A Sequential Targeting Strategy Interrupts AKT-Driven Subclone-Mediated Progression in Glioblastoma

Patients with glioblastoma receiving standard-of-care treatment almost always experience disease progression under primary therapy that is surgery and combined radiotherapy (RT)- and temozolomide (TMZ)-based chemotherapy (RT/TMZ; ref. 1). Research on mechanisms of RT and TMZ resistance is rapidly progressing (e.g., refs. 2–5), but there is little evidence for improvement of clinical outcome in the care of glioblastoma relapse (6, 7). Longitudinal heterogeneity of potential cellular targets in recurrent versus primary tumor tissue is a known phenomenon (8) and there is evidence for subclonal trajectories in the progress towards clinical relapse (cR). Bulk sequencing of paired tissue, obtained before and after primary therapy, reveals oligoclonal regrowth patterns without a stereotype of recurrence-specific gene alterations (9–11). This challenges the view of ongoing genetic evolution as a primary cause of resistance at relapse, despite some evidence that TMZ itself might contribute to resistance by selecting for specific protein activities, e.g., involved in AKT signaling (11, 12).

To further elucidate the cellular dynamics of intra-tumor progression and to define rational targeting strategies at glioblastoma relapse, this study relied on the use of paired clinical samples representing the status of treatment-naïve primary disease and the posttreatment relapse status of the same patient. Discovery, exploration, and validation of respectively paired data suggests an enrichment and AKT-based acquired resistance in subclonal ALDH1A1+ cell hierarchies that appears to be specifically an adaptation to TMZ exposure. We here show the consequent resilience to further TMZ and that henceforth subclonal progression can be interrupted, in preclinical models, by sequential inhibition of AKT providing a yet unconsidered functional targeting approach for future clinical exploration.

Prospectively collected tumor tissue of patients with isocitrate dehydrogenase wild-type (IDH^WT) glioblastoma at the University Bonn Medical Center was used for derivation of the vital cell samples of the ‘discovery cohort’ (n = 15 samples from n = 7 patients; Supplementary Table S1). The ‘validation cohort’ consisted of paired formalin-fixed, paraffin-embedded (FFPE) tissue from treatment-naïve and cR (following initial treatment) of n = 38 patients with IDH^WT glioblastoma (Supplementary Fig. S1). Specimens were provided by the Department of Neuropathology, University of Bonn Medical Center (patient IDs: ‘BN’), the central nervous system (CNS) tumor tissue bank at the Department of Neuropathology, University Hospital Düsseldorf and the biobank of the University Hospital Essen (patient IDs: ‘DE’; Supplementary Table S2). Studies were conducted in accordance with recognized ethical guidelines (Declaration of Helsinki). Respective local ethics committees (institutional review boards) approved the study; all patients provided informed written consent. Tumor classification relied on the World Health Organization guidelines from 2021 (13). The IDH mutation status was determined by IHC with an IDH-R132H–specific antibody followed by sequencing of IDH1 codon 132 and IDH2 codon 172 (14). The O6-methylguanine DNA methyltransferase (MGMT) promoter methylation status of the samples was determined by methylation-specific PCR or by pyrosequencing, according to (15, 16). The ‘external reference cohort’ comprised data of n = 37 IDH^WT glioblastoma patients (treatment-naïve vs. TMZ-exposed relapse) extracted from (ref. 10; see External validation of differentially regulated genes).

Gene expression status of the ‘discovery cohort’ was determined at passage 6±1 applying the GeneChip Human Transcriptome Array 2.0 (Affymetrix Inc., San Diego) at the Department of Genomics, Life & Brain Center, University Bonn, Germany. Briefly, total RNA was isolated from vital cell samples using the RNeasy Mini Kit (Qiagen, Hilden, Germany). Complementary cDNA synthesis of 500 ng total RNA was performed using the Expand Reverse Transcriptase Kit including RNase Inhibitor (Roche, Basel, Switzerland). Sample preparation, hybridization, and quality control were performed according to Affymetrix's recommended protocols. Raw CEL files were imported into the Affymetrix Expression Console, and gene-level analysis was performed (NCBI/GEO accession No. GSE145128). Mapping of probesets to genes was based on the Affymetrix ‘HTA-2_0.na34.hg19.transcript’ annotation file. To assess differential gene expression, we used the R-package ‘limma’. Candidate genes were queried for direct protein–protein interactions using the human-specific interaction database HIPPIE (17). Network layout and data visualization was performed with Cytoscape v.3.7.1 (www.cytoscape.org). Interactions with a low confidence score (<0.49) were discarded. In the resulting protein–protein interaction network, nodes represent genes, labeled nodes represent differentially regulated candidate genes (Supplementary Table S3). We analyzed paired RNA expression data of 37 patients with IDH^WT glioblastoma derived at tumor diagnosis and first tumor relapse (GLASS cohort; ref. 10). We selected the 175 genes identified on the basis of known association with TMZ treatment resistance, from pathways involved in stemness, DNA repair, and AKT signaling. Eight of the genes were differentially regulated by three criteria: (i) Statistically significant difference (FDR < 5%); (ii) Mean fold change (mFC) was > 2 or < −2; (iii) More than 2-fold differentially regulated in the same direction in at least 9 sample pairs.

