Main

Gliomas are a heterogeneous group of primary brain tumors that are associated with diffuse brain infiltration and premature death1,2. Central nervous system (CNS) World Health Organization (WHO) grade 2 and 3 gliomas initially grow at a slower rate than glioblastomas (CNS WHO grade 4) but later transform into aggressive tumors with neovascularization and contrast enhancement on magnetic resonance imaging (MRI)1,3,4,5. Diffuse gliomas in adults cannot be cured by surgery, radiation or chemotherapy and are associated with considerable disease- and treatment-associated morbidity1,6. There is an unmet need for new therapeutic options with favorable safety profiles and the potential for longer treatment duration1,6.

Most LGGs in adults harbor mutations in the genes encoding the metabolic enzyme IDH1 or, rarely, IDH2 (refs. 1,7). Cancer-associated mutations confer the enzyme with the neomorphic ability to catalyze the production of 2-HG8,9. 2-HG accumulates in tumor tissue and inhibits 2-oxoglutarate-dependent enzymes, a family of enzymes that includes the TET family of 5-methylcytosine (5mC) hydroxylases, the JmjC family of histone demethylases and many other enzymes controlling a wide range of cellular functions10. Compared with gliomas without IDH mutations, mIDH gliomas follow a distinct molecular pathogenesis, with a characteristic pattern of genomic and epigenetic alterations11,12,13,14.

Small-molecule inhibitors of mIDH enzymes have emerged as a new strategy for the treatment of mIDH cancers. Ivosidenib, an inhibitor of the mIDH1 enzyme, is approved for the treatment of subsets of mIDH1 acute myeloid leukemias and previously treated, locally advanced/metastatic cholangiocarcinomas, and has shown preliminary antitumor activity in patients with mIDH1 glioma and chondrosarcoma15,16,17. Vorasidenib, a dual inhibitor of mIDH1 and mIDH2 enzymes, was designed for improved penetrance of the blood–brain barrier and has also shown preliminary antitumor activity in patients with mIDH glioma18,19. Before advancing ivosidenib or vorasidenib to randomized phase 3 evaluation, we conducted the current perioperative study to document inhibition of the mIDH enzyme and mIDH pathway-related pharmacodynamic (PD) effects in on-treatment tumor biopsies in a side-by-side evaluation of both agents. We examined two different dosing schedules for each agent to identify the optimal biological dose in patients with mIDH1 glioma.

Results

Patient characteristics

Patients were assessed for eligibility between March 2018 and April 2019 across seven sites in the USA. Enrollment was completed in April 2019 and follow-up of the study remains ongoing. As of 29 April 2020 (analysis cutoff date), 49 patients overall were randomized before surgery. Patients in cohort 1 were randomized in a 2:2:1 ratio to ivosidenib 500 mg q.d., vorasidenib 50 mg q.d. or no treatment before surgery. After documenting inhibition of the mIDH1 enzyme in cohort 1, cohort 2 was opened to test alternative dose regimens and patients were randomized 1:1 to ivosidenib 250 mg twice daily (b.i.d.) or vorasidenib 10 mg q.d. Treated patients received 28 (+7) d of drug up to and including the day of surgery. All patients had the option to receive postoperative treatment until disease progression or unacceptable toxicity (Fig. 1). Intra-patient dose escalation was permitted per protocol. Tumor and blood samples were collected and analyzed per protocol. On-treatment tumor tissue was compared with tumor tissue from a previous surgery whenever possible (Extended Data Fig. 1).

Fig. 1: Study design.
figure 1

An overview of the study design is shown. All patients could opt to receive the study drug postoperatively. After surgery, patients in the untreated control group were re-randomized 1:1 to either vorasidenib 50 mg q.d. or ivosidenib 500 mg q.d. Based on the PD and pharmacokinetic results of cohort 1, alternative dose regimens of vorasidenib and/or ivosidenib were to be tested in cohort 2.

Full size image

Overall, 24 patients received at least one dose of vorasidenib and 25 patients received at least one dose of ivosidenib. Patients who had been randomized to the control arm before surgery were re-randomized after surgery to either vorasidenib 50 mg q.d. (n = 2) or ivosidenib 500 mg q.d. (n = 3) (Extended Data Fig. 2). At the time of the analysis cutoff date, 17 (70.8%) patients remained on vorasidenib treatment. Five (20.8%) patients discontinued vorasidenib owing to disease progression and two (8.3%) discontinued per investigator decision; 15 (60.0%) patients remained on ivosidenib treatment, three (12.0%) did not continue ivosidenib postoperatively, six (24.0%) discontinued ivosidenib owing to disease progression and one (4.0%) discontinued owing to an adverse event (AE).

