Disabling Uncompetitive Inhibition of Oncogenic IDH Mutations Drives Acquired Resistance

Cytosolic isocitrate dehydrogenase (IDH) 1 and mitochondrial IDH2 catalyze the oxidative decarboxylation of isocitrate (ICT) to α-ketoglutarate (αKG) with the concomitant production of nicotinamide adenine dinucleotide phosphate (NADPH). Somatic mutations in the catalytic arginine residues of IDH occur frequently in acute myeloid leukemia (AML) and other cancers (1–3) through gain-of-function activity by catalyzing NADPH-dependent αKG reduction to the oncometabolite (R)-2-hydroxyglutarate (R-2HG, or 2HG; refs. 4–6). Accumulating R-2HG causes altered histone and DNA methylation by impairing αKG-dependent dioxygenases (7–10). Allosteric inhibitors of mutant IDH suppress R-2HG production and induce differentiation of leukemic blasts, providing clinical responses in approximately 40% of treated AML patients with mutant IDH (11–13). These findings provide the basis of targeted therapies for AML (ivosidenib and enasidenib), establishing the first example of personalized therapy based on cancer metabolism. However, the development of treatment resistance to IDH inhibitors emerges as a new challenge. Most patients who initially responded eventually relapsed, whereas a thorough analysis of the mutational landscape and mechanisms for acquired resistance to IDH inhibition has been difficult.

Small-molecule inhibitors comprise about half of cancer-targeting drugs, which commonly involve the competitive, noncompetitive, or uncompetitive mechanism of inhibition (14). Competitive inhibitors resemble the normal substrates and thus prevent substrate binding to the enzymes, whereas noncompetitive inhibitors bind equally well to the enzyme and the enzyme–substrate complexes to impact enzymatic activity. By contrast, uncompetitive inhibitors bind only to the complex formed between the enzyme and the substrate, thus requiring the formation of a stable enzyme–substrate complex prior to inhibitor binding. It remains elusive whether and how the unique biochemical features of IDH inhibitors contribute to the development of clinical resistance.

Studies of small cohorts revealed a subset of acquired mutations at the IDH dimer interface that restore R-2HG production by interference with allosteric inhibitor binding (15). The emergence of mutations in the IDH homolog during the inhibition of the other mutant IDH protein, or isoform switching, also contributes to resistance by R-2HG restoration (16). These findings underscore a pivotal role for maintaining R-2HG production in IDH-mutant (IDH^mut) malignancies and establish R-2HG–restoring mutations as a crucial mechanism in therapy resistance. Although mutations conferring resistance to IDH inhibition are increasingly identified (17, 18), the entire repertoire of IDH second-site mutations with the potential to induce therapy resistance has not been systematically evaluated. Moreover, our understanding of the mechanisms for R-2HG–restoring mutations remains incomplete. This is due in large part to the lack of preclinical cell models and an unbiased analysis of clinically relevant IDH second-site mutations.

Here we developed a panel of isogenic leukemia cell lines containing the most common IDH1 or IDH2 oncogenic mutations by CRISPR base editing. We performed saturation mutagenesis screens of IDH single amino acid variants and identified second-site mutations conferring resistance to IDH inhibition in base-edited leukemia cells. We integrated these results with targeted sequencing of patients with AML, structural modeling, enzymology, and functional validation in humanized mouse models. Our findings not only validate known mutations associated with resistance to IDH inhibition but also uncover a new class of pathogenic mutations and the underlying mechanisms for resistance to IDH-targeting cancer therapies.

To survey the prevalence of R-2HG–restoring mutations in human AML, we evaluated 31 relapsed or refractory IDH^mut AML cases previously described to have sustained R-2HG despite treatment with mutant IDH1 or IDH2 inhibitors (15–19). We identified 16 of 31 (or 51.6%) cases of isoform switching, 13 of 31 (41.9%) cases of second-site mutations, and 2 of 31 (6.5%) cases of both mutation types (Fig. 1A; Supplementary Table S1). Importantly, among the IDH second-site mutations, 40% (6 of 15) are known to affect the allosteric binding of IDH^mut inhibitors to the dimer interface including IDH1^S280F, IDH2^Q316E, and IDH2^I319M (15), whereas the mechanisms for the other mutations are unknown (Fig. 1A), illustrating a deficiency in our current understanding of therapy resistance to IDH inhibition.

Due to the lack of leukemia cell models harboring endogenous IDH mutations, prior studies of mutant IDH have relied on overexpressing IDH^mut proteins in cells with wild-type IDH1 and IDH2 genes (20–22), which may not fully recapitulate the molecular events and therapy response driven by mutant IDH. To address this, we used CRISPR base editing to engineer clinically relevant isogenic cell models with monoallelic IDH1 or IDH2 mutations. The GM-CSF–dependent human TF-1 erythroleukemia cells were previously used as a model to evaluate the biological function and allosteric inhibition of IDH^mut proteins (12, 22, 23). We used the optimized CRISPR base editor (Cas9-BE) with the fusion of Cas9n (Cas9^D10A), cytidine deaminase, and a uracil glycosylase inhibitor (UGI) to introduce C•G-to-T•A transitions (24, 25) using single-guide RNAs (sgRNA) designed to target individual IDH hotspot mutations including IDH1^R132H, IDH2^R140Q, IDH2^R140W, and IDH2^R172K, respectively (Fig. 1B; Supplementary Fig. S1A). Upon base editing, we screened single cell–derived clonal lines and confirmed monoallelic IDH1^WT/mut or IDH2^WT/mut genotypes by targeted sequencing of genomic DNA (Supplementary Fig. S1B; Supplementary Table S2).

To assess the functional impact of base-edited IDH^mut genes in leukemia cells, we first determined R-2HG levels by mass spectrometry in the representative IDH1^R132H (clone D5, hereafter IDH1^R132H-BE) and IDH2^R140Q (clone B4, hereafter IDH2^R140Q-BE) cells. Compared with unedited parental TF-1 cells, R-2HG levels were significantly increased in the pellets and medium of IDH1^R132H-BE and IDH2^R140Q-BE cells, similar to that observed in TF-1 cells stably overexpressing IDH1^R132H or IDH2^R140Q (hereafter IDH1^R132H-OE and IDH2^R140Q-OE; Fig. 1C), illustrating the gain-of-function activity of base-edited IDH^mut proteins. Moreover, IDH1^R132H-BE and IDH2^R140Q-BE cells acquired GM-CSF cytokine-independent cell growth and resistance to EPO-induced differentiation to fetal hemoglobin (HbF)–positive erythroid cells, consistent with the role of oncogenic IDH mutations in promoting “stemness” by blocking cell differentiation (refs. 7, 9, 22, 23, 26; Fig. 1D and E).

Mutant IDH causes global epigenetic alterations by impairing histone and DNA demethylases (7–10). We observed increased histone methylation and 5-methylcytosine (5mC) and decreased 5-hydroxymethylcytosine (5hmC) in base-edited IDH1^mut and IDH2^mut TF-1 cells (Supplementary Fig. S1C and S1D). We also generated base-edited IDH1^WT/mut or IDH2^WT/mut K562 and MOLM-13 clonal leukemia cells and confirmed the significant increases of R-2HG (Supplementary Fig. S1B and S1E). These results demonstrate that the biological effects caused by base-edited IDH^mut are not dependent on specific cellular contexts or genetic background. Together, we created a panel of clinically relevant IDH1 or IDH2 monoallelic mutations in 24 independent clonal cell lines that faithfully recapitulate the naturally occurring heterozygous IDH^mut cancer cells, thus establishing the preclinical cell models for dissecting the biological consequences and determinants of therapy response to IDH inhibition.

Having validated the base-edited IDH1^R132H and IDH2^R140Q leukemia cells, we designed orthogonal mutagenesis screens to identify IDH second-site mutations capable of conferring resistance to IDH^mut inhibitors. We reasoned that if IDH^mut cells acquire R-2HG–restoring mutations upon treatment with IDH^mut inhibitors (AG-120 or ivosidenib for IDH1^R132H; AG-221 or enasidenib for IDH2^R140Q), the cells would be resistant to EPO-induced differentiation and progressively enriched in undifferentiated HbF-negative cell populations. To this end, the IDH2^R140Q-BE cells were treated with 1 μmol/L AG-221 continuously for 16 weeks, followed by EPO-induced differentiation for 8 days (2 U/mL) before FACS-sorting of HbF-negative cells. Targeted sequencing of the complete protein-coding sequences of the endogenous IDH2 gene was performed in genomic DNA isolated in cells after short-term (6 weeks) or long-term (16 weeks) AG-221 treatment (Fig. 2A). From two replicate screens, we identified 109 enriched second-site mutations at 91 amino acid positions (fold change ≥2 in cells at 16 relative to 6 weeks) clustered at the small domains and the C-terminal large domain of IDH2 (Fig. 2B; Supplementary Fig. S2A; Supplementary Table S3).