Differentially enriched gene sets derived from comparing microarray gene expression data of relapse cells (TMZ→eR or cR) and treatment-naïve cells of the ‘discovery cohort’ were subjected to gene set enrichment analysis (GSEA; version 4.0.3 from the Broad Institute at MIT) using the MSigDB C2 database (GMT file c2.all.v6.2.symbols.gmt), 1,000 permutations, and a gene set size between 15 and 500 genes (n = 3,765 gene sets). Normalized enrichment score and FDR were used to determine statistical significance. Only gene sets with an FDR value < 0.05 (n = 246) were considered for further analysis.

DNA isolation was performed on the ‘discovery cohort’ cell samples at passage 6±1 using Allprep DNA/RNA Kit (Qiagen) und Dneasy Blood & Tissue Kit (Qiagen). Library preparation on 50 ng DNA used Twist Human Core Exome Plus (Twist Bioscience). DNA sequencing was conducted on a NovaSeq 6000 system with 2×100 bp as flow cell type. Sequences with Q30 value > 93% were kept for analysis. Demultiplexing was performed with Illumina bcl2fastq (2.20), adapters were trimmed with Skewer (version 0.2.2; ref. 18). Quality trimming of the reads was not performed. DNA-Seq: Trimmed raw reads were aligned to the human reference genome (hg19) using the Burrows-Wheeler Aligner (BWA-mem version 0.7.17-cegat) and ABRA (version 2.18) for realignment of reads in target regions to facilitate accurate indel calling. In the reference hg19 the pseudoautosomal regions (PAR) on chromosome Y were masked (chrY:10001–2649520, chrY:59034050–59363566) to prevent reads that map to this region from being discarded due to mapping to two different chromosomes. Reads that aligned to more than one locus were discarded. Variant calling: A proprietary software (CeGaT, Tübingen, Germany) was used for variant detection including variants with reasonable frequencies (observed frequency of alternative alleles from 10% upwards of sequenced reads). Variants were annotated on the basis of public data bases. Only variants with a pathogenic Functional Analysis through Hidden Markov Models (FATHMM) score were considered for further analysis (v2.3; http://fathmm.biocompute.org.uk). The quality of FASTQ files was analyzed with FastQC (version 0.11.5).

Antigen labeling in FFPE tissue (4-μm sections) relied on standard protocols (Table 1). For permanent 3,3′-diaminobenzidine (DAB)-based labeling, automated processing was performed on the Benchmark system (Ventana Medical Systems, Roche, Mannheim, Germany). Negative controls and IgG isotype controls were processed in parallel with each batch of staining. Quantification used established methods (19). In brief, five visual fields (200x magnification) that contained the highest density of immunoreactive cells were selected for evaluation from every tissue specimen. Areas of unspecific immunoreactivity, necrosis, and macrophages were excluded. Two blinded investigators (B. Scheffler, F.K.H. van Landeghem) estimated the quantity of immunoreactive cells as a cell score (<1% = 0; 1%–5% = 1; 6%–25% = 2; 26%–50% = 3; > 50% = 4) avoiding areas of unspecific immunoreactivity, necrosis, and macrophages. Immunofluorescence (IF) analysis on deparaffinized 4-μm tissue sections used standard protocols for antigen detection, based on the conditions listed in Table 1. Nuclear staining was performed with Hoechst 33342 (1 μg/mL - New England Biolabs, Ipswich, MA) or with 4′,6-diamidino-2-phenylindole (DAPI)-containing mounting medium (Abcam, Cambridge, UK). A Zeiss inverted microscope was used for documentation (Zeiss, Oberkochen, Germany). Co-expression analysis for ALDH1A1 and pAKT involved paired samples from naïve versus RT (n = 2×12) and paired naïve versus RT/TMZ (n = 2×13) treatment conditions. Areas of vital tumor tissue and ALDH1A1+ cell accumulations were selected by two blinded investigators (D. Trageser, B. Scheffler). Ten high power fields (HPF; 400x magnification) were assessed per sample quantifying the presence and pAKT co-expression status, as well as the total number of DAPI+ nuclei (mean of 336 ± 30 ALDH1A1+ cell counts per sample). ALDH1A1/Ki67 co-expression in tumors of the BN46–patient-derived xenograft (PDX) mouse model was determined at the respective survival endpoint in four slides per condition. Six HPFs/slide were assessed for treatments with DMSO [median overall survival (mOS) = 90 days; range 87–93 days] and TMZ+MK2206 (mOS = 142 days; range: 127–148 days). Three to six HPFs/slide were quantified in TMZ-treated PDX brain samples (mOS = 124.5 days; range: 123–126). The total number of DAPI+ and co-expressing ALDH1A1+/Ki67+ cells were determined and used to calculate the relative percentage of double positive cells per HPF.

Short-term expanded vital cell populations were derived from the ‘discovery cohort’ of glioblastoma tissue and maintained in controlled media conditions (see below; Supplementary Table S1). Mycoplasma-negative status of all cultured cells were verified by standard PCR-based testing in 4-week intervals. Paired samples comprised of treatment-naïve cells from primary disease (^naïvecells) and cR or experimental relapse (eR) cells from the same patient. cR samples represent the status post clinical treatment (e.g., RT/TMZ). eR samples represent the status post in vitro exposure to TMZ (TMZ→eR) or irradiation (RT→eR). ^TMZ→eRcells were collected from regrowth of ^naïvecell populations following serial exposure to TMZ (500 μmol/L) for 16 days every other day. ^RT→eRcells were derived from regrowth of ^naïvecell populations following irradiation challenge 24 hours after plating, supplying 2 Gy ionizing irradiation daily for 12 days (6-MeV medical linear accelerator Mevatron MD2, Siemens, Munich, Germany). Recurring growth of ^RT→eRcells was noted 38±11 days after the first irradiation dose.