Demographic and baseline characteristics were similar for the vorasidenib and ivosidenib cohorts (Table 1). Most patients had WHO grade 2 tumors (43 of 49; 87.8%) based on the most recent pathology before screening. All patients had at least one previous surgery; 24 (49.0%) received previous systemic therapy and 14 (28.6%) received previous radiation therapy.

Table 1 Demographic and baseline characteristics
Full size table

Tumor 2-HG and drug concentrations

Tumors from 40 of 49 patients, including all five untreated patients, were included in the tissue analyses. Nine patients were excluded from the tissue analysis because they did not have enough remaining tissue (n = 2), mIDH1 was not confirmed in resected tissue (n = 3) or they received incorrect drug doses before surgery (n = 4).

In the external control group of archival tumors, the mean (s.d.) tumor 2-HG concentration was 3.7 (3.1) µg g−1 in wild-type (WT) IDH gliomas and 276.8 (231.4) µg g−1 in mIDH1 gliomas. The mean (s.d.) tumor 2-HG concentration in patients who did not receive study treatment before surgery (untreated controls) was 154.9 (146.9) µg g−1. Mean (s.d.) tumor 2-HG concentrations in patients who received treatment before surgery were 8.9 (4.1) µg g−1 (vorasidenib 50 mg q.d.), 67.5 (65.4) µg g−1 (vorasidenib 10 mg q.d.), 20.9 (30.7) µg g−1 (ivosidenib 500 mg q.d.) and 16.8 (18.1) µg g−1 (ivosidenib 250 mg b.i.d.) (Fig. 2a).

Fig. 2: Tumor 2-HG concentration and associated molecular changes.
figure 2

a, 2-HG concentrations in external control and evaluable on-study resected tumor samples. Horizontal lines denote median values; boxes denote the 25th to 75th percentiles and the whiskers go from the smallest to the largest values. Colors indicate dose cohorts represented in Fig. 1. b, The percentage of Ki-67+ cells from on-treatment tumor samples in association with 2-HG. c, Levels of DNA hydroxymethylation from on-treatment tumor samples in association with 2-HG. The percentage of 5hmC was calculated as the ratio 5hmC:cytosine. d,e, Modulation of selected molecular pathways on 2-HG suppression. P values are adjusted for multiplicity as described in Methods. f, An unsupervised clustering of 2-HG–associated genes in the IFN-α/γ response pathways. Box plot: the horizontal lines denote median values, the boxes denote 25th to 75th percentiles and the whiskers go from the smallest to the largest values (n = 30 for cluster 1 and n = 11 for cluster 2; two-sided P value generated with Student’s t-test). g, The densities of CD3+ and CD8+ T cells from on-treatment tumor samples in association with 2-HG. For b, c and g, two-sided P values are simple linear regressions. IVO, ivosidenib; VOR, vorasidenib.

Full size image

The posterior mean percentage reduction in tumor 2-HG relative to the combined data from all untreated control tumors was 92.6% (95% CrI, 76.1–97.6) with vorasidenib 50 mg q.d. and 91.1% (95% CrI, 72.0–97.0) with ivosidenib 500 mg q.d., respectively. Correction of tumor 2-HG concentrations for tumor cellularity (‘normalized 2-HG’) reduced the variability in each treatment cohort (Extended Data Fig. 3).

Tumor concentrations of vorasidenib and ivosidenib were well above the reported half-maximal inhibitory concentration (IC50) for inhibition of the mIDH1-R132H allele for vorasidenib and ivosidenib, respectively18,20. Tumor:plasma ratios were considerably higher for vorasidenib than for ivosidenib (Supplementary Table 1).

Safety

All patients proceeded to surgery as planned without any treatment-related delays. Both drugs were well tolerated. AEs were similar to previous studies of these agents and are listed in Table 2. Treatment-emergent AEs are listed in Supplementary Table 2.