The AG-221 selection screen uncovered candidate second-site mutations associated with acquired resistance but did not provide a complete assessment of all possible variants due to low frequency of spontaneous mutations. We next designed the second screen using a saturation variant library consisting of full-length IDH2 cDNAs harboring amino acid substitutions. We validated the presence of 7,742 individual IDH2-mutant cDNAs in which each amino acid at positions 41 to 452 (excluding mitochondrial localization signal and stop codon) was replaced by one of the other 19 amino acids (Fig. 2C; Supplementary Table S4). The saturation variant library was then cloned into a lentiviral vector and transduced into IDH2^R140Q-BE cells at a multiplicity of infection (MOI) ≤ 0.3, such that the vast majority of cells received only one mutant cDNA. The transduced cells were selected, treated with 1 μmol/L AG-221 (0 to 10 weeks), and induced by EPO for 8 days before harvesting genomic DNA from HbF-negative cells. By targeted sequencing of IDH2-mutant cDNA in cells after AG-221 treatment (0, 2, 4, 6, 8, and 10 weeks or T0 to T5, respectively; Fig. 2C), we identified 245 progressively enriched second-site mutations at 187 amino acid positions (fold change ≥2 in T5 relative to T0) from two replicate screens (Fig. 2D; Supplementary Fig. S2B; Supplementary Table S5).

Together, second-site mutations at 43 amino acid positions in IDH2 were consistently identified in AG-221 selection and saturation variant screens. The identification of known mutations at the dimer interface (IDH2^Q316E and IDH2^I319M) provides support for the performance of both screens (ref. 15; Fig. 2D and E). More importantly, by mapping these positions on the structure of inhibitor-bound IDH2^R140Q homodimer (12), we uncovered a previously unrecognized class of mutations at residues N136, E343, E345, A347, H348, V351, T352, and R353, which physically surround and interact with NADPH cofactor at the enzyme active site via a series of hydrogen bonding, hydrophobic, electrostatic, and/or Van Der Waals interactions (Fig. 2E). For example, the carboxyl group of residue E343 forms sodium ion and E345-mediated indirect interactions with the amino group of the reduced nicotinamide moiety of NADPH, whereas the main chain atoms of residue A347 contact the same chemical group of NADPH through solvent-mediated hydrogen bonds (Fig. 2E). Moreover, the enrichment of mutations at the NADPH binding sites was comparable with or higher than the enrichment of mutations at the dimer interface (Fig. 2F).

To assess the clinical relevance, we analyzed a cohort of 24 relapsed or refractory AML cases with IDH1 or IDH2 baseline mutations who developed resistance after treatment with IDH inhibitors (Supplementary Fig. S2C; Supplementary Table S6). By targeted sequencing of the complete protein-coding sequences of IDH1 and IDH2 genes in matched pre- and posttreatment bone marrow specimens, we identified seven cases with IDH1^R132C, IDH1^R132H, or IDH2^R140Q baseline mutations at diagnosis that acquired R-2HG–restoring mutations in relapse samples. These include one case of a dimer interface mutation (IDH2^I319M in patient UT6) and two cases of isoform-switching mutations (IDH1^R132C to IDH2^R140Q in UT5 and IDH1^R132H to IDH2^R140Q in UT9; Supplementary Fig. S2D–S2F), consistent with the role of these mutations in driving acquired resistance to IDH inhibition (15, 16).

Importantly, we identified three cases of IDH2 second-site mutations at the NADPH binding sites, including A347T in UT8, V351I in MDA2, and E343V in MDA7 (Fig. 2G–I; Supplementary Fig. S2G), which were among the top candidates identified in AG-221 selection or saturation variant screens (Fig. 2A–F). Structural analysis indicates that the side chain packing of residue A347 is changed by this mutation such that residues A347 and H348 move away from their locations in the wild-type conformation to attenuate NADPH association (Fig. 2G). Importantly, the other mutations (N136S, E343V, E345G, H348Q, V351I, T352A, and R353H) were also predicted to impact NADPH binding by disrupting hydrogen bond, electrostatic, and/or Van Der Waals interactions (Fig. 2H and I; Supplementary Fig. S2H). Moreover, we analyzed longitudinal samples of UT8, MDA2, and MDA7 harboring NADPH binding site mutations. We observed progressive increases in variant allele frequencies (VAF) across posttreatment samples, indicating a positive selection of second-site mutations in response to IDH inhibition (Fig. 2G–I). Besides these mutations, N136S was previously described in an AML case with IDH2^R172K baseline mutation (27).

We next performed PCR and Sanger sequencing using genomic DNA and confirmed the cis configuration for IDH2^WT + IDH2^R140Q+A347T in UT8 and IDH2^WT + IDH2^R140Q+E343V in MDA7 but the trans configuration for IDH2^R140Q + IDH2^V351I in MDA2, respectively (Fig. 2J). To determine whether the same mutation can occur in cis or trans, we analyzed the mutation patterns in cDNAs of base-edited leukemia cells from the AG-221 selection screens by amplicon sequencing (Supplementary Fig. S2I; Supplementary Table S7). We found that A347T, V351I, and E343V were detected in both cis and trans configurations in different sequencing reads (Supplementary Fig. S2J).

To corroborate these findings, we performed similar screens in IDH1^R132H-BE cells after treatment with 1 μmol/L AG-120 for 16 weeks (Supplementary Fig. S3A). We identified IDH1^S280F, a mutation known to affect allosteric binding of AG-120 at the dimer interface (15, 17), as one of the top candidates (Supplementary Fig. S3B and S3C; Supplementary Table S8). More importantly, by comparing the acquired mutations from AG-120 inhibitor screens with R-2HG–restoring mutations identified in IDH1-mutant AML patients treated with AG-120 (Supplementary Tables S1 and S6; ref. 17), we not only confirmed the recurrent dimer interface mutation at IDH1^S280F (15, 17) but also uncovered two NADPH binding–associated mutations at IDH1^T313I and IDH1^H315D, respectively (Supplementary Fig. S3D). IDH1^T313I displayed progressive increases of VAFs in patient MDA4 following AG-120 treatment (Supplementary Fig. S3E), whereas IDH1^H315D was detected in a relapsed AML case with VAF around 30% and associated with restored R-2HG in the relapsed samples following AG-120 monotherapy (17). These mutations are predicted to weaken NADPH binding by disturbing hydrogen bond interactions based on the structure of inhibitor-bound IDH1^R132H (PDB #5L57; ref. 28; Supplementary Fig. S3F) and were detected in both cis and trans configurations in AG-120–treated TF-1 cells (Supplementary Fig. S3G). Therefore, our orthogonal screens in base-edited IDH^mut leukemia cells revealed a compendium of clinically relevant second-site mutations, enabling a detailed dissection of the underlying mechanisms for acquired resistance to IDH inhibition.

The identification of recurrent second-site mutations affecting NADPH cofactor binding raises the question about their effects on R-2HG production and the catalytic function of IDH^mut proteins. To address this, we generated TF-1 cells stably expressing IDH2^R140Q with or without each second-site mutation and measured R-2HG levels after treatment with control (DMSO) or 1 μmol/L AG-221 for 3 days. AG-221 treatment significantly decreased R-2HG in IDH2^R140Q-expressing cells without second-site mutation, consistent with the effective inhibition of IDH2^mut-catalyzed R-2HG production (4–6). By contrast, expression of IDH2^R140Q harboring second-site mutation associated with NADPH binding blocked AG-221–mediated inhibition, resulting in restored R-2HG similar to that observed in DMSO-treated cells or cells with the dimer interface mutations (Fig. 3A). Similarly, the IDH1^T313I and IDH1^H315D mutations associated with NADPH binding also blocked AG-120–mediated inhibition in IDH1^R132H-expressing TF-1 cells, causing R-2HG accumulation to the level observed with IDH1^S280F at the dimer interface (Fig. 3B).

As mutant IDH catalyzes the reduction of αKG to R-2HG by oxidizing NADPH (4, 5), we determined the effect of second-site mutations on IDH catalytic activity by measuring NADPH consumption in IDH2^R140Q-expressing TF-1 cells (Fig. 3C). Of note, AG-221 effectively inhibited the catalytic activity of IDH2^R140Q, resulting in significantly decreased NADPH consumption in IDH2^R140Q-expressing cells. By contrast, the presence of second-site mutations at NADPH binding sites resulted in similar rates of NADPH consumption as observed with dimer interface mutations (IDH2^Q316E and IDH2^I319M) or DMSO-treated cells (Fig. 3C; Supplementary Fig. S4A–S4H). Similar results were obtained for IDH1^T313I, IDH1S280F, or IDH1^H315D in IDH1^R132H-expressing TF-1 cells (⁰; Supplementary Fig. S4K and S4L). The mutation of residues not associated with NADPH binding did not affect IDH^mut activity or R-2HG production in the presence of inhibitors (Supplementary Figs. S4I–S4N and S5A and S5B). Moreover, these second-site mutations alone did not induce R-2HG production and had no effect on cytokine-dependent cell growth or EPO-induced HbF expression (Supplementary Fig. S5C–S5F).