For data generation, vital samples were used at cell culture passages 5–12. Unless otherwise specified, cells were grown in laminin/poly-L-ornithine coated dishes in defined proliferative media containing Neurobasal, 0.5X N2 supplements, 0.5X B27 supplements, 2 mmol/L l-glutamine, and 1 μg/mL laminin (ref. 20); 20 ng/mL EGF/bFGF (R&D Systems, Wiesbaden, Germany) were provided upon plating and 10 ng/mL EGF/bFGF every other day thereafter; cell culture media, reagents, and analytical compounds were obtained from Life Technologies (Carlsbad, CA) or Sigma-Aldrich (St. Louis, MO).

Unless noted otherwise, vital cells were seeded 24 hours prior to drug exposure onto laminin/poly-L-ornithine coated cell culture dishes. Compounds were provided fresh with every media change and as specified in the procedures thereafter [TMZ, 500 μmol/L; N,N-diethylaminobenzaldehyde (DEAB), 50 μmol/L; MK2206, 1 μmol/L/5 μmol/L; Selleckchem, Munich, Germany]. Cell confluence was determined by software-based cell recognition (Cellavista or Nyone, Synentec, Elmshorn, Germany). AlamarBlue-based cell viability data were obtained at indicated time points as instructed by the manufacturer (Life Technologies) using a Tecan Infinite F200 instrument. For all paradigms, DMSO was applied in corresponding concentrations to control cultures. Experiments were conducted in triplicates unless noted otherwise.

For analysis, 2.5×10⁵ to 5×10⁵ Trypsin-dissociated (Thermo Fisher, Waltham, MA) single cells were provided with 10% FBS (HyClone, Cytiva, Marlborough, MA) centrifuged, resuspended in FACS buffer (2 mmol/L EDTA, 0.1% BSA in 1x PBS), and incubated with human Fc Block (1:100; BD Biosciences, Franklin Lakes, NJ) for 15 minutes at room temperature. The BD Cytofix/Cytoperm Fixation/Permeabilization Kit was used (BD Biosciences) according to the manufacturer's protocol. Conjugated antibodies were used for labeling (Table 2). Incubation of antibodies was performed for 30 minutes on ice. Prior to the analysis, cells were washed and resuspended in FACS buffer. Samples were assayed on a FACSCelesta Flow Cytometer (BD Biosciences) using the FACSDiva software version v8.0.1.1 (BD Biosciences) and data analysis was performed using the FlowJo software version v10.5.3 (BD Biosciences). Unless mentioned otherwise, experiments were conducted in triplicates. The Aldehyde Dehydrogenase assay was performed using the ALDEFLUOR Assay Kit (STEMCELL Technologies, Vancouver, CA) as using manufacturer conditions. ALDH^br cells were quantified on a FACSCalibur or a FACSCanto II system (BD Biosciences) and analyzed by FlowJo according to manufacturer suggestions with DEAB reagent as a negative control.

Thirty microgram of total protein was separated by SDS-PAGE and transferred onto nitrocellulose membranes (Sigma-Aldrich) by electroblotting. Membranes were blocked in 5% milk powder (Carl Roth GmbH, Karlsruhe, Germany) and TBS with 0.05% TWEEN-20 (Carl Roth GmbH) for 1 hour, supplied with primary antibodies overnight at 4°C or for 1 hour at room temperature (β-actin, Sigma- Aldrich) and detected with a horseradish peroxidase (HRP)-coupled secondary antibody using Luminata HRP Substrate (Burlington, MA) and the ChemiDoc System (Bio-Rad, Hercules, CA; Table 3). Intensity quantification was performed using ImageJ software.

Knockdown experiments were performed using a MISSION short hairpin RNA (shRNA) plasmid against ALDH1A1 and a non-targeting control (Sigma-Aldrich) by lentiviral transduction. Lentiviruses were produced in 293FT cells using packaging reagents following the manufacturer's instructions (Sigma-Aldrich). For overexpression, coding sequences of ALDH1A1 or control-GFP were cloned into pLenti6.2/V5-DEST Gateway Vector (Life Technologies). Transduced cells were selected for up to 2 weeks with blasticidin (5 μg/mL; Santa Cruz Biotechnology). Stable expression/knockdown was confirmed by Western blots or by qPCR using standard protocols and the following primer sequences: ALDH1A1, GGACCAGTGCAGCAAATCAT (forward), TACCACGCCATAGCAATTCA (reverse); GAPDH, CTGCTTTTAACTCTGGTAAAGT (forward), GCGCCAGCATCGCCCCA (reverse). For transient, siRNA-based inhibition of ALDH1A1 expression, 10,000 cells/cm² were seeded in ABX-free media one day prior to siRNA transfection. The transfection was performed using Lipofectamine RNAiMAX Transfection Reagent (Thermo Fisher), Opti-MEM Reduced Serum Medium (Thermo Fisher), 5x siRNA buffer (Dharmacon, Lafayette, CO), and the specific siRNAs (20 μmol/L): ON-TARGETplus Non-targeting Pool (5 nmol; D-001810–10–05) and ON-TARGETplus Human ALDH1A1 (216) siRNA – SMARTpool (5 nmol; L-008722–00–0005), both from Horizon Discovery (Waterbeach, UK). After 5 days, cells were seeded for neurosphere formation as described below and an additional siRNA transfection was performed while plating. The siRNA knockdown was confirmed by qPCR using standard protocols and the following primer sequences: ALDH1A1, GCACGCCAGACTTACCTGTC (forward), CCTCCTCAGTTGCAGGATTAAAG (reverse); RPL37A, GACGTACAATACCACTTCCGC (forward), GGAGCGTCTACTGGTCTTTCA (reverse).