Table 2 Adverse events
Full size table

IDH pathway-related molecular and cellular changes

Staining of resected tumor tissue with an antibody against the Ki-67 antigen, a marker for tumor cell proliferation, showed a positive correlation between tumor cell proliferation and tumor 2-HG concentrations (Fig. 2b), even after tumor 2-HG concentrations were corrected for cellularity (Extended Data Fig. 4a).

We observed an inverse correlation between tumor 2-HG concentrations and DNA 5-hydroxymethylcytosine (5hmC) content (Fig. 2c). There was no correlation between tumor 2-HG and DNA 5-methylcytosine (5mC) content (Extended Data Fig. 4b).

Genome-wide RNA expression profiling and data analysis, using established molecular pathway annotations (Molecular Signatures Database, C2 Pathways)21, showed that low tumor 2-HG was associated with a reversal of the ‘proneural’ gene expression signature (Fig. 2d and Supplementary Table 3), a molecular hallmark of mIDH gliomas22 and downregulation of genes linked to stem cell properties in a variety of cancers (Extended Data Fig. 5a,b). The last finding is consistent with the reported effect of 2-HG on cellular differentiation23,24,25,26,27.

Among gene sets that represent specific biological states or processes (Molecular Signatures Database, Hallmark Pathways), the interferon (IFN)-α and IFN-γ pathways were the most highly upregulated pathways at low tumor 2-HG concentrations. In contrast, genes associated with cell-cycle progression (G2M_Checkpoint) were suppressed at low tumor 2-HG levels (Fig. 2e and Supplementary Table 4).

Using linear regression analysis, we identified 762 genes that were induced or repressed in response to 2-HG suppression (Extended Data Fig. 5c,d). This list included genes associated with IDH mutations in both astrocytomas and oligodendrogliomas (Extended Data Fig. 5e,f), genes associated with cellular differentiation in the CNS (Extended Data Fig. 5a,b) and gene sets related to immune cell activation (Fig. 2f).

Examination of formalin-fixed, paraffin-embedded (FFPE) tumor tissue from the resection showed an inverse correlation between tumor 2-HG and tumor-infiltrating CD3+ and CD8+ T cells (Fig. 2g), and again showed an association between 2-HG suppression and upregulation of antigen presentation and the IFN pathways (Extended Data Fig. 6).

Matched-pair analysis from surgery 1 (archival tumor tissue from previous surgery) and surgery 2 (on-treatment surgery) suggested that more complete tumor 2-HG suppression was required to promote tumor infiltration with CD3+/CD8+ T cells (Extended Data Fig. 7a,b) and inhibit tumor cell proliferation (Extended Data Fig. 7c,d).

Preliminary assessment of antitumor activity

Investigator-assessed tumor response showed a decrease in tumor size after postoperative treatment with either vorasidenib or ivosidenib (Fig. 3a). Patients without residual disease (n = 11) were considered to have the best response of stable disease if disease progression had not been documented. The objective response rate (ORR) for vorasidenib 50 mg q.d. was 42.9% (95% confidence interval (CI), 17.7−71.1), including two partial responses (PRs) and four minor responses (mRs), and 10.0% (95% CI, 0.3−44.5) for vorasidenib 10 mg q.d. (one mR). The ORR for ivosidenib 500 mg q.d. was 35.7% (95% CI, 12.8−64.9), including three PRs and two mRs, and 12.5% (95% CI, 0.3 − 52.7) for ivosidenib 250 mg b.i.d. (one PR) (Supplementary Table 5). The median postoperative treatment duration was 14.3 months (range 0.9–22.6 months) for vorasidenib and 15.1 months (range 1.8–22.1 months) for ivosidenib (Fig. 3b).