To further evaluate whether the identified mutations promote therapy resistance in vivo, we xenografted IDH2^R140Q-expressing TF-1 cells with or without second-site mutations into the NSG-SGM3 (or NSGS) transgenic mice expressing human IL3, GM-CSF, and SCF that support the stable engraftment of human myeloid leukemia cells (29). These mice were engrafted with mCherry and luciferase-marked TF-1 cells expressing IDH2^R140Q alone or with second-site mutation (IDH2^E343V or IDH2^A347T). After disease onset, the mice were treated with vehicle or 30 mg/kg AG-221 and monitored for leukemia development (Fig. 3D). AG-221 treatment impaired the propagation of leukemia cells, resulting in less tumor burden, fewer leukemic cells in peripheral blood and bone marrow, and prolonged survival compared with vehicle-treated controls. By contrast, the presence of a second-site mutation rendered IDH2^R140Q-expressing cells resistant to AG-221 inhibition, resulting in leukemia burden comparable with vehicle-treated controls (Fig. 3E–H). The presence of NADPH binding–associated IDH1^T313I or IDH1^H315D in IDH1^R132H-expressing TF-1 cells also induced resistance to AG-120 treatment in xenotransplanted mice (Supplementary Fig. S6A–S6E). Collectively, these results establish a functional role for NADPH binding site–associated mutations in restoring R-2HG production to drive resistance to IDH inhibition.

Having validated the in vivo efficacy of NADPH binding–associated mutations in conferring therapy resistance, we determined the extent to which the second-site mutation affects inhibitor function. To this end, we first determined R-2HG production in TF-1 cells expressing IDH2^R140Q with or without A347T second-site mutation in the presence of AG-221 (10⁻⁵ to 100 μmol/L). AG-221 impaired R-2HG production in cells with IDH2^R140Q alone starting at 1 nmol/L with maximal inhibition at ≥1 μmol/L concentration. By contrast, cells coexpressing IDH2^R140Q and A347T increased the threshold of AG-221 inhibition to 50 μmol/L or higher concentration (Fig. 4A). We next measured the enzyme activity of IDH2^R140Q with or without A347T in the presence of AG-221 (10⁻⁵ to 100 μmol/L) in TF-1 cells. AG-221 effectively blocked IDH2^R140Q activity at 0.1 to 1 nmol/L with maximal inhibition at 0.1 to 1 μmol/L concentration (Fig. 4B). By contrast, IDH2^R140Q with A347T significantly increased the threshold of AG-221 inhibition. These studies indicate that the NADPH binding site mutations increase the threshold of inhibitor binding, thus requiring higher inhibitor concentrations to block IDH^mut-catalyzed R-2HG production. Moreover, the NADPH binding site mutations also restored IDH^mut-mediated R-2HG production in TF-1 cells treated with other IDH^mut inhibitors including AG-881 (vorasidenib, ref. 30; IDH-305, ref. 31; or BAY-1436032, ref. 32; Fig. 4C and D; Supplementary Fig. S6F and S6G).

Wild-type IDH enzymes convert ICT to αKG and generate NADPH, whereas oncogenic IDH^mut proteins gain the activity of reducing αKG to R-2HG and convert NADPH to NADP⁺. Consistent with the neomorphic activity of IDH^mut enzymes, IDH1^R132H- and IDH2^R140Q-expressing cells significantly decreased intracellular NADPH and increased NADP⁺ in TF-1 cells (Fig. 4E). IDH1^mut and IDH2^mut inhibitors restored NADPH levels in IDH1^R132H- and IDH2^R140Q-expressing cells, respectively, indicating effective blocking of IDH^mut enzyme activity (Fig. 4F). However, NADPH levels remained low in inhibitor-treated TF-1 cells coexpressing IDH1^R132H with IDH1^H315D (or IDH2^R140Q with IDH2^A347T) second-site mutation (Fig. 4F). These findings demonstrate that secondary mutations at NADPH binding sites restore the neomorphic activity of IDH^mut enzymes in the presence of IDH inhibitors.

Both AG-120 and AG-221 show slow–tight binding inhibition to IDH1^R132H and IDH2^R140Q, respectively (12, 33). The binding of IDH inhibitors to the dimer interface of IDH^mut homodimers stabilizes the inactive open conformation and prevents conformational change to block IDH^mut-catalyzed R-2HG production (12, 22, 34). Moreover, AG-221 is an uncompetitive inhibitor (Fig. 5A) with regard to NADPH and NADP⁺ based on the studies using IDH2^mut and IDH2^WT homodimers, respectively (12). Cancer-associated IDH mutations occur in a single allele, resulting in the formation of a mixture of IDH^WT/mut heterodimers and IDH^mut homodimers. As the IDH^WT/mut heterodimers produce R-2HG more efficiently than IDH^mut homodimers (35), they are considered the major molecular targets for therapeutic inhibition. Despite these findings, previous biochemical and structural studies were based on IDH^mut homodimers (12, 22, 33, 34), whereas the mode of inhibition for IDH^WT/mut heterodimers has not been evaluated.

To this end, we tested whether IDH inhibitors exhibit uncompetitive inhibition with regard to NADPH using purified IDH1^WT/R132H and IDH2^WT/R140Q heterodimers. By titration series of substrates versus fixed concentrations of inhibitors, we observed that the linear regression fits of the reciprocal plots yielded parallel lines for both IDH1^WT/R132H–AG-120 and IDH2^WT/R140Q–AG-221 complexes (Fig. 5B and C), consistent with an uncompetitive model of inhibition (36). Therefore, both AG-120 and AG-221 act as uncompetitive inhibitors with regard to NADPH on the clinically relevant IDH^WT/mut heterodimers. A prerequisite for the mode of uncompetitive inhibition is that the IDH^mut–NADPH complex must be formed to allow the binding of the IDH^mut inhibitor and that inhibitor binding facilitates the formation of the stable IDH^mut–inhibitor complex (Fig. 5A).

Based on the biochemical studies, we hypothesized that the binding of NADPH is critical for the formation of stable IDH^mut–inhibitor complexes and that secondary mutations at NADPH binding sites promote acquired resistance by disabling uncompetitive enzyme inhibition. Given the more potent activity of IDH^WT/mut heterodimers in catalyzing R-2HG production (35), we focused on IDH1^WT/R132H and IDH2^WT/R140Q for subsequent studies.

We isolated IDH^WT/+mut heterodimers without (IDH^WT + IDH^mut1) or with (IDH^WT + IDH^mut1+mut2) second-site mutations by coexpressing GST-tagged IDH^WT and 6xHis-tagged IDH^mut, followed by tandem affinity purification in Escherichia coli Rosetta2 (DE3) cells (Fig. 5D). Specifically, we purified heterodimers containing IDH2^R140Q alone or with secondary mutations at the dimer interface (IDH2^R140Q+Q316E and IDH2^R140Q+I319M) or NADPH binding sites (IDH2^R140Q+E343V and IDH2^R140Q+A347T; Fig. 5E). Of note, the binding affinity (K_m) of NADPH to IDH2^R140Q+Q316E and IDH2^R140Q+I319M was comparable with IDH2^R140Q alone (Fig. 5F and G), consistent with the role of these mutations in preventing allosteric inhibitor binding to dimer interface without affecting NADPH binding at the active site. By contrast, NADPH binding affinity to IDH2^R140Q+E343V and IDH2^R140Q+A347T was lower than IDH2^R140Q alone (K_m = 5.66 ± 0.27 μmol/L and 4.64 ± 0.77 μmol/L for IDH2^R140Q+E343V and IDH2^R140Q+A347T vs. K_m = 3.51 ± 0.19 μmol/L for IDH2^R140Q; Fig. 5F and G). Moreover, the half maximal inhibitory concentration (IC₅₀) was higher for heterodimers containing dimer interface or NADPH binding site mutations than IDH2^R140Q alone (Fig. 5F and H), illustrating impaired inhibitor binding by both types of mutations. Lastly, although IDH^mut enzyme activity was impaired by AG-221–mediated inhibition of IDH2^R140Q, second-site mutations at the dimer interface or NADPH binding sites restored IDH2^R140Q activity in the presence of AG-221 (Fig. 5I). It is important to note that the modest changes in NADPH and inhibitor binding in vitro were associated with significant increases in R-2HG production in TF-1 cells, likely due to the differences in substrate and enzyme concentrations used for in vitro biochemical assays.