0.5 cells/well were seeded in five 96-well plates. Using automated software-based cell recognition (Nyone) the distribution of single cells/well was subsequently confirmed. After 16 days of expansion, the developed colonies and cells were photodocumented for each colony and counted.

Similar to described conditions for the assay (20–22), single-cell suspensions of 5 × 10⁵ cells/cm² were plated in 1:1 Methylcellulose (Sigma-Aldrich)/proliferative media, provided with 20 ng/mL EGF/bFGF, and inoculated into nonadhesive culture dishes. 10 ng/mL EGF/bFGF was added every other day thereafter. At 12 days, first-generation (1°) neurospheres were documented (Nyone) and quantified. From these, a single-cell suspension was plated under identical conditions for analysis of second-generation (2°) neurospheres on day 24 of the assay. To evaluate for generation of neuronal and glial phenotypes, a representative fraction of 2° neurospheres was plated onto laminin/poly-L-ornithine coated dishes and, after attachment, provided with proliferative media devoid of mitogens to allow differentiation for 2 weeks. Samples were then fixed (ROTIHistofix, 4%) and processed for standard IF analysis using the antibodies and dilutions listed in Table 4. Cells were mounted using Aqueous Fluoroshield mounting medium (Abcam) and imaged on a Zeiss inverted microscope. Experiments were conducted in triplicates.

Animal studies were approved by regulatory authority (LANUV#84–02.04.2014.a316). 10-week-old Fox Chase SCID/beige mice (Charles River, Germany) were engrafted with 1×10^{6 naïve}BN46 cells, unilaterally injected into the right striatum (0.7 mm anterior to the bregma, 2.6 mm lateral from the midline, at a depth of 3.0 mm). Drug regimens were initiated at day 60 posttransplantation. Mice received TMZ (50 mg/kg, i.p.; 5 days per week), MK2206 (100 mg/kg, orally, 6 times once every other day) or control solvent (DMSO in PBS); as specified. For TMZ→TMZ/MK2206, mice were first treated with TMZ alone (50 mg/kg, i.p.; 5 days/week) for 3 weeks. Subsequently, TMZ was administered alternating with MK2206 (100 mg/kg, orally; 6 times every other weekday) before mice continued to receive TMZ-only (50 mg/kg, i.p.; 5 days/week). Alternatively, for TMZ/MK2206→TMZ, mice were treated with alternating TMZ and MK2206 for 2 weeks, followed by TMZ-only (5 days/week). Mice were monitored daily and euthanized in the event of clinical deterioration (when neurologic symptoms such as hemiparesis or gait disorder developed or significant weight loss occurred, defined as the loss of > 25% of pretreatment body weight).

Unless otherwise specified, data are presented as mean ± SD. Statistical methods, including corrections for multiple testing are described in each set of data. The significance threshold was set at 0.05. R (The R Foundation for Statistical Computing, v4.1.3) and GraphPad Prism v8 software was used for data analysis and visualization.

The genomic sequencing data have been deposited at NCBI/GEO under the accession no. GSE145128. All computational structural information is available upon reasonable request.

Building on previous strategies to study subclonal dynamics ex vivo (22), we generated a ‘discovery cohort’ from n = 7 patients with glioblastoma (IDH^WT; Supplementary Table S1). The cohort consisted of primary cells isolated from treatment-naïve tissue at initial diagnosis (^naïvecells; n = 7) paired with primary cells that were generated from relapse surgery of the same patients after clinical RT/TMZ treatment (^cRcells; n = 5). In two cases, where relapse tissue was not available, and in one index case (BN118), paired eR cells were generated by serially exposing ^naïvecells to TMZ in vitro (^TMZ→eRcells; n = 3; Fig. 1A). A comparative analysis of transcriptomes derived from these paired samples (n = 15) uncovered the expression of 175 potential markers linked to drug resistance, of which eight were differentially expressed (Fig. 1B; Supplementary Table S3). The differential expression of the top upregulated gene ALDH1A1 (mFC, +6.3) could be validated, among the 175 tested, on a cohort of publicly available RNA sequencing data of paired treatment-naïve and RT/TMZ relapse glioblastoma tissue samples (ref. 10; Fig. 1C). We next studied ALDH1A1 expression by IHC in FFPE tissue from an independent ‘validation cohort’ of n = 38 patients with glioblastoma with available paired primary and relapse tissues (IDH^WT; Supplementary Figs. S1 and S2). Initial treatment of these patients consisted of surgery and either RT/TMZ or radiotherapy-alone (RT-only; Supplementary Table S2). In these samples, we observed distinctive patterns of subclonal enrichment. At primary disease, ALDH1A1+ cells were faintly stained and sparsely dispersed within the tumor. By contrast, focal accumulations of strongly immunoreactive ALDH1A1+ cells were evident in RT/TMZ relapse tissues (Fig. 1D), but not in RT-only relapse tissues (Figs. 1E and F). This finding indicated a relationship between clinical TMZ exposure and the expansion of a putatively clonal ALDH1A1+ cell population in progressing glioblastoma.