Fig. 3: Postoperative tumor response assessment and treatment duration.
figure 3

a, The best percentage change in SPD compared with postoperative baseline MRI and the overall ORRs by assigned treatment. Eleven patients had an on-study gross total resection without residual disease and were considered to have best response of stable disease as long as disease progression had not been documented. Patient 22 in the vorasidenib group had a >50% reduction in tumor size compared with baseline that was not confirmed and was therefore categorized as stable disease. Two patients were not included in the plot (patient 24 in the vorasidenib group and patient 22 in the ivosidenib group) owing to changes in selected target lesions measured at one or more postoperative time points that affected best percentage change assessment. aComplete response, PR or mR. b, Postoperative treatment duration by treatment group. The histology at the time of initial diagnosis and the WHO grade based on pathology of the on-study resected tumor are provided for six patients who achieved a PR after postoperative treatment with vorasidenib (n = 2) or ivosidenib (n = 4), according to the investigator’s assessment of response using RANO-LGG. bThe tumor grade for this patient changed from grade 2 at screening to grade 3 at surgery. c, On-treatment DNA hydroxymethylation levels in the tumor regression (n = 12) and tumor growth (n = 9) groups. Box plot: the horizontal lines denote the 25th, 50th and 75th percentiles and the upper/lower whiskers extend from the hinge to the largest/smallest value that is within 1.5 × the interquartile range from the hinge; the two-sided P value was generated using Student’s t-test. d, Modulation of selected molecular pathways comparing the tumor regression and tumor growth groups; P values adjusted for multiplicity as described in Methods. A, astrocytoma; O, oligodendroglioma; SPD, sum of products of tumor lesion diameters.

Full size image

At the time of manuscript submission, fewer than half the 32 patients who remained on treatment as of the data cutoff date (29 April 2020) have discontinued treatment and the progression-free survival data continue to mature. Preliminary Kaplan–Meier curves of progression-free survival are shown in Extended Data Fig. 8.

We also examined the relationship between the molecular changes in on-treatment biopsies and subsequent radiographic response. This exploratory post-hoc analysis included a manual consensus review of all MR images by multiple investigators and excluded patients whose tumors neither grew during drug treatment nor could be assessed for tumor regression owing to a previous complete tumor resection (Extended Data Fig. 9). Tumor regression was associated with high tumor DNA 5hmC content (Fig. 3c) and reduced expression of cell-cycle-associated genes (Fig. 3d, Extended Data Fig. 10 and Supplementary Table 6) in the on-treatment biopsies.

Discussion

IDH mutations are believed to play a prominent role in the development of glioma, but the role of the mutant enzyme in recurrent glioma is currently unclear. Our data demonstrate that the mIDH enzyme remains active in recurrent disease, because 2-HG reduction was associated with reduced tumor cell proliferation, increased DNA 5hmC content (mediated by TET 5mC hydroxylase activity)23 and a reversal of gene expression programs typically associated with IDH mutations in LGGs. Although a net loss of malignant glioma cells might explain a loss of proneural transcripts, this would not account for the specific upregulation of mesenchymal transcripts and increase in cellular differentiation markers that we observed.

Our results point toward opportunities for combination therapies with IDH inhibitors in glioma. The association between radiographic tumor response and increased DNA 5hmC content, for example, suggests that DNA methylation is an important contributor to tumor maintenance by the mIDH enzyme in glioma, reminiscent of the contribution of TET enzymes in mIDH-associated tumorigenesis in other cancers28. This raises the hypothesis that DNA-hypomethylating agents, such as the DNA methyltransferase inhibitor azacitidine, might augment the antitumor activity of IDH inhibitors, as has recently been shown in acute myeloid leukemia29,30. Our observation that 2-HG reduction was associated with the induction of genes associated with antitumor immunity and a modest increase in tumor infiltration with CD8+ T cells is consistent with an immunosuppressive effect of 2-HG on the glioma microenvironment, as suggested by previous studies31,32, and raises the intriguing possibility that inhibition of the mutant enzyme might synergize with other strategies to promote antitumor immunity against mIDH gliomas.

Surgical window-of-opportunity trials typically rely on the comparison of matched pre-treatment and on-treatment biopsy pairs from the same patients. It was not feasible for patients enrolled in our current study to undergo a pre-treatment biopsy because we could not ensure that sufficient tumor tissue could be collected through a needle biopsy for 2-HG analysis and key related molecular studies (for example, RNA-sequencing (RNA-seq) or 5hmC analyses) and complications from this procedure could have resulted in a delay of the planned tumor resection. Our comparisons of immunohistochemistry (IHC) results (for example, Ki-67, immune cell infiltration) between on-study tumor tissue and tumor tissue collected during the original diagnostic tumor resection have to be interpreted with caution because many patients received chemotherapy or radiation in the interval between the two surgeries, and these tumor samples were often collected and stored under different conditions. Given these limitations, we included a large number of external untreated control tumors with a borrowing methodology in our primary analysis. The integration of internal and external controls in our clinical trial design may serve as a template for future side-by-side comparisons of investigational agents with similar mechanisms of action for brain tumor patients.