To corroborate these findings, we examined the effects of NADPH binding site–associated IDH1^T313I or IDH1^H315D mutation in IDH1^R132H heterodimers. IDH1^T313I or IDH1^H315D also decreased NADPH and inhibitor binding, resulting in restored IDH1^R132H activity upon AG-120 treatment (Supplementary Fig. S7A–S7E). Therefore, secondary mutations at NADPH binding sites decreased NADPH binding to prevent the formation of stable IDH^mut–inhibitor complexes, resulting in restored IDH^mut activity in the presence of inhibitors.

Acquired resistance caused by dimer interface mutations can occur in cis on the IDH^mut allele or in trans on the opposite IDH^WT allele in human patients (15). The NADPH binding site mutations were also detected in cis or trans configuration in AML patients and cells. As our functional and biochemical studies of IDH second-site mutations focused on the cis configuration, it remains unknown whether and how disabling uncompetitive inhibition confers acquired resistance by acting in trans. IDH^WT/mut heterodimer consists of an IDH^WT monomer capable of catalyzing NADP⁺-dependent ICT to αKG conversion and an IDH^mut monomer catalyzing NADPH-dependent αKG to R-2HG conversion (4). We noted that the IDH^mut inhibitors AG-120 and AG-221 also exhibited uncompetitive inhibition with regard to NADP⁺ on purified IDH1^WT/R132H and IDH2^WT/R140Q heterodimers, respectively (Fig. 6A and B), consistent with previous findings using IDH1^WT and IDH2^WT homodimers (12, 33). Based on these results, we reasoned that second-site mutations may act in trans by disabling NADP⁺-dependent uncompetitive inhibition of IDH^WT monomer to impair inhibitor binding and restore the catalytic function of IDH^WT/mut heterodimers.

To test this, we purified IDH2 heterodimers containing IDH2^R140Q alone or with second-site mutations in trans (IDH2^R140Q + IDH2^E343V or IDH2^R140Q + IDH2^A347T; Fig. 6C and D). We first noted decreased NADP⁺ binding affinity to IDH2^E343V and IDH2^A347T but not IDH2^Q316E and IDH2^I319M homodimers, consistent with impaired NADP⁺ binding due to E343V and A347T mutations (Fig. 6E and F). Moreover, we observed lower inhibitor binding affinity to IDH2^R140Q + IDH2^E343V and IDH2^R140Q + IDH2^A347T heterodimers relative to IDH2^R140Q alone, resulting in restored IDH2^R140Q activity upon AG-221 treatment (Fig. 6G–I). IDH1^T313I or IDH1^H315D in the context of IDH1^R132H also decreased NADP⁺ and inhibitor binding, causing restored enzyme activity in the presence of AG-120 (Supplementary Fig. S8A–S8F). Lastly, we used differential scanning fluorimetry (DSF) to assess the effect of IDH2^A347T on AG-221 binding to the purified IDH2^R140Q heterodimers. In line with the enzymology results, the IDH2^A347T second-site mutation, located either in cis or trans, markedly weakened AG-221 binding to IDH2^mut enzymes relative to IDH2^R140Q alone (Supplementary Fig. S9A–S9C). Hence, second-site mutations can act in cis through disabling NADPH-dependent uncompetitive inhibition or in trans through disabling NADP⁺-dependent uncompetitive inhibition of IDH^WT/mut heterodimers, resulting in restored R-2HG production to promote therapy resistance to IDH inhibition.

Collectively, by generating clinically relevant IDH^mut base-edited leukemia cells and mutagenesis screens of IDH second-site mutations at single amino acid resolution, we uncovered new mechanisms for acquired resistance to IDH inhibition by disabling uncompetitive enzyme inhibition. IDH^mut inhibitors AG-120 and AG-221 are slow–tight binders of the dimer interface on IDH1^mut and IDH2^mut enzymes, respectively. The formation of the IDH^mut–inhibitor complex stabilizes the inactive open conformation, prevents conformational change, and blocks IDH^mut-catalyzed R-2HG production (Fig. 7, models 1 and 2; Supplementary Movies S1 and S2). More importantly, AG-120 and AG-221 are NADPH- and NADP⁺-dependent uncompetitive inhibitors of IDH^WT/mut heterodimers. Second-site mutations at NADPH binding sites act in cis or trans to prevent the formation of stable enzyme–inhibitor complexes, restore mutant IDH enzyme activity in the presence of inhibitors, and drive acquired resistance to IDH inhibition (Fig. 7, models 3 and 4; Supplementary Movies S3 and S4). Hence, our findings not only identify a new class of pathogenic mutations responsible for resistance to IDH inhibition but also establish previously unrecognized mechanisms for acquired resistance to targeted cancer therapies.

Therapy resistance to allosteric IDH^mut inhibitors emerges as a new challenge; however, due to the lack of preclinical cell models harboring cancer-associated IDH mutations, the mutational repertoire and the underlying mechanisms responsible for acquired resistance to IDH inhibition have remained incompletely understood. The development of sustainable cell lines containing endogenous IDH mutations has been difficult (37). As such, prior functional and biochemical studies were conducted by overexpressing mutant IDH1 or IDH2 in cells harboring wild-type IDH genes (20–22). Hence, establishing clinically relevant cell models with monoallelic IDH mutations is a prerequisite to dissect the biological events involved in tumorigenesis and therapy response.

Here, we used CRISPR base editing to generate a panel of isogenic cell models harboring monoallelic IDH1 or IDH2 oncogenic mutations commonly found in human cancers and validated the molecular and functional effects of base-edited IDH mutations in cell and mouse models. The isogenic IDH^mut cells recapitulating the naturally occurring IDH mutations will provide valuable models for investigating IDH^mut-driven biological events. Using these models, we identified amino acid substitutions conferring resistance to IDH inhibition through saturation variant screens. By integrating these results with targeted sequencing of AML samples and functional studies, we uncovered new mechanisms for acquired resistance by disabling uncompetitive enzyme inhibition. It is important to note that, although our studies were performed in the context of myeloid leukemia, the integrative approaches and reagents should be generally applicable to other IDH^mut malignancies, including gliomas, intrahepatic cholangiocarcinomas, and chondrosarcoma, in which IDH inhibitors are being evaluated as targeted therapies (33, 38). Therefore, the development of base-edited cell models and the discovery of second-site mutations associated with acquired resistance will likely have broad implications for understanding the molecular basis of therapy resistance to IDH inhibition in human cancers.

Acquired resistance to targeted cancer therapies can arise through distinct mechanisms. Resistance to kinase inhibitors often involves second-site mutations that modulate drug binding or copy-number alterations of the mutated kinases (39–43). In IDH^mut malignancies, recurrent second-site mutations at the IDH dimer interface confer resistance by interfering with allosteric inhibitor binding (15). Isoform-switching mutations also contribute to resistance by restoring R-2HG production (16). Additional mechanisms involving the selection of non-IDH mutations or clonal evolution may contribute to primary or acquired resistance to IDH inhibition by restoring differentiation arrest (17–19, 44).

IDH^mut inhibitors ivosidenib (AG-120) and enasidenib (AG-221) represent a new class of cancer therapy due to their unique biochemical features (12, 33). Instead of competitively binding to enzyme active sites, they bind to allosteric sites enclosed within the dimer interface of IDH^mut homo- or heterodimers. Consequently, the mutant enzyme adopts an inactive open conformation that is incapable of catalyzing αKG to R-2HG conversion. As such, IDH^mut inhibitors show noncompetitive inhibition for αKG but uncompetitive inhibition against IDH^mut homo- and heterodimers for NADPH and NADP⁺ cofactors, respectively (ref. 12; Figs. 5A–C and 6A and B).

The mode of uncompetitive inhibition posits that IDH^mut inhibitors bind only to the IDH^mut–NADPH complex formed between the enzyme and the cofactor (Fig. 5A). Importantly, IDH^mut tumor cells have limited availability of NADPH due to the catalytic activity of IDH^mut enzymes (Fig. 4E and F; ref. 4). Inhibition of IDH^mut restores the NADPH level to facilitate the formation of IDH^mut–NADPH-inhibitor complexes due to uncompetitive inhibition. As such, acquired second-site mutations that decrease NADPH binding, such as the mutations identified here, may prevent the formation of stable IDH^mut–NADPH-inhibitor complexes without compromising IDH^mut-catalyzed αKG reduction to R-2HG (Fig. 7). Hence, tumor cells that acquire recurrent mutations at NADPH binding sites may be positively selected by inhibitor treatment to drive clonal evolution and therapy resistance. It is important to note that, whereas our studies focused on IDH^WT/mut heterodimers, the mechanism of disabling uncompetitive inhibition may also be relevant to IDH^mut homodimers to promote acquired resistance to IDH inhibition. In addition, we noted that the VAFs of the second-site mutations are often lower than the primary mutations. Unlike the primary mutations acquired early during clonal evolution, the second-site mutations are likely acquired or selected in subclones. Consistent with this notion, we observed progressively increased VAFs in the longitudinal samples. It is also plausible that the R-2HG–restoring mutations may drive drug resistance in subclones by exporting R-2HG to affect other cells without acquired mutations.