ALDH1A1 has frequently been described as a marker for stemness in human cancers (23), even though early investigations found it dispensable for stem cell function in murine hematopoietic and nervous systems (24). Studies on human glioblastoma cell lines suggested the enzyme as a mediator of TMZ resistance (25). Yet, ALDH1A1 immunoscoring in glioma/glioblastoma tissue from cohorts of primary disease inconsistently predicted patient survival (26, 27). To further explore the role of ALDH1A1 under TMZ subclonal selection pressure, we performed a series of genetic and pharmacologic interference assays in vitro.

We first noted that serial exposure of ^naïvecells to TMZ (generating ^TMZ→eRcells) resulted in increased ALDH-bright (ALDH^br) cell frequencies and ALDH1A1 protein levels in eR cell populations. For example, in the ^TMZ→eRBN46 model, the ALDH^br cell fraction increased from means of 6% (^naïveBN46) to 55%, accompanied by a 37-mFC of ALDH1A1 in the flow cytometry analysis (Fig. 2B; Supplementary Fig. S2B). This effect could not be observed upon serial in vitro exposure to ionizing irradiation (resulting in ^RT→eRcells, see Methods; Fig. 2A; Supplementary Fig. S2A). These data further indicated that the observed increase of ALDH1A1+ cells at glioblastoma relapse may be a specific effect of TMZ.

We next tested the ^naïveBN46 patient cells in a protracted in vitro assay and observed that their recurrent growth upon TMZ exposure could be repressed by pharmacologic inhibition of ALDH with DEAB or by knockdown of ALDH1A1 expression with shRNA (Fig. 2C and D). We speculated that, in line with the literature (25), increasing ALDH1A1 would mediate increasing TMZ resistance. In overexpression experiments, we observed that ALDH1A1^{high naïve}cells recovered up to three weeks earlier from serial TMZ exposure compared with control ^naïvecells transduced with green fluorescent protein (GFP Ctrl; Fig. 2E). However, cocultures of ALDH1A1^high and GFP Ctrl cells (1:1 ratio) revealed similar initial drug response kinetics (Fig. 2F). This argued for a similar degree of TMZ sensitivity, at least during initial exposure, in both cell types and against an intrinsic TMZ resistance in treatment-naïve ALDH1A1^high cells. Furthermore, population-based knockdown or overexpression of ALDH1A1 in ^naïveBN46 patient cells did not strongly alter stemness-associated features (Supplementary Fig. S2C). Moreover, we did not observe significant effects on clonogenicity and clonal cell growth as determined by the limiting dilution assay (Fig. 2G; Supplementary Fig. S2D). Similarly, in the neurosphere assay, the frequency of neurosphere-forming cells (indicator for clonogenicity), the ratio between primary (1°) and secondary (2°) neurospheres (indicator for self-renewal), and the ability to generate neuronal and glial phenotypes (indicator for potency), did not vary significantly among the tested conditions (Fig. 2H; Supplementary Fig. S2E–S2G).

This set of data confirms on the one hand the association of TMZ and ALDH1A1 expression in glioblastoma, but on the other hand suggests that ALDH1A1 is not necessary to drive subclonal progression of primary disease.

Because glioblastoma stem cells have been shown to adapt under treatment (28), and in particular, ALDH1A1+ cells have been suggested to adapt in response to clinical EGFR inhibitor therapy (29), we speculated that ALDH1A1+ cells may acquire adaptive molecular features upon TMZ exposure. To test this hypothesis, we compared ALDH1A1 cells from treatment-naïve and paired relapse conditions in vitro.

Clinical RT/TMZ relapse cell samples demonstrated consistent, yet variably strong increases of ALDH1A1 compared with their paired ^naïvecell populations (Fig. 3A; Supplementary Fig. S3A). By flow cytometry, we determined respective mFCs of 84.0 (^cRBN91), 32.3 (^cRBN118), 0.9 (^cRBN123), and 2.2 (^cRBN132). The data distribute around the mFC of 37.0 recorded for eR ^TMZ→eRBN46 cells (Supplementary Figs. S2B and S3B), which we further used as a reference for analysis. Performance of these ‘naturally’ ALDH1A1-enriched cR samples was then evaluated in the neurosphere assay (Supplementary Figs. S3C and S3D). We noted that knockdown of ALDH1A1 in ^cRBN118 and ^cRBN123 cells resulted in a uniformly strong decrease of clonogenic activity, by 70%±4% and 69%±7% for 1° and 2° neurospheres, which is in strong contrast to the almost unchanged activity after knockdown of ALDH1A1 in the paired treatment-naïve samples (Fig. 3B). Similarly, for the ^TMZ→eRBN46 cells, knockdown of ALDH1A1 reduced 1° and 2° neurospheres by 48%±2% and 56%±7% respectively, while the clonogenic activity of the parental ^naïveBN46 cells remained hardly affected (Supplementary Figs. S3D and SD3E). These data further supported the hypothesis that exposure of TMZ leads to an enrichment of ALDH1A1+ cells characterized by altered functional activity.