Preliminary indications of clinical activity in our study must be interpreted with caution because the progression-free survival data continue to mature and follow-up time is short for LGGs. A daily dose of vorasidenib 50 mg showed the most consistent inhibition of the mutant enzyme and the greatest preliminary antitumor activity. Based on the current data, vorasidenib (50 mg q.d. of the uncoated tablet formulation used in the present study) was selected for the initiation of the ongoing global phase 3 INDIGO study in grade 2 mIDH nonenhancing glioma (ClinicalTrials.gov, NCT04164901; a coated-tablet formulation was later introduced into the INDIGO study at a dose exposure equivalent of 40 mg q.d.). Of note, ivosidenib showed considerably lower CNS penetration than vorasidenib, in line with preclinical studies, but nevertheless reached adequate tumor concentrations to inhibit the mIDH enzyme in patients owing to its high plasma exposure. The latter observation highlights that even drugs with low CNS penetrance may warrant a detailed pharmacokinetic/PD evaluation before excluding them from further development for CNS tumors.

In conclusion, our study established the clinical, PD and translational rationale to select a late-stage molecule to target one of the most prevalent mutations in LGGs and provides a rich data resource to advance our understanding of the role of IDH mutations and the mechanism of action of IDH inhibitors in recurrent gliomas.

Methods

Trial design and oversight

This is a randomized, controlled, multicenter, open-label, perioperative study of vorasidenib and ivosidenib in recurrent, nonenhancing mIDH1 LGGs (ClinicalTrials.gov, NCT03343197). The randomization schedule was generated by an independent statistical group and randomization assignment was implemented by an interactive web response system. Additional details are provided in the study protocol and statistical analysis plan.

The study was conducted according to the International Council on Harmonisation of Good Clinical Practice guidelines and the principles of the Declaration of Helsinki. All patients provided written informed consent before screening and enrollment.

Patients

Key eligibility criteria included patients with recurrent, CNS WHO 2016 grade 2/3, mIDH1-R132H oligodendroglioma or astrocytoma who were surgical candidates. Additional eligibility criteria included age ≥18 years, adequate hepatic and renal function, a Karnofsky performance status score ≥60%, no previous IDH inhibitor treatment, ≥6 months since any radiation and measurable nonenhancing lesion by central radiology review. Patients were recruited by the authors at Memorial Sloan Kettering Cancer Center (by I.K.M., n = 13), University of California, Los Angeles Medical Center (by T.F.C., n = 10), University of Texas Southwestern (by E.A.M., n = 8), University of California, San Francisco Division of Neuro-Oncology (by J.W.T., n = 7), Dana Farber Cancer Institute (by P.W.Y., n = 6), Massachusetts General Hospital (by I.A.-R., n = 3) and Duke University Medical Center (by K.B.P., n = 2). The study protocol was approved by the institutional review board/independent ethics committee at each of these study locations. Participants’ sex was assigned by the site and no sex- or gender-based analyses were performed because these would have been post hoc and insufficiently powered to enable meaningful conclusions.

Outcomes

The primary end-point of 2-HG concentration in resected tumors was evaluated by comparing concentrations in patients with mIDH1 glioma treated with vorasidenib or ivosidenib with concentrations in tumors from untreated on-study patients (internal contemporaneous control) and additional tumors from untreated patients with WT IDH (n = 15) and mIDH1 (n = 61) glioma (external control using previously banked tumor samples).

Secondary end-points included safety, tumor and plasma pharmacokinetics, and preliminary clinical activity. Exploratory end-points included IDH pathway-related molecular and cellular changes that correlated with 2-HG and radiographic response, where feasible.

Safety was assessed by monitoring all AEs from the time of consent through 28 (+ ≤5) d after the last dose using the National Cancer Institute Common Terminology Criteria for Adverse Events, v.4.03 (ref. 33).

Tumor measurements were collected using a standardized MRI protocol at screening and surgery. A brain MR image was collected postoperatively as the new baseline before resuming treatment. Additional brain MR images were collected every 56 (±2) d thereafter (that is, while receiving study treatment) and at the end of treatment. Antitumor activity was assessed by the investigator using the Response Assessment in Neuro-Oncology criteria for low-grade glioma (RANO-LGG)34.