Uncompetitive inhibition contributes to various biological processes under physiologic and pathologic conditions. For instance, removing membrane lipids decreases the α-helix content in mitochondria, leading to changes in ATPase that resemble uncompetitive inhibition (45). The N-methyl-D-aspartate receptor (NMDAR) binds glutamate and glycine to control ion transport in response to amino acid binding in neurons. Various blockers have been developed to modify NMDAR activity such as the uncompetitive inhibitor memantine used to treat Alzheimer disease (46).

Here we describe a new class of pathogenic mutations and establish an example that uncompetitive enzyme inhibition can be subverted as a mechanism to drive therapy resistance in human cancers. Through acquisition of secondary mutations to disable uncompetitive inhibition, IDH^mut tumors leverage the specific metabolic rewiring to escape allosteric inhibition, restore mutant enzyme activity, and promote therapy resistance. Furthermore, our studies highlight the importance of thorough cataloging of drug-resistance mutations in clinically relevant cell models for developing improved therapies for IDH^mut cancers, as exemplified by the development of second- and third-generation inhibitors of BCR–ABL, EGFR, and ALK, or the identification of rational combination therapies (47). The comprehensive identification of second-site mutations with acquired resistance to IDH inhibition enables further investigation of candidate mutations and associated mechanisms, leading to new strategies for monitoring disease progression and preventing or overcoming resistance to targeted cancer therapies.

Selected clot section specimens of bone marrow biopsy used in this study were obtained from relapsed or refractory AML patients positive for IDH1 or IDH2 mutation after AG-120 or AG-221 treatment with written informed consent in accordance with the Declaration of Helsinki. The collection of these samples was approved by the Institutional Review Board (IRB) of the University of Texas Southwestern Medical Center (UTSW; IRB STU 122013-023 and IRB STU 2019-0815). We analyzed a cohort of 24 AML cases by the following selection criteria: harboring IDH1 or IDH2 baseline mutation associated with R-2HG production; treated with IDH inhibitor and achieved partial or complete response; relapsed and/or developed resistance to IDH inhibition; and both diagnosis (pretreatment) and posttreatment samples are available for the analysis of IDH mutations.

NSG-SGM3 (NSGS) transgenic mice expressing human IL3, GM-CSF (CSF2), and SCF (29) were obtained from the Jackson Laboratory (stock no. 013062). Both male and female mice were used unless otherwise specified. All mice were housed under a 12-hour light–dark cycle, 75°F, and 35% humidity in the Animal Resource Center at UTSW. All mouse experiments were performed under protocols approved by the Institutional Animal Care and Use Committee of UTSW.

Human TF-1 cells (ATCC, CRL-2003) were cultured in RPMI 1640 medium supplemented with 2 ng/mL recombinant human GM-CSF (PeproTech, 300-03). TF-1 cells with different IDH mutations were cultured without GM-CSF. Human K562 and MOLM-13 cells were cultured in RPMI 1640 medium. HEK293T cells were cultured in DMEM. To generate single cell–derived clones, FACS-sorted cells were plated in 96-well plates and screened by genotyping using targeted sequencing to identify IDH heterozygous mutations. All cell lines were cultured in the medium containing 10% FBS and 1% penicillin/streptomycin at 37°C with 5% CO₂. AG-120 (ApexBio, B7805), AG-881 (Selleck Chemicals, S8611), IDH-305 (MedChem Express, HY-104036), and BAY-1436032 (MedChem Express, HY-100020) were used to treat cells with IDH1^R132H mutation alone or with secondary mutations. AG-221 (ApexBio, B7804) and AG-881 were used to treat cells with IDH2^R140Q mutation alone or with secondary mutations. All cell lines used in this study tested negative for Mycoplasma contamination. No cell line used in this study was found in the database of commonly misidentified cell lines that is maintained by the International Cell Line Authentication Committee and the National Center for Biotechnology Information BioSample.

The luciferase cassette was cloned into the XbaI site of the pLVX-EF1α-IRES-mCherry vector (Clontech, 631987) using the In-Fusion HD Cloning Kit (Takara Bio, 638911). To generate pLVX-EF1α-IRES-TagBFP vector, the sequences of IRES and TagBFP were first fused by In-Fusion cloning. The IRES-zsGreen1 cassette of pLVX-EF1α-IRES-zsGreen1 vector (Clontech, 631982) was then replaced with IRES-TagBFP sequence by In-Fusion cloning. IDH1^WT or IDH2^WT sequence was cloned to the EcoRI and BamHI sites of the pLVX-EF1α-IRES-zsGreen1 and pLVX-EF1α-IRES-TagBFP vectors, respectively. IDH1^R132H, IDH2^R140Q, IDH2^R140W, and IDH2^R172K mutations were generated by the Q5 Site-Directed Mutagenesis Kit (New England Biolabs, E0552S) following the manufacturer's instructions and using the pLVX-EF1α-IDH1^WT-IRES-zsGreen1 and pLVX-EF1α-IDH2^WTIRES-zsGreen1 vectors as templates. Second-site mutations at IDH1 or IDH2 were generated using pLVX-EF1α-IDH1^R132H-IRES-zsGreen1 and pLVX-EF1α-IDH2^R140Q-IRES-zsGreen1 as the templates, respectively. IDH1^WT, IDH2^WT, IDH1^R132H, and IDH2^R140Q sequences were amplified and cloned to BamHI and XhoI sites of pGEX-4T-1 vector (Sigma-Aldrich, GE28-9545-49). IDH1 or IDH2 sequences with primary or second-site mutations were amplified and cloned to BamHI and HindIII sites of pET28a vector (EMD Biosciences) with an N-terminal 6xHis-SUMO tag. Primers used for cloning and generating the primary or second-site mutations are listed in Supplementary Table S9.

We used the optimized CRISPR base-editing method (25) to generate hotspot mutations including IDH1^R132H (CGT to CAT), IDH2^R140Q (CGG to CAA), IDH2^R140W (CGG to TGG), and IDH2^R172K (AGG to AAA) in TF-1, K562, and MOLM-13 human leukemia cells, respectively. The CRISPR–Cas9 base editor (Cas9-BE; Addgene, 110869) contains BE3^RA [Cas9n with codon optimization and deletion of premature poly(A) sites], cytidine deaminase APOBEC, UGI domain, and two nuclear-localization signal sequences at the N- or C-terminus. sgRNAs were designed to target individual hotspot mutations in IDH1 or IDH2 and cloned into a pSLQ1651 vector (Addgene, 100549) with mCherry. Leukemia cells were transduced with lentiviruses expressing Cas9-BE and sgRNA for each mutation, which were packaged in HEK293T cells as previously described (48), and sorted for EGFP and mCherry-positive cells. Genomic DNA (gDNA) was extracted from single cell–derived clones and used for targeted sequencing. Raw fastq files were mapped to IDH1 or IDH2 reference sequences using Bowtie2 with default parameters (49). Mapped reads were extracted from mapping files and aligned together. For each position of piled reads, the counts of bases (A, T, C, and G) and corresponding mutation frequencies for each position were calculated. The mutation frequencies at the base-edited sites in single cell–derived clones (two independent clones for each mutation) are shown in Supplementary Fig. S1B. The sequences for sgRNAs and primers are listed in Supplementary Table S9.

Parental and base-edited TF-1 cells were induced for erythroid differentiation by EPO as previously described (23, 50). Briefly, cells were washed four times with RPMI 1640 without serum and starved overnight in a medium containing 10% FBS lacking of GM-CSF. Recombinant human EPO (Syd Labs, BP000180-CYT-201) was added to a final concentration of 2 U/mL. After 8 days of induction, erythroid differentiation was assessed by FACS analysis of HbF expression. For intracellular staining of HbF, cells were fixed with 0.05% glutaraldehyde (grade II, Sigma-Aldrich, G6257) for 10 minutes at room temperature and centrifuged for 5 minutes at 600 × g. After permeabilization with 0.1% Triton X-100 in PBS with 0.1% BSA for 5 minutes at room temperature and centrifugation at 600 × g for 15 minutes, cells were incubated with APC-conjugated anti-human HbF antibody (Invitrogen, MHFH05) at 4°C for 20 minutes. Unbound antibodies were washed out using PBS supplemented with 2% FBS. Cells were analyzed using a FACSAria or FACSCanto flow cytometer with the FACSDiva v.8.0.2 software (BD Biosciences).