To discover potential underlying mechanisms, we used GSEA on the paired expression data of the ‘discovery cohort’ revealing PI3K–AKT pathway genes more frequently enriched across the leading edges of all significantly enriched pathways (Fig. 3C). To test if this enrichment underlies genetic events, we analyzed single nucleotide variants in the paired samples using whole-exome sequencing. FATHMM, retaining all variants documented as pathogenic with a major allele frequency of at least 10%, did not reveal significant differences in the mutational profile of relevant PI3K–AKT pathway driver genes in treatment-naïve versus relapse samples (P = 0.9743; Supplementary Table S4). However, we did observe increased pathway activity demonstrated in paired ^naïvecells vs. experimental and cR samples by Western blot and flow cytometry (Fig. 3D and E; Supplementary Figs. S3F and S3G). In particular, pAKT+/ALDH1A1+ double-positive cells represented only 5%±0.04% of the treatment-naïve cells and up to 49% (mean 28%±13%) of the total cell population in the paired cR samples. This subclonal enrichment was consistent in 4 of 4 pairs of naïve versus clinical RT/TMZ relapse samples with mFCs of 3.8 (^cRBN91), 7.9 (^cRBN118), 2.8 (^cRBN123), and 3.8 (^cRBN132; Fig. 3F; Supplementary Fig. S3H) and in naïve versus eR cells (^TMZ→eRBN46; mFC = 3.1; Supplementary Figs. S3I and S3J), but not in in vitro exposed naïve cells to ionizing irradiation (Supplementary Fig. S3J). Furthermore, the emergence of pAKT+/ALDH1A1+ cellular enrichment could be verified as well in TMZ-treated tumor xenografts (Fig. 3G) and in original patient tissue (Fig. 3H). IF analysis of FFPE specimens of the ‘validation cohort’ (Fig. 3H and Supplementary Table S2) revealed in the treatment-naïve as well as in the paired clinical RT-only relapse tumor tissue, rare ALDH1A1+ cells intermixed with heterogeneously distributed pAKT+ cells. By contrast, we noted characteristic clusters formed by abundantly co-expressing ALDH1A1+/pAKT+ cellular phenotypes in the paired clinical RT/TMZ relapse tissue, but only few in the RT-only relapse tissue (Fig 3I), indicating that this effect was specific to TMZ.

pAKT is known for its heterogeneous distribution in human glioblastoma (30) and there is a large body of evidence on glioblastoma-promoting AKT activity (31). Early research had already suggested an involvement of AKT in stem-like properties of cancer cells and during glioma formation (32, 33). A subclonally confined process of aberrant AKT signaling that evolves dynamically under TMZ exposure from primary disease towards glioblastoma relapse has hitherto not been described.

To explore drug targeting under this premise, we tested the clinical-grade allosteric AKT inhibitor MK2206 (34) in vitro, in a paradigm involving continuous TMZ exposure (Fig. 4A). TMZ resistance levels were consistently increased in the (TMZ preexposed) experimental and cR cells vs. their paired treatment-naïve samples, as quantified by read-out of cell confluence and viability on days 8 and 10, respectively (Fig. 4B and C). Combinatorial application of MK2206 (5 μmol/L) did not alter the sensitivity of treatment-naïve cells to TMZ, but it efficiently reverted the high levels of TMZ resistance in relapse cells to the more sensitive status of paired treatment-naïve cells (Supplementary Fig. S4A). Intriguingly, we noted that the combinatorial application of MK2206 in vivo also reverted the increasing levels of proliferative activity in ALDH1A1+ cells characteristically observed in TMZ-treated tumor xenografts (Fig. 4D and E; Supplementary Fig. S4B). This correlated with increasing levels of active caspase-3, a classic apoptosis marker (ref. 35; Fig. 4F). Notably, the pro-apoptotic effect of MK2206 + TMZ in relapse cells even appeared to exceed the antitumor effects that TMZ-alone had in treatment-naïve cell samples (Fig. 4G). And, in line with the hypothesis of an AKT-driven subclonal progression of ALDH1A1+ cells, we observed that the application experiments of MK2206 + TMZ targeted the vast majority of ALDH1A1+ cells within the experimental and cR cell populations (Fig. 4H; Supplementary Fig. S4A).

On the basis of the collective findings of our study, we next designed a rational targeting strategy for ALDH1A1+ glioblastoma cells. To interfere with the dynamics of ALDH1A1+ subclonal progression, we assayed TMZ and MK2206 in the long-term in vitro approach originally used to generate experimental ^TMZ→eRcells. We observed that MK2206 alone did not inhibit cellular proliferation, it rather prevented eR when applied subsequently to TMZ exposure in vitro (Fig. 4I). Sequential treatment was therefore modeled in follow-up xenograft experiments. Schedules were initiated in SCID/beige mice at day 60 after orthotopic implantation of ^naïveBN46 cells, corresponding to 67% of the vehicle control mOS time (DMSO = 90 days; Fig. 4J; Supplementary Fig. S4C). MK2206 monotherapy did not alter survival, but TMZ did, in line with clinical expectations [mOS (TMZ) = 112 days; mOS (DMSO vs. TMZ), P = 0.0005]. The combination of MK2206, within the first 2 weeks of TMZ treatment, did not have an added benefit. However, delayed addition of MK2206, during weeks 4 to 5, of TMZ exposure extended survival significantly [mOS (TMZ→TMZ/MK2206) = 136 days; mOS (TMZ vs. TMZ→TMZ/MK2206), P < 0.0001; Fig. 4D]. In summary, these results showed that the appropriately timed treatment sequences of TMZ + MK2206 yielded a > 2-fold survival benefit over the already effective TMZ monotherapy, illustrating a sequential targeting paradigm for rational interference with AKT-driven subclonal progression in glioblastoma (Fig. 4K).

