class drug effects, if applicable.

Among 65 (without repetition) successfully screened INFORM samples, 47 (72%) demonstrated at least one drug hit, with some samples exhibiting even 10 or more hits. We could not identify any drug hits for 18 samples (28%) (Fig. 6a). The most frequently reported drug class was apoptotic modulators (e.g., navitoclax), followed by conventional chemotherapy (Fig. 6b). Samples demonstrating sensitivity to the pan BCL2 family inhibitor navitoclax were, in general, also sensitive to the BCL2 selective inhibitor venetoclax and other investigational selective BCL2 family inhibitors (Supplementary Fig. 14a; Supplementary Data Table 4). However, in most cases, venetoclax did not qualify as a hit due to a percent inhibition at C_max below the 50% cutoff, except for two of four neuroblastoma samples. This finding is in line with cell culture and preclinical data from several research groups, which demonstrated the strong effectiveness of BCL2 inhibitors in neuroblastoma models. Hence, venetoclax is currently being evaluated in clinical phase I trials for treating neuroblastoma (clinicaltrials.gov NCT03236857)^16,17,18,19.

a Bar diagram displaying the number of reported hits per sample and in %. b Bar diagram depicting the report frequency (in %) of single drugs and drug classes. Color code integrated. c Unsupervised hierarchical clustering based on DSS_asym quantiles for the MDM2 inhibitors AMG-232 and idasanutlin. The color scale shown at the top represents quantile values. The TP53 status is highlighted with green (mutant) and white (wild-type) bars on the left side. d Pie diagrams reflecting the proportion of samples with reported hits with or without reported NGS (very) high-evidence targets for the total DSP cohort and different diagnoses.

To analyze whether the response to BCL2 inhibitors is linked to the expression of distinct BCL2 family members, we performed a Manova analysis of samples with the highest response (DSS_asym quantile >75%) versus less responsive samples. The analysis identified proapoptotic BBC3 (PUMA) and BCL2L11 (BIM) as significantly upregulated in responsive samples when comparing BCL2 inhibitor-responsive versus less-responsive samples (Supplementary Fig. 14b, c), indicating the expression of these genes as potential biomarkers for response prediction. Conversely, high expression of the antiapoptotic BCL2L12 significantly correlated with a low response to these drugs (Supplementary Fig. 14b, c).

As a first step toward evaluating our DSP platform’s clinical utility and predictivity, we looked at cases with strong genetic driver alterations identified by molecular profiling and the vulnerability to matching drugs. Indeed, in 9/14 patient cases, we identified at least one drug sensitivity hit matching the tumor-driving alteration. Of the five cases with no hit identified, one had an NTRK-fusion with clinical resistance to several lines of TRK-inhibitors (see also Fig. 7), and three harbored a CDK4/6 amplification known to confer resistance to CDK4/6 inhibitors²⁰ (summarized in Supplementary Data Table 5). The matching drug sensitivity becomes especially evident when comparing the sensitivity of the sample of interest to the response of the rest of the DSP cohort, visualized in the DSS_asym quantile waterfall plots (Supplementary Fig. 15; molecular aberration matched drugs are marked with arrows). NGS matching drugs present with the highest quantile ranks. Furthermore, DSP gives additional value to NGS target base identification of vulnerabilities, as it gives information on which drug of the respective drug class is more effective (Supplementary Fig. 15).

a Violin dot plots displaying the mean DSS_asym values for the whole cohort (upper graph) and each drug for the respective sample (lower graph; blue line: mean). The sample of interest (here, case #1 relapse 3) is marked in orange. b Waterfall plot sorting all tested drugs for the case #1 relapse 3 sample upon their DSS_asym values, starting with the highest on the left. c Dot plots depicting the DSS_asym values for the indicated drugs for all successfully screened samples. The sample of interest (here, case #1 relapse 3) is marked in orange. All three MET inhibitors (merestinib, crizotinib, foretinib) show above-average responses, whereas the three NTRK inhibitors all have a DSS_asym of or close to 0. d Copy-number profiles of the FF samples reveal an acquired (or selected) MET amp in the plot for relapse 3.