Inhibitor selection screens were performed to identify enriched second-site mutations at IDH1 or IDH2 that restored R-2HG production in the presence of IDH^mut inhibitor AG-120 or AG-221, respectively. We reasoned that if the second-site mutations conferred resistance to IDH^mut inhibitor, R-2HG production would be restored to block EPO-induced erythroid differentiation. As such, the second-site mutations would be progressively enriched in HbF-negative undifferentiated cell populations. Base-edited cells were treated with IDH^mut inhibitor (AG-120 for IDH1^R132H-BE cells and AG-221 for IDH2^R140Q-BE cells) at 1 μmol/L continuously for 16 weeks. EPO at 2 U/mL was then added to induce erythroid differentiation. Cells were stained for HbF expression, and HbF-negative cells were FACS-sorted. gDNA was extracted, and the exon sequences of IDH1 and IDH2 genes were PCR-amplified using primers listed in Supplementary Table S9. The DNA fragments for each exon were purified with QIAquick spin columns (Qiagen, 28106), mixed at equal molarity (500 nmol/L), and fragmented to 200 bp using Covaris acoustic shearing following the manufacturer's instructions. Sonicated DNA (25 μL) was processed for library generation using NEBNext Ultra II DNA Kits (New England Biolabs, E7645L) following the manufacturer's protocol. Libraries with different index sequences were quantified, pooled, and sequenced on an Illumina NextSeq 500 system using the 75-bp high-output sequencing kit. Sequencing fastq files were aligned to IDH1 or IDH2 reference sequences using BWA mem with default parameters. Reads that are supplementary aligned or not primary aligned were removed using SAMtools (51). SAMtools mpileup was used to generate pileup files from mapping files. pileup2base (https://github.com/riverlee/pileup2base) was used to parse SAMtools pileup file and to obtain the counts of bases (A, T, C, and G) for each position. Mutation frequencies at each position were then calculated from the pileup2base output. The mutation with the highest enrichment at each site is shown in Fig. 2B, and the other enriched mutations at the same site are listed in Supplementary Tables S3 and S8.

Saturation variant screens were performed to identify all possible IDH2 second-site mutations that conferred resistance to AG-221 inhibition. The IDH2 saturation variant library was designed as described previously (52–55) by substituting each wild-type amino acid at position 41 to 452 (excluding mitochondrial localization signal at positions 1–40 and stop codon at position 453) with one of the other 19 possible amino acids. The saturation variant library was designed and synthesized by Twist Bioscience. Total 7,828 IDH2-mutant cDNAs containing a single amino acid substitution were designed and 7,742 mutant cDNAs were confirmed to pass the quality control (Supplementary Table S4). The library containing IDH2-mutant alleles was cloned into pLVX-EF1α-IRES-TagBFP vector by Gibson assembly (New England Biolabs, E2611S) following the manufacturer's protocol. The Gibson assembly reaction (1 μL) was electroporated to 25 μL of E. coli 10G ELITE SUPREME Electro-competent cells (Lucigen, 60080-2) using a Bio-Rad MicroPulser (Bio-Rad Laboratories). The transformation product was plated onto prewarmed 24.5 cm² bioassay plates with ampicillin. The colonies on the plates were counted to calculate library coverage (>200×). All colonies were collected from the plates and maxiprep was performed to isolate the saturation variant library. Lentiviruses were produced as described previously (48). IDH2^R140Q-BE cells were transduced with the saturation variant library at low MOI (≤ 0.3) such that the majority of cells received only one mutant allele. Over 5 × 10⁶ BFP⁺ cells were FACS-sorted to get enough coverage of each mutant allele (>500×). Cells were collected before AG-221 treatment as the baseline time point (T0) and then treated with AG-221 at 1 μmol/L for the selection of second-site mutations associated with drug resistance. After 2, 4, 6, 8, and 10 weeks (T1 to T5) following AG-221 treatment, cells were induced for erythroid differentiation by EPO for 8 days and HbF-negative cells were FACS-sorted. gDNA was isolated, and IDH2 cDNA sequences were PCR-amplified using Phusion High-Fidelity DNA Polymerase (New England Biolabs, M0530L) with the primers listed in Supplementary Table S9. DNA fragments were diluted to 1 ng/μL and fragmented to 200 bp using Covaris acoustic shearing following the manufacturer's instructions. Twenty-five microliters of sonicated DNA were processed for library generation. Libraries with different index sequences were sequenced on an Illumina NextSeq 500 system using the 75-bp high-output sequencing kit. Initial data processing and alignment were performed by ORFCall (https://github.com/tedsharpe/ORFCall). ORFCall aligns reads to the IDH2 reference sequence. The counts for all 64 possible codons at each codon position were tallied and normalized by total codon counts. Codon enrichment was then calculated as the ratio of normalized codon counts between different time points (T1–T5) and T0. The mutations with progressive enrichment from T1 to T5 were selected by MaSigPro (56). The mutation with the highest enrichment at each site is shown in Fig. 2D, and the other enriched mutations enriched at the same site are listed in Supplementary Table S5.

Targeted sequencing of IDH1 and IDH2 genes was performed using bone marrow specimens from matched diagnosis and relapsed AML patients with IDH baseline mutation before or after IDH^mut inhibitor treatment. gDNA was extracted from the bone marrow clot sections using the QuickExtract FFPE DNA extraction kit (Epicenter Biotechnologies, QEF81805) following the manufacturer's instructions. For each sample, the exon sequences of IDH1 and IDH2 genes were PCR- amplified using primers located in the intronic regions and purified with QIAquick spin columns. Library preparation, sequencing, and data analysis were performed as described in the inhibitor selection screens.

Allele-specific amplicon sequencing was performed to identify the cis or trans configuration of the primary and secondary mutations in patients UT8, MDA2, and MDA7. Specifically, gDNA from these samples was isolated and used for targeted sequencing. Primers were designed to amplify the genomic region spanning exons 4 and 8 of IDH2. The 3.6-kb PCR products were cloned into EcoRI and BamHI sites of the pLVX-EF1α-IRES-zsGreen1 vector using the In-Fusion HD Cloning Kit. After transformation and verification by colony PCR and real-time PCR, the plasmid DNA was isolated and used for Sanger sequencing. For the analysis of cis or trans configuration in leukemia cells, the base-edited cells from inhibitor selection screens were collected to isolate RNA by TRIzol (Life Technologies, 15596018). cDNA was generated using an iScript DNA synthesis kit (Bio-Rad, 1708891) following the manufacturer's instructions and used for PCR cloning of IDH1 or IDH2 regions containing the primary and secondary mutations. Amplicon libraries with different index sequences were quantified and pooled, followed by paired-end sequencing on an Illumina NextSeq 500 system. Sequencing reads were aligned to the WT or mutant reference sequences and counted. The numbers and percentages of reads with mutations identified in cis or trans are listed in Supplementary Table S7. The primers used for cloning and sequencing are listed in Supplementary Table S9.

R-2HG was measured by LC/MS or the assay kit (BioVision, K213-100) following the manufacturer's instructions. Intracellular R-2HG was quantified in cell pellets from 2 × 10⁶ cells. For R-2HG measurement in culture medium, 200 μL medium was mixed with 800 μL methanol and dried with a SpeedVac into a pellet. The dried pellet was resuspended in 100 μL of freshly mixed 80:20 acetonitrile:acetic acid solution with 50 mg/mL diacetyl-L-tartaric anhydride (DATAN, Sigma-Aldrich, 358924) and U¹³C D/L-2HG solution (Cambridge Isotope Laboratories, CLM-10351-0.001) as internal standards. After sonication and heating at 75°C for 30 minutes, samples were cooled to room temperature and centrifuged. The resulting supernatant was dried with a SpeedVac to a pellet. The pellet was reconstituted in 100 μL of 1.5 mmol/L ammonium formate aqueous solution with 10% acetonitrile. LC/MS analysis of R-2HG levels was performed on an AB Sciex 5500 QTRAP liquid chromatography/mass spectrometer (Applied Biosystems SCIEX) and analyzed by the AB Sciex Analyst 1.6.1 Software as previously described (57). Waters Acquity UPLC HSS T3 column (150 × 2.1 mmol/L, 1.8 μmol/L) was operated at 35°C for separation. Different 2-hydroxyglutarate-diacetyl tartrate derivatives, 363/147 (CE: −14 V) and 368/152 (internal standard, CE: −14 V), were checked with multiple reaction monitoring.

NADPH and NADP⁺ levels in unmodified TF-1 cells or TF-1 cells stably expressing primary and/or second-site IDH mutations were measured using the NADP/NADPH Glo-Assay (Promega, G9081; refs. 58, 59) following the manufacturer's instructions. After the standard curves were generated using the purified NADPH (Sigma-Aldrich, N7505) and NADP⁺ (Sigma-Aldrich, N5755) in the same buffer as the experimental samples, the absolute amount of NADPH and NADP⁺ was obtained for each sample.