TMZ-based treatment schemes represent a cornerstone of the primary care measures in glioblastoma (1). Resistance to TMZ is considered the limiting factor of effective therapy; a variety of mechanisms and molecular constellations (e.g., an unmethylated MGMT promotor status) likely contribute to resistance of tumor cells against the actions of this chemotherapeutic (5). Restricted populations of preexisting glioma stem cells had been considered as intrinsically resistant and as drivers for progression of disease (e.g., ref. 37). It was tempting to likewise consider the rare ALDH1A1+ cells at primary disease as prototypical resistant clones, because they accumulate in paired recurring tumor tissue after clinical treatment. Our investigation could not reveal evidence for primary resistance, we instead observed that these cells can adapt, specifically to TMZ exposure, and then proliferate under the continued influence as a more TMZ resistant hierarchy, driven by aberrant AKT signaling. This occurred irrespective of the MGMT promotor status and without association of mutational profiles in relevant PI3K–AKT pathway drivers in the investigated cohort. For an overall statement on the potential impact of molecular constellations, including the status of MGMT expression, appropriately powered translational and clinical studies are needed. Nevertheless, the dynamics of this process resembled the behavior of drug-tolerant persister cells found in response to chemotherapies and targeted anticancer agents, e.g., in melanoma, lung and colon cancer (38, 39). Drug tolerance and stem cell plasticity have already been described in glioblastoma, at least following receptor tyrosine kinase (RTK) inhibition, via epigenetically mediated Notch signaling (40). Resistance mechanisms that operate at the phenotypic level conferring resilience through adaptive plasticity represent a new theme in glioblastoma research, because this could explain why limited genetic evolution is observed at disease recurrence (9–11, 41, 42).

ALDH1A1+ cells may represent an archetype of glioma stem cells in that regard and extensive follow-up investigation could aim at characterizing the molecular machinery and the microenvironmental interactions that these cells use to resist treatment. In this study, our main concern was the therapeutic consequence of their subclonal progression, particularly considering the very limited range of clinical options at glioblastoma relapse (1, 43, 44). The most obvious choice would be the targeting of ALDH1A1 itself. There are drugs already available broadly inhibiting aldehyde dehydrogenases (e.g., disulfiram), which unfortunately can show severe side effects and toxicity (45). Isoform-specific inhibitors are under development (23), but their safety profile remain to be elucidated under the premise that ALDH1A1 is also expressed during development and in the adult brain (27, 46). We therefore approached the alternative, provided by the marked upregulation of pAKT in the target cells. Our study however lacks detailed analyses on the mechanistic links between ALDH1A1 and AKT, which may be part of a complex regulatory mechanism involving epigenetic regulation, lipid and retinol metabolism, mitochondrial function, and pleiotropic downstream effects on oncogenic signaling (23, 25, 29). We also considered, but referred for future investigation, to clarify whether re-irradiation may have a potential benefit under this constellation at glioblastoma relapse, which could have an impact on current clinical practice (47–49).

AKT is a known suppressor of apoptosis (50), and clinical-grade inhibition can be associated with improved progression-free and overall survival in advanced, pathway-altered solid tumors (51, 52). Targeting the related axis of RTK–PI3K–AKT–mTOR signaling is an ongoing matter in clinical neuro-oncology trials as well (53). AKT inhibitors can be well tolerated, but appear to remain ineffective as a monotherapy, at least in heavily pretreated patients (54). This may be attributed to the pharmacokinetic challenges of the blood-brain-barrier and CNS penetration (55), but it also could be related to the need of an adequate combinatorial treatment (56) and, as we would like to suggest here, to the additional need to stratify patients based on the dynamics of subclonal progression.

Targeting subclonal progression is not yet standard of care in glioblastoma. Our data on clinical specimens suggest that the expansion of individual subclones may be a direct consequence of a therapeutic component in the preceding treatment schedule. The cells that initiate subclonal enrichment may not be intrinsically resistant to that specific therapeutic. Thus, for the detection of aberrant subclones, analysis of recurring tumor tissue may be mandatory. As we here reveal ALDH1A1+ cells as a dynamic and TMZ-adaptive population during first-line treatment, future second-line trials would have to determine the content of ALDH1A1+/pAKT+ target cells prior to the addition of AKT inhibitors. To optimize treatment response to this dynamic and indirect targeting approach, AKT inhibition should only be initiated when the target population is considerably expanded. To determine the most appropriate time point and duration of the sequence, adaptive trial designs may be needed.

Combinatorial drug therapy is widely considered a mainstay of treatment in malignant cancers (57), and heterogeneity of driving disease mechanisms and associated targets may serve as a well-grounded basis for the choice of treatment combinations (58). Our work additionally highlights the fact that the timing of combinatorial schemes may affect the dynamics of intra-tumor subclones and via their expansion the occurrence of treatment resistance and, eventually, the course of disease. Recent evidence in a subset of malignant glioma has already exposed that concomitant vs. adjuvant TMZ schedules may impact on survival outcome (59). This further emphasizes the importance of the chronological sequence of compound combinations for rational interference with glioblastoma progression. Targeting the predictable subclonal trajectory of ALDH1A1+ cells by temporally restricted interference with AKT may be swiftly applicable to clinical practice and outperform contemporary TMZ-based treatment standards.