To further validate our platform, we made use of the fact that TP53-wild-type tumors can be responsive to MDM2 inhibitors, as the tumor suppressor function of p53 is frequently impaired due to enhanced activity of its upstream negative regulator MDM2. However, tumors with inactivating mutations in TP53 are resistant to most MDM2 inhibitors (reviewed in refs. ^21,22,23). To determine whether this correlation was reflected in our DSP results, we visualized MDM2 inhibitor (idasanutlin and AMG-232) sensitivity (determined as DSS_asym quantiles) and the TP53 status of our DSP samples in an unsupervised hierarchical clustering heatmap. Samples harboring a TP53 mutation were less sensitive to MDM2 inhibitors than TP53 wild-type samples. Moreover, samples with clear above-average idasanutlin and AMG-232 sensitivity ranking in the top 25% of the cohort (above 75% quantile, blueish color; Fig. 6c) were exclusively TP53-wildtype tumors. Conversely, samples exhibiting resistance to MDM2 inhibition within the cohort predominantly harbored TP53 mutations. Quantile waterfall plots for representative cases with different diagnoses are depicted in Supplementary Fig. 16.

Importantly, in ~81% (38/47) of all successfully screened INFORM cases with a reported drug hit (n = 47), a drug hit was present in samples in which WES and RNA-Seq did not identify an actionable target with a high or very high evidence level¹ demonstrating added information from ex vivo DSP in the majority of cases. This finding especially holds true for tumor diagnoses in which no or only a few very high or high evidence level targets¹ are typically detected, such as ependymomas, rhabdomyosarcomas or, most evident, Ewing sarcomas (Fig. 6d). An exemplary DSP result, visualized as a DSS_asym quantile plot, for a nephroblastoma sample for which only borderline or very low evidence targets¹ (BCL2, XPO, and HDAC2 overexpression) were identified is shown in Supplementary Fig. 15. In addition to detecting drugs matching the identified molecular alterations (marked by arrows), namely, BCL2 inhibitors navitoclax and venetoclax), XPO inhibitor selinexor and all four HDAC inhibitors present in the library), the DSP demonstrates an unexpected in-class effect for all MEK inhibitors in the library, ranking with the highest quantiles for this sample (marked with asterisks). Biocomputational analyses addressing the underlying (molecular) mechanism of this unexpected sensitivity to MEKi are ongoing.

Overall, our INFORM drug sensitivity pipeline can identify drug hits, matching corresponding molecular driver alterations and, more importantly, can identify unexpected drug sensitivities in a high proportion of pediatric solid and brain tumors lacking clinically relevant molecular targets.

Selected clinical case reports demonstrating ex vivo–in vivo correlation of drug sensitivity profiles

To look into potential correlations between the output of our ex vivo drug sensitivity platform and clinical courses of patients in vivo, we evaluated samples from three patients for whom clinical follow-up data were available. Figure 7 demonstrates case #1, a seven-year-old patient with high-grade glioma enrolled in INFORM in whom a BCR:NTRK2 fusion was identified by NGS (relapse 1). Consequently, the patient was enrolled in an NTRK-inhibitor trial (larotrectinib) and, following progression (relapse 2), a 2nd generation NTRK inhibitor trial (selitrectinib). Following further progression (relapse 3), the patient received another biopsy to obtain tissue for molecular analysis and ex vivo DSP, which passed the QC. Overall, the tumor sample was quite resistant, with a mean DSS_asym below 3.0 (Fig. 7a), and showed complete resistance against all NTRK inhibitors in the library (larotrectinib, selictretinib, and entrectinib), reflecting the clinical course of the patient. However, the sample exhibited high sensitivity to several MET-targeting inhibitors of the library, namely, merestinib, crizotinib, and foretinib (Fig. 7b, c; Table 1). Subsequent NGS analyses revealed an acquired (or selected) MET amplification as a likely resistance mechanism to NTRK inhibition, which is consistent with the drug screening results (Fig. 7d). This finding is in line with other recent findings describing acquired MET amplification as a potential resistance mechanism to NTRK inhibitor therapy, similar to that described for targeted EGFR inhibitor therapy in NSCLC patients^24,25. In addition, unexpected drug hits were identified in this case, including sensitivity to ALK (lorlatinib) and JAK1/2 (ruxolitinb) inhibitors.