NADPH oxidation analysis was performed as previously described (4, 15). Specifically, cells were collected and cleaved by M-PER lysis buffer (Thermo Scientific, 78503) supplemented with protease and phosphatase inhibitors (Sigma-Aldrich, PPC1010). After centrifugation at 21,000 × g for 10 minutes, the supernatant was collected, quantified by BCA assay (Thermo Scientific, PI-23225), and normalized by total protein concentration. Purified proteins were also quantified by BCA assay for measuring enzymatic activity. Three micrograms of cell lysates or purified proteins were added to 200 μL of assay buffer containing 33 mmol/L Tris-acetate (pH 7.4, NovaTeinBio, NBB-2410), 1.3 mmol/L MgCl₂, 25 μmol/L β-NADPH, 40 mmol/L NaHCO₃, and 0.6 mmol/L αKG (Sigma-Aldrich, 75890), and absorbance at 340 nm was measured every minute for 5 hours. Datapoints were averaged among 5 time points centered at every 5 minutes and represented by the mean activity of 3 replicates.

TF-1 cells with stable luciferase expression were generated as previously described (60). Briefly, cells were transduced with lentiviruses produced with the pLVX-EF1α-Luc2-IRES-mCherry vector and mCherry-positive cells were FACS-sorted. Cells were then transduced with lentiviruses expressing IDH^WT with BFP and IDH^mut1 or IDH^mut1+mut2 with zsGreen1, and FACS-sorted. Eight- to 10-week-old NSGS mice were intravenously transplanted with mCherry, BFP, and zsGreen1 triple-positive cells (1 × 10⁶ per mice) resuspended in PBS (200 μL per mice). Following intraperitoneal injection of D-Luciferin (Gold Biotechnology, LUCK-1G) at 150 mg/kg, bioluminescence imaging was performed 4 hours after transplant to confirm fluorescent signals and ensure successful transplantation procedures. After disease onset for 7 days, the engrafted mice were randomly assigned for treatment with vehicle solution (control) or 30 mg/kg of AG-120 or AG-221 as described previously (12). When the mice became moribund after treatment with vehicle or inhibitors for 28 days, bioluminescence imaging was performed and intensity was quantitated by the Living Image v.3.2 acquisition and analysis software (Caliper Life Science). Total flux values were determined by regions of interest with an identical size for each mouse and presented in photons per second (p/s). Leukemia cell frequencies in peripheral blood and bone marrow were monitored by flow cytometry as previously described (61).

IDH1^WT/R132H and IDH2^WT/R140Q heterodimers were generated by cotransformation of plasmids containing GST-tagged IDH^WT and 6xHis-tagged IDH^mut1 into E. coli Rosetta2 (DE3) cells. For IDH^WT + IDH^mut1+mut2 heterodimers with second-site mutations in cis, plasmids containing GST-tagged IDH^WT and 6xHis-tagged IDH^mut1+mut2 were cotransformed. For IDH^mut1 + IDH^mut2 heterodimers with second-site mutations in trans, plasmids containing GST-tagged IDH^mut1 and 6xHis-tagged IDH^mut2 were cotransformed. Cells were selected with ampicillin and kanamycin. Protein expression and purification were performed as described previously (62). Specifically, E. coli Rosetta 2 (DE3) cells cotransformed with IDH-expressing plasmids were grown in LB medium with antibiotics at 37°C until OD₆₀₀ reached 0.6 to 0.8. Then, 0.5 mmol/L isopropyl-b-d-thiogalactoside (IPTG) was added to induce protein expression at 18°C overnight. After centrifugation at 4,000 × g for 10 minutes at 4°C, cell pellets were resuspended in buffer containing 300 mmol/L NaCl, 50 mmol/L Tris-HCl pH 8.0, 10% glycerol, 2 mmol/L β-mercaptoethanol (βME), and 1 mmol/L PMSF and were lysed by sonication. The supernatant was collected by centrifugation at 12,000 × g for 10 minutes at 4°C and flowed through Ni-NTA resin (G-Biosciences, 786-940) five times for the binding of 6xHis-tagged proteins. After washing with buffer 1 (20 mmol/L imidazole, 1 M NaCl, 50 mmol/L Tris-HCl pH 8.0, 10% glycerol, and 2 mmol/L βME) and buffer 2 (20 mmol/L imidazole, 100 mmol/L NaCl, 50 mmol/L Tris-HCl pH 8.0, 10% glycerol, and 2 mmol/L βME), proteins were eluted with buffer 3 (250 mmol/L imidazole, 100 mmol/L NaCl, 50 mmol/L Tris-HCl pH 8.0, 5% glycerol, and 2 mmol/L βME). The eluate was incubated with TEV protease at 4°C overnight followed by incubation with GST resin (Thermo Scientific, 16101) for 2 hours at 4°C with mixing. After washing with buffer 1 (0.1% NP40, 500 mmol/L NaCl, 50 mmol/L Tris-HCl pH 8.0, 10% glycerol, and 2 mmol/L DTT), buffer 2 (1 M NaCl, 50 mmol/L Tris-HCl pH 8.0, 10% glycerol, and 2 mmol/L DTT), and buffer 3 (100 mmol/L NaCl, 50 mmol/L Tris-HCl pH 8.0, 10% glycerol, and 2 mmol/L DTT), protein-bound GST resin was incubated with HRV-3C protease at 4°C overnight. The flow-through fractions were collected by eluting with buffer 4 (100 mmol/L NaCl, 20 mmol/L Tris-HCl pH 8.0, and 2 mmol/L DTT). The proteins were quantified by BCA assay and stored as aliquots at −80°C.

Different IDH heterodimers isolated with tandem affinity purification were validated by native polyacrylamide gel electrophoresis (PAGE), followed by coomassie blue staining and western blot. The purified proteins were quantified by a BCA assay. Protein samples were prepared with native sample buffer (Bio-Rad, 1610738) and separated on a PAGE gel with nonreducing buffer in the absence of SDS. The proteins on the gel were stained with Colloidal Blue Staining Kit (Thermo Scientific, LC6025) following the manufacturer's instructions or transferred to Amersham Hybond P 0.45 PVDF blots (GE Healthcare, 10600023). After blocking with 5% nonfat milk in TBST (20 mmol/L Tris-HCl pH 7.5, 150 mmol/L NaCl, and 0.1% Tween-20), the blots were incubated with antibodies against IDH1 (Proteintech, 12332-1-AP) or IDH2 (Abcam, ab55271) with 1:500 dilution at 4°C overnight with shaking, followed by secondary antibodies at room temperature for 1 hour. The blots were washed with TBST three times before and after antibody incubation and developed with Plus-ECL (PerkinElmer, NEL104001EA). To measure histone methylation, western blot analysis was performed as previously described (63) with antibodies against H3K4me3 (Abcam, ab8580), H3K9me2 (Abcam, ab1220), H3K9me3 (Abcam, ab8898), H3K27me3 (Millipore Sigma, 07-449), H3K36me3 (Abcam, ab9050), H3K79me2 (Abcam, ab3594), and total H3 (Abcam, ab1791) with 1:1,000 dilution.

For uncompetitive inhibition by AG-120 and AG-221 with regard to NADPH, IDH1^WT+R132H and IDH2^WT+R140Q heterodimer proteins were diluted to 100 nmol/L in 120 μL assay buffer (20 mmol/L Tris pH 7.5, 150 mmol/L NaCl, 10 mmol/L MgCl₂, 10% glycerol, and 0.03% BSA) containing vehicle or IDH^mut inhibitor (100 nmol/L AG-120 for IDH1^WT+R132H or 75 nmol/L AG-221 for IDH2^WT+R140Q), 1 μmol/L NADP⁺, and NADPH dilution series. For uncompetitive inhibition by AG-120 and AG-221 with regard to NADP⁺, heterodimer proteins were diluted to 50 nmol/L in 120 μL assay buffer (20 mmol/L Tris pH 7.5, 150 mmol/L NaCl, 10 mmol/L MgCl₂, 10% glycerol, and 0.03% BSA) containing vehicle or IDH^mut inhibitor (1 μmol/L AG-120 for IDH1^WT+R132H or 25 μmol/L AG-221 for IDH2^WT+R140Q), 10 μmol/L NADPH, and NADP⁺ dilution series. After 1-hour incubation at 25°C to allow the formation of complexes containing inhibitor, IDH heterodimer, NADP⁺, and NADPH, the reaction was initiated by adding 30 μL assay buffer with 6.25 mmol/L αKG, incubated for 2 minutes at 25°C, and terminated by adding 50 μL assay buffer with 36 μg/mL of diaphorase and 18 μmol/L resazurin. The remaining NADPH levels were measured on a FLUOstar Omega plate reader (BMG Labtech) via fluorescence (λ_ex = 544 nm, λ_em = 590 nm). Three independent measurements were performed for each dilution of NADPH or NADP⁺. The curves were generated in GraphPad Prism software.