S. Kebir reports grants and personal fees from Novocure outside the submitted work. H. Fröhlich reports grants from German Federal Ministry for Science and Education during the conduct of the study. J. Matschke reports grants from Federal Ministry of Education and Research (BMBF), German Research Foundation DFG, and European Union's Framework Program for Research and Innovation Horizon 2020 (2014–2020) under Marie Skłodowska-Curie (ITN THERADNET) during the conduct of the study. V. Jendrossek reports grants from Deutsche Forschungsgemeinschaft (DFG), Federal Ministry of Education and Research, and European Commission during the conduct of the study as well as other support from Novartis Pharma AG and Lediant Biosciences Inc. outside the submitted work. B.M. Grüner reports grants from German Research Foundation — 341532333, German Research Foundation, German Cancer Aid, and German Research Foundation — 405344257 during the conduct of the study as well as grants from Federal Ministry for Economic Affairs and Energy outside the submitted work. M. Weller reports grants from Quercis and Versameb and personal fees from Novocure, Bayer, Merck (EMD), Orbus, and Philogen outside the submitted work. O. Brüstle is a cofounder and CEO of and has shares in Life & Brain GmbH; in addition, O. Brüstle is a coinventor on a patent for peripheral zone tumor cells, methods for their preparation and use issued (EP2324111; AU2009291203; US9103819). M. Glas reports personal fees from Roche, Novartis, Daiichi Sankyo, Bayer, Janssen-Cilag, Merck, Kyowa Kirin, and Seagen; grants and personal fees from Novocure; and grants from Life & Brain GmbH outside the submitted work; in addition, M. Glas has a patent for peripheral zone tumor cells, methods for their preparation and use, issued. B. Scheffler reports nonfinancial support and other support from DKTK and grants from DFG, BMBF, Bayer, and VW Foundation during the conduct of the study as well as grants from Life & Brain GmbH outside the submitted work; in addition, B. Scheffler has a patent for peripheral zone tumor cells, methods for their preparation and use issued. No disclosures were reported by the other authors.

S. Kebir: Data curation, formal analysis, investigation, visualization, methodology, writing–original draft, writing–review and editing. V. Ullrich: Data curation, formal analysis, investigation, visualization, writing–review and editing. P. Berger: Data curation, formal analysis, investigation, visualization, writing–review and editing. C. Dobersalske: Data curation, formal analysis, investigation, visualization, writing–review and editing. S. Langer: Data curation, investigation, visualization, writing–review and editing. L. Rauschenbach: Formal analysis, methodology, writing–review and editing. D. Trageser: Investigation, writing–original draft. A. Till: Investigation, writing–original draft. F.K. Lorbeer: Data curation, investigation. A. Wieland: Formal analysis. T. Wilhelm-Buchstab: Methodology. A. Ahmad: Data curation. H. Fröhlich: Supervision, methodology. I. Cima: Formal analysis, writing–review and editing. S. Prasad: Formal analysis, methodology. J. Matschke: Methodology. V. Jendrossek: Methodology. M. Remke: Formal analysis. B.M. Grüner: Formal analysis. A. Roesch: Conceptualization, formal analysis, methodology. J.T. Siveke: Conceptualization, formal analysis. C. Herold-Mende: Conceptualization, formal analysis. T. Blau: Data curation, formal analysis, methodology. K. Keyvani: Resources, formal analysis, methodology. F.K.H. van Landeghem: Data curation, formal analysis, methodology. T. Pietsch: Resources, formal analysis. J. Felsberg: Data curation, formal analysis, methodology. G. Reifenberger: Resources, formal analysis, methodology. M. Weller: Conceptualization, formal analysis. U. Sure: Conceptualization, resources. O. Brüstle: Conceptualization, formal analysis. M. Simon: Resources, methodology. M. Glas: Conceptualization, formal analysis, funding acquisition, methodology, writing–review and editing. B. Scheffler: Conceptualization, formal analysis, supervision, funding acquisition, methodology, writing–original draft, project administration, writing–review and editing.

The authors wish to thank Sabine Normann and Mihaela Keller for their assistance in cell culture work and Andreas Waha for their assistance in determining MGMT promoter status. Roman Reinartz, Raggu Saman, Gabor Petzold, Joao Dinis, and Emre Kocakavuk were involved in explorative studies and helped to collect preliminary data not considered for this manuscript. Rolf Fimmers has provided initial advice on the statistical evaluation of preliminary data. Cartoons and drawings generated using Servier Medical Art, provided by Servier, licensed under a Creative Commons Attribution 3.0 unported license. Composition of figures in Affinity Designer v1.10.5.

This work was supported by the German Cancer Consortium (DKTK) and funded by the Deutsche Forschungsgemeinschaft (DFG) – SCHE656/2–1, proj#418179322 and SCHE656/2–2, proj#405344257, and by a grant of the Federal Ministry of Education and Research (BMBF; FKZ03V0785). Additional grants were provided to individual investigators.

B. Scheffler: Lichtenberg program / VW Foundation, AZ82779.

I. Cima/B. Scheffler: Wilhelm Sander Stiftung (2017.148.2)

B.M. Grüner: supported by the DFG Emmy Noether program (GR4575/1–1)

J.T. Siveke: supported by the German Cancer Aid (grant no. 70112505, PIPAC; grant no. 70113834 PREDICT-PACA), the Wilhelm-Sander Foundation (grant no. 2019.008.1), the DFG through grant SI1549/3–1 (Clinical Research Unit KFO337) and SI1549/4–1, and the Federal Ministry of Education and Research (BMBF; SATURN3 consortium).

The publication costs of this article were defrayed in part by the payment of publication fees. Therefore, and solely to indicate this fact, this article is hereby marked “advertisement” in accordance with 18 USC section 1734.