In case #2 with relapsed EWSR1:FLI1-positive Ewing sarcoma, serial viable tissue sampling with DSP before and after therapy allowed us to monitor the evolution of drug resistance under multiagent chemo- and targeted therapy. The first sample for DSP was obtained after the 11-year-old patient experienced relapse during treatment according to the EWING2008 study protocol (before RIST). The patient then received a RIST multidrug treatment regimen (rapamycin, irinotecan, dasatinib (Sprycel), temozolomide)²⁶, and was biopsied again fourteen months later with a further relapse (after RIST). Figure 8 shows a shift of drug sensitivity from a more sensitive profile (mean DSS_asym 6.4) to a generally drug-resistant profile (mean DSS_asym 4.0) (Fig. 8a). In particular, DSP revealed the emergence of complete resistance to all four RIST regimen drugs at the second relapse, collected after RIST therapy (Fig. 8b; Table 2). In this case, we identified HDAC inhibitors (e.g., entinostat) as drug hits with potential clinical use and the investigational BET inhibitor I-BET151.

a Violin dot plots of case #2, a relapsed EWSR1:FLI1-positive Ewing sarcoma (EWS) before and after RIST therapy. The dot plots display the mean DSS_asym values for the whole cohort (left) and each drug for the respective sample before RIST treatment (middle) and after RIST treatment (right). The sample of interest before RIST is marked in orange, and the sample after RIST is marked in red. The blue line represents the mean. b Drug dot plots depicting the DSS_asym values for the indicated drugs for all successfully screened samples. The sample of interest before RIST is marked in orange, and the sample after RIST is marked in red. All four drugs displayed a strong decrease in DSS_asym after RIST therapy. c Violin dot plots of case #3, a relapsed CNS-HGNET brain tumor. Shown are DSP 1 and DSP 2 before and after MEMMAT therapy. The dot plots display the mean DSS_asym values for the whole cohort (left) and each drug for the respective case (DSP 1 in the middle; DSP 2 on the right). The sample of interest DSP1 before MEMMAT is marked in orange, and DSP2 after MEMMAT is marked in red. The blue line represents the mean. d Drug dot plots depicting the DSS_asym values for the single treatment (navitoclax or BCL-XL inhibitor A-1155463) and the combination (navitoclax plus dactinomycin or A-1155463 plus dactinomycin). The sample of interest (case #3, CNS-HGNET) is marked in orange. e Dose-response curves of a single compound (navitoclax) and combinatorial treatment (navitoclax plus IC₂₅ 10 nM dactinomycin). The overlay of both curves, reflecting a shift in sensitivity upon combinatorial treatment, is depicted. The combination screen was performed with cryopreserved cells cultured for four days after thawing for drug testing. % inhibition: normalized inhibition of metabolic activity.

Case #3 relates to a patient diagnosed with a CNS-HGNET-MN1 tumor at the age of nine. In the following years, two relapses were surgically removed, the patient received proton therapy and, after the second relapse, a six-month chemotherapy regimen with etoposide and trophophosphamide. We received fresh tumor tissue from relapses three and four (at 13 and 14 years of age, respectively). Between these two relapses, the patient was treated according to the MEMMAT (Medulloblastoma European Multitarget Metronomic Anti-Angiogenic Trial)²⁷ therapy regimen, which included etoposide, cyclophosphamide, and cytarabine (Table 2). The two DSPs revealed a similar shift in the mean DSS_asym as case #2 from 5.2 to 2.5 (Fig. 8c), indicating the emergence of multidrug resistance after treatment. As dactinomycin was one of the top hits reported for the first sample (Table 2), we performed a combination screen with dactinomycin (IC₂₅: 10 nM) against our standard drug library. Potential beneficial combination partners for dactinomycin were identified with the differential combination drug sensitivity score dcDSS_asym¹³, which corresponds to the largest difference between the DSS_asym values for each drug in the presence and the absence of 10 nM dactinomycin (Supplementary Data Table 6). The BCL2 family inhibitor navitoclax showed the largest shift in drug sensitivity (dcDSS_asym 13.4), supported by the in-class effect of the investigational BCL-XL inhibitor A-1155463 (dcDSS_asym 9.3; Fig. 8d). Venetoclax, a selective BCL2 inhibitor, failed to substantially increase drug sensitivity (dcDSS_asym 0.8), indicating a functional role of BCL-XL and not BCL2 in the observed combination activity. This result is also evident from the overlay of the drug response curves for navitoclax in the absence and presence of dactinomycin and the almost complete inhibition of metabolic activity at C_max of 98% through the combination (Fig. 8e).

Overall, these cases demonstrate striking parallels between clinical courses and our ex vivo functional precision medicine platform, suggesting an “imprinting” of drug resistance patterns in the ex vivo drug sensitivity profiles; however, further prospective evaluation of the predictivity of our DSP platform in a larger patient cohort is needed.