Analysis of binding affinity between NADPH and IDH heterodimers or NADP⁺ and IDH homodimers with or without second-site mutations was performed as previously described (4, 12). Briefly, to determine the binding affinity of NADPH to heterodimers with or without second-site mutations, 20 to 100 nmol/L of purified proteins were diluted in the assay buffer (20 mmol/L Tris pH 7.5, 150 mmol/L NaCl, 10 mmol/L MgCl₂, 10% glycerol, and 0.03% BSA) with NADPH dilution series. The reactions were initiated by adding 6.25 mmol/L αKG in the assay buffer. After incubation for 10 minutes at room temperature, the reactions were then terminated by adding assay buffer with 36 μg/mL of diaphorase and 18 μmol/L resazurin, and mixed for 1 minute with a shaker. The remaining NADPH levels were measured on a plate reader by fluorescence (λ_ex = 544 nm, λ_em = 590 nm). To determine the binding affinity of NADP⁺ to homodimers with first- and/or second-site mutations, 1 to 5 μmol/L purified proteins were diluted in the assay buffer with NADP⁺ dilution series. The reactions were initiated by adding assay buffer with 0.2 mmol/L ICT, 60 μg/mL diaphorase, and 200 μmol/L resazurin. After running for 30 minutes at room temperature, the reactions were terminated by adding 6% SDS with mixing for 1 minute. NADPH generated by the reactions was measured on a plate reader. Absolute NADPH concentration for each sample was obtained using an NADPH standard curve generated in the assay buffer alone. Three independent measurements were performed for each dilution of NADPH or NADP⁺. The curve, best-fit values ± SE, and 95% confidence interval of Km for each sample were generated in GraphPad Prism.

IDH heterodimer proteins with or without second-site mutations were diluted to 0.25 μg/mL in 120 μL assay buffer (20 mmol/L Tris pH 7.5, 150 mmol/L NaCl, 10 mmol/L MgCl₂, 10% glycerol, and 0.06% BSA) containing IDH^mut inhibitor dilution series, 42.5 μmol/L NADP⁺, 5 μmol/L NADPH, and 2.5 mmol/L βME. The mixture was incubated at 25°C for 1 hour or 16 hours. The reaction was initiated by adding 30 μL assay buffer with 6.25 mmol/L αKG. The reactions were run for 50 minutes at 25°C and terminated by 50 μL assay buffer with 36 μg/mL of diaphorase and 18 μmol/L resazurin. After shaking the reaction for 1 minute, the NADPH levels were measured on a plate reader by fluorescence (λ_ex = 544 nm, λ_em = 590 nm). Three replicates were performed for each dilution of IDH^mut inhibitors. The assay values with AG-120 or AG-221 at 50 μmol/L maximum concentration were scaled to 100%. The curve and value of IC₅₀ for each sample were generated in GraphPad Prism.

Structural analysis was performed based on the crystal structures of 20a (an AG-120 analogue)-bound IDH1 (PDB #5L57) and AG-221–bound IDH2 (PDB #5I96; refs. 12, 28). Structure figures were rendered in PyMOL Molecular Graphics System (Version 1.2r3pre; Schrödinger, LLC). The biologically relevant IDH1 dimer was constructed by combining the deposited structure of one protomer with the structure of another symmetry-related protomer. Point mutations were generated in Coot (64), and mutated residues were represented as rotamers with the highest probability that do not sterically clash with other residues in the structure.

IDH2^R140Q proteins with or without second-site mutations at 1 μmol/L concentration were mixed with 50 μmol/L NADP⁺, 5 μmol/L NADPH, and the indicated concentrations of AG-221 (Cayman, 21277) in 20 μL assay buffer (20 mmol/L Tris pH 7.5, 150 mmol/L NaCl, 10 mmol/L MgCl₂, and 2% DMSO). The mixtures were incubated at 25°C for 16 hours before adding SYPRO orange (Sigma, S5692) with a 1,000-fold dilution of the stock. The dye-containing mixtures were transferred to a 96-well plate to acquire fluorescence signals at temperatures ranging from 20°C to 85°C on a Roche LightCycler480 real-time PCR machine. Tm values corresponding to the maxima of the negative first derivative of fluorescence signals as a function of temperature were manually recorded. The apparent K_d values were determined using GraphPad Prism.

Statistical details including N, mean, and statistical significance values are indicated in the text, figure legends, or methods. Error bars in the experiments represent SEM, SE, or SD from either independent experiments or independent samples. All statistical analyses were performed using GraphPad Prism software unless otherwise specified, and the detailed information about statistical methods is specified in figure legends or methods. The numbers of independent experiments or biological replicate samples and P values (*, P < 0.05; **, P < 0.01; ***, P < 0.001; ****, P < 0.0001; n.s., not significant) are provided in individual figures. P < 0.05 was considered statistically significant. Figures 5E and 6D and Supplementary Figs. S1C, S7A, and S8A show a representative image at least three independent experiments or biological replicate samples with similar results.

All processed datasets for targeted sequencing and screens are available in Supplementary Tables S2–S6. Raw sequencing data are deposited in the European Nucleotide Archive under accession number PRJEB48042. Targeted sequencing analyses were performed using Bowtie2 v2.2.8, BWA v0.7.17, SAMtools v 0.1.19, pileup2base (https://github.com/riverlee/pileup2base), MaSigPro, and ORFCall (https://github.com/tedsharpe/ORFCall). Code for analyses using other indicated software is available from the websites of the corresponding software.

Y.F. Madanat reports other support from BluePrint Medicines, GERON, OncLive, Sierra Oncology, Stemline Therapeutics, Novartis, and Morphosys outside the submitted work. P. Patel reports other support from Servier and Bristol Myers Squibb/Celgene outside the submitted work, and is an employee of Servier BioInnovation but was at the University of Texas Southwestern at the time of this work. A.S. Mims reports other support from The Leukemia & Lymphoma Society, personal fees from Servier Pharmaceuticals, Syndax Pharmaceuticals, Astellas Pharmaceuticals, Jazz Pharmaceuticals, Daiichi Sanyko, Bristol Myers Squibb, Ryvu Therapeutics, Genetench, and AbbVie outside the submitted work. U. Borate reports grants and personal fees from Bristol Myers Squibb and AbbVie outside the submitted work. S.F. Cai reports other support from Imago Biosciences outside the submitted work. R.J. DeBerardinis reports personal fees from Agios Pharmaceuticals outside the submitted work. J. Xu reports grants from the NIH, the Cancer Prevention & Research Institute of Texas, The Leukemia & Lymphoma Society, the American Society of Hematology, and the Welch Foundation during the conduct of the study. No disclosures were reported by the other authors.

J. Lyu: Conceptualization, data curation, formal analysis, validation, investigation, methodology, writing–original draft, writing–review and editing. Y. Liu: Data curation, software, formal analysis, methodology. L. Gong: Data curation, formal analysis, investigation, methodology. M. Chen: Resources, data curation, methodology. Y.F. Madanat: Resources, methodology. Y. Zhang: Data curation, software, formal analysis, methodology. F. Cai: Formal analysis, investigation, methodology. Z. Gu: Investigation, methodology. H. Cao: Investigation, project administration. P. Kaphle: Investigation. Y.J. Kim: Investigation. F.N. Kalkan: Resources, methodology. H. Stephens: Resources, methodology. K.E. Dickerson: Methodology. M. Ni: Resources, software, investigation, methodology. W. Chen: Resources, data curation. P. Patel: Resources, methodology. A.S. Mims: Resources, methodology. U. Borate: Resources, methodology. A. Burd: Resources, methodology. S.F. Cai: Resources, data curation, methodology. C.C. Yin: Resources, data curation, methodology. M.J. You: Resources, data curation. S.S. Chung: Resources, supervision, methodology, project administration. R.H. Collins: Resources, supervision, project administration. R.J. DeBerardinis: Resources, supervision, funding acquisition. X. Liu: Software, formal analysis, supervision, funding acquisition, investigation, methodology, writing–original draft. J. Xu: Conceptualization, resources, supervision, funding acquisition, methodology, writing–original draft, project administration, writing–review and editing.

We are grateful to Ralf Kittler, Paul Yenerall, and Rahul Kollipara for helping with data analysis and discussions; Ryan Huang for assay design; Ashley Yocum and Timothy Chen for assistance with Beat AML samples; and other Xu laboratory members for technical support. J. Xu is a Scholar of The Leukemia & Lymphoma Society and an American Society of Hematology Scholar. This work was supported by NIH grants R01DK111430, R01CA230631, R01CA259581, and R21AI158240 to J. Xu; NIH grants R01GM121662 and R01GM136308 to X. Liu; Cancer Prevention & Research Institute of Texas grants RP180504, RP190417, RP220337, and RP220375 (to J. Xu); and Welch Foundation grants (I-1942 to J. Xu and I-1790 to X. Liu).

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.