Patient sample collection

Tumor samples were collected and immediately processed for all clinical trial participants. The tissue was immediately split into two parts: the first sample was snap frozen in liquid nitrogen and the second sample was processed into an FFPE block. Pre-treatment, archival FFPE slides were also collected whenever possible and included in selected biomarker analyses (Extended Data Fig. 1). Concentrations of 2-HG and vorasidenib or ivosidenib were measured in tumor and plasma using liquid chromatography with tandem mass spectrometry (LC–MS/MS). Frozen tumor tissue was analyzed for tumor content and cellularity (hematoxylin and eosin (H&E) staining), DNA alterations, genome-wide RNA expression (RNA-seq) and DNA methylation. FFPE tumor tissue was analyzed for tumor content (H&E staining), IHC and expression profiling of selected genes (Extended Data Fig. 1). Patients who missed two or more doses in two weeks before surgery, or for whom no or inadequate mIDH-containing tumor tissue was received, were replaced in the 2-HG analysis.

DNA-seq

DNA-seq of archival FFPE slides and frozen surgical samples was performed by next-generation sequencing using the ACE Extended Cancer Panel (Personalis).

DNA methylation profiling

To determine the levels of 5hmC, DNA extracted from frozen surgical samples were digested with DNA degradase (Zymo Research) to generate single nucleosides. LC–MS/MS was used to quantify 5hmC, 5mC and cytosine (C). The percentage of 5hmC was calculated as the ratio 5hmC:C.

RNA expression profiling

Transcriptional profiling (RNA-seq) of frozen surgical samples was conducted using the ACE Research Transcriptome assay (Personalis). Paired-end reads in FASTQ format were aligned to the human genome (GRCh38, release 85 (ref. 35)) with HISAT, v.2.0.5 (ref. 36). SAM-to-BAM conversion and sorting were performed using Samtools v.1.4 (ref. 37). Transcript assembly with RefSeq annotation in GTF format and gene abundance estimation were carried out using StringTie v.1.3.3b and the built-in prep_DE.py Python script38, producing gene-level raw count expression values as well as transcripts per million (that is, counts corrected for gene length and sequencing depth). All subsequent analyses were conducted in the R environment v.4.1 (ref. 39).

To evaluate the association between gene expression and 2-HG concentrations, differential expression analysis was conducted on raw count expression values for a complete set of 33,121 genes. A negative binomial generalized linear model, as implemented in the DESeq2 v.1.24.0R package40, was fit to identify genes with an expression associated with z-scored log10(transformed 2-HG levels), after correcting for histology and treatment status. P values were adjusted for multiple testing using the false discovery rate (FDR)/Benjamini–Hochberg method41. For pathway enrichment analysis, genes were first sorted by the significance and the direction of their association with 2-HG, according to the formula:

$${\mathrm{rank}} = -{\mathrm{log}}_{10}\left( {{\mathrm{unadjusted}}\,P\,{\mathrm{value}}} \right) \times {\mathrm{sign}}\left( {{\mathrm{log}}_2({\mathrm{FC}})} \right)$$

where log2(FC) represents log2(transformed moderated fold-changes)42. The ranking was then used as input in gene set enrichment analysis (GSEA)43,44. Enrichment scores were calculated using the fast GSEA (FGSEA) v.1.12.0R package45 against MSigDB curated gene set (C2) and Hallmark pathways (H), with gene-set sizes ranging from 15 to 500, using 1,000 permutations. Pathways with a positive normalized enrichment score (NES) contained genes downregulated with the suppression of 2-HG. Pathways with a negative NES contained genes upregulated with the suppression of 2-HG.

To visually demonstrate the relationship between gene expression and 2-HG concentration, a heatmap of z-scored variance stabilization-transformed (VST) values was generated for the differentially expressed genes belonging to select top-scoring, significantly enriched pathways using pheatmap v.1.0.12 (ref. 46). Complete linkage hierarchical clustering method with Euclidean distance was used to cluster the genes in rows and cluster the samples in columns; the two topmost column clusters were considered to represent patients with high 2-HG and decreased gene expression and patients with low 2-HG and increased gene expression.

To identify genes that are associated with IDH mutations in LGGs, we used HTSeq count data from 413 WT and 94 mIDH LGG primary tumor samples from the Cancer Genome Atlas-LGG transcriptome profiling dataset47,48. To remove genes expressed at low levels, we converted raw count expression values to counts per million (c.p.m.), using the edgeR v.26.4 (ref. 49) R package to account for sequencing depth. We retained genes with c.p.m. ≥ 0.76 in ≥94 samples. Of 56,404 genes, 18,416 passed the filter and were used in subsequent analyses. Gene expression analysis was conducted on raw count expression values. We used the DESeq2 package to fit a negative binomial generalized linear model to identify genes expressed differentially in mIDH and WT samples, after correcting for 1p19q codeletion status. P values were adjusted for multiple comparisons using FDR.

To identify genes differentially expressed between samples from patients whose tumors responded to treatment (tumor regression) and those whose tumors did not (continued tumor growth), gene expression analysis was conducted on raw count expression values for a complete set of 33,121 genes. We used DESeq2 v.1.24.0 to fit a negative binomial generalized linear model, correcting for histology. P values were FDR corrected. For pathway enrichment analysis, genes were first sorted by the significance and direction of their association with tumor response. The ranking was then used to calculate enrichment scores in the FGSEA package45 against MSigDB Hallmark pathways (H) using the same parameters as detailed above. Pathways with a positive NES contained genes upregulated with tumor regression. Pathways with a negative NES contained genes downregulated with tumor regression.

To visually demonstrate the relationship between gene expression and tumor response, a heatmap of z-scored VST values was generated for the differentially expressed genes belonging to select significantly enriched pathways using the pheatmap v.1.0.12 (ref. 46) R package. Complete linkage hierarchical (semisupervised) clustering method with Euclidean distance was used to cluster samples in columns (in the two tumor response groups separately) and cluster genes in rows.

NanoString gene expression assay was performed on RNA extracted from FFPE tumor biopsies after macrodissection to enrich tumor content and then scanned using the nCounter Digital Analyzer as per the manufacturer’s instructions (NanoString Technologies). Gene expression was analyzed using nSolver software 4.0 (NanoString Technologies) and the expression levels of each gene were normalized to those of control genes.

Immunohistochemistry

IHC for Ki-67, CD3 and CD8 was performed by Mosaic Laboratories and quantification derived from an annotation, including all tumor and intervening stroma in the tumor nest. CD3 (mouse clone LN10) immunoglobulin (Ig)G1 antibody (catalog no. NCL-L-CD3-565) was purchased from Leica Biosystems. The CD8 (mouse clone C8/144B) IgG1κ antibody (catalog no. M7103) and the Ki-67 (mouse clone MIB-1) IgG1κ antibody (catalog no. M7240) were purchased from Dako. Antibodies were diluted per the manufacturer’s instructions. All antibodies were stored at 2–8 °C.

Statistical analysis

Unless otherwise specified, graphs and statistical analyses were performed using GraphPad Prism.

All randomized patients with mIDH1-R132H glioma, as confirmed by IHC with an antibody specific to mIDH1-R132H or DNA-seq, were included in the primary end-point analysis. A Bayesian hierarchical normal model50 was used to compare 2-HG concentrations, on a log10 scale, in evaluable treated and untreated control tumors. The model was used to dynamically borrow 2-HG concentration from externally banked, untreated, frozen control samples and the enrolled, untreated control patients. Independent, noninformative, normal distributions and inverse-gamma distributions were used for scale and variance parameters, respectively. The Markov Chain Monte Carlo Gibbs sampling51 was used to estimate the posterior distributions of unknown parameters. The posterior mean and 95% CrI of the treatment effect on the 2-HG percentage reduction relative to untreated control groups were provided for each treatment arm. The sample size was determined using extensive simulations on Bayesian analysis for the primary end-point; the sample size achieves approximately 94% probability of detecting that the 2-HG concentrations of the treated group are less than those of the untreated group.

The safety analysis set comprised randomized patients who received at least one dose of vorasidenib or ivosidenib either pre- or postoperatively. Baseline disease characteristics and safety data were summarized by treatment and dose.

Concentrations of vorasidenib and ivosidenib in tumor and plasma were reported as geometric mean and tumor:plasma ratios summarized by treatment and dose. Association of 2-HG with 5hmC, Ki-67 and CD3+/CD8+ T cells was assessed using simple linear regression. Student’s t-test was used to compare two groups. All reported P values are two sided.

Reporting summary

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