How do 3D-DNA architecture changes shape IDH1- and DNMT3A-mutant AML?

What HiC adds beyond methylation data

The paper asks whether mutations in two epigenetic regulators, IDH1 and DNMT3A, reshape the physical folding of DNA in leukemic cells, and whether that reshaping creates targets that methylation data alone would miss. Both mutations are known to alter DNA methylation globally, and IDH1 mutations had already been tied to changed DNA-loop formation and oncogene upregulation. This study scales that question up to the whole genome and compares the two genotypes side by side.

The central measurement is chromatin-conformation-capture, or HiC: it reads out which distant DNA segments physically contact each other in the nucleus, producing genome-wide interaction maps. From those maps the authors derive two levels of structure. Compartments are large-scale regions labeled active (A) or inactive (B); DNA loops are shorter-range contacts, often anchored at CTCF binding sites, that bring enhancers near promoters. A loop score quantifies how strongly a given loop is present, so a rising score means a contact became more frequent and a falling score means it weakened.

Shared and genotype-specific loops in both models and patients

The authors built a cellular model by inserting either IDH1 p.R132H or DNMT3A p.R882H into KG1a myeloid leukemia cells, which lack other known driver mutations. They used 11 CRISPR-modified clones: three IDH1 wild-type controls, three IDH1-mutant clones, two DNMT3A wild-type controls, and three DNMT3A-mutant clones. Eight primary AML samples at diagnosis were analyzed in parallel — three wild-type for both genes, three with IDH1 p.R132H/C, one with DNMT3A p.R882H/C, and one carrying both mutations — and served as an exploratory and validation dataset.

In patient samples, 9,839 loops were identified in total, of which 3,617 were common across genotypes. Gains were partly shared (888 loops gained in both mutants) and partly specific: 934 gained only in IDH1-mutant and 1,378 only in DNMT3A-mutant samples. Losses followed the same pattern, with 724 lost in IDH1-mutant, 872 in DNMT3A-mutant, and 1,426 lost in both. The single double-mutant patient carried both IDH1- and DNMT3A-associated gained and lost loops, which the authors read as support for mutation-specific architectural signatures rather than a single generic leukemia signature.

The paper also links loop changes back to methylation: hypermethylated CTCF anchors were enriched among lost loops, and hypomethylated loops among gained loops. That is a correlation between methylation state and loop status, not a demonstration that methylation caused the loop change.

Altered loop anchors carry differentially expressed genes

To connect structure to transcription, the authors asked where differentially expressed genes sit. Genes with expression changes in IDH1- or DNMT3A-mutant KG1a cells were significantly more often located at anchors of lost or gained loops than at anchors of loops common to both genotypes. Loops whose anchors hosted at least one differentially expressed gene also showed larger loop-score changes across genotypes.

Loop-score change correlated positively with gene-expression change for both genotypes in KG1a, and the direction of change was largely consistent between cell lines and patients: 53% of genes with overlapping loop-score changes in IDH1-mutant KG1a and patient samples, and 60% in DNMT3A-mutant samples, also showed similar expression changes. The overlap is partial, which matters for how far the cell-line results can be pushed.

Among upregulated genes in altered loops shared between KG1a and patients, the authors list EXT1, CHD7, and IGF1R for IDH1-mutant samples, and IKZF3, ETS1, SEPTIN9, and IRS2 for DNMT3A-mutant samples. For IGF1R, increased interaction intensity around the gene locus was confirmed in both KG1a and AML samples and coincided with reduced CpG methylation at the downstream CTCF anchor. The authors present altered looping as a potential driver of the transcriptional changes; the evidence is a consistent association across genotypes and sample types, not a direct causal test at this stage.

A drug-correlation screen ranks loop genes as targets

To turn loop-associated genes into candidate vulnerabilities, the authors used the BeatAML 2.0 dataset, which pairs ex vivo drug-response profiles for more than 300 compounds with gene expression from 805 AML patients; 543 samples had both datasets. Because mutation-specific drug-response data were limited, they correlated drug AUC values with loop-gene expression across all samples rather than within each genotype, then combined the Spearman coefficient with gene-expression log2 fold changes from KG1a and patient samples into a rank product.

In IDH1-mutant AML, 48 overlapping loop genes were ranked; in DNMT3A-mutant AML, 57. For IDH1-mutant samples, ST8SIA6 and ADAMTS3 ranked highest against the LSD1 inhibitor GSK-2879552 (Spearman coefficients −0.432, p = 4.09 × 10⁻⁵ and −0.289, p = 0.0076), and IGF1R appeared among the top candidates for the IGF1R inhibitor BMS-754807 (coefficient −0.424, p = 0.0057). For DNMT3A-mutant samples, TBC1D16 and FMN1 ranked highest against ralimetinib (coefficient 0.381, p = 0.0133) and the HDAC inhibitor panobinostat (coefficient −0.306, p = 1.28 × 10⁻⁷), respectively.

Ranking by the number of loop-gene targets put venetoclax and BMS-754807 at the top for IDH1-mutant AML, and elesclomol and panobinostat for DNMT3A-mutant AML. The BeatAML data also showed IDH1-mutant patients were more sensitive to BMS-754807 than IDH1 wild-type patients, and ralimetinib ranked among drugs with better response in DNMT3A-mutant patients. These are correlative rankings across a mixed patient set, so they nominate candidates rather than establish genotype-specific efficacy.

Cell-line testing supports two genotype-specific candidates

The authors then tested the nominated drugs directly in their CRISPR models with WST-1 viability assays. IDH1-mutant clones showed increased sensitivity to BMS-754807 monotherapy, and combining BMS-754807 with the IDH1-specific inhibitor ivosidenib produced synergistic effects under a Loewe synergy model. Sensitivity to GSK-2879552 was visible but modest.

DNMT3A-mutant cells showed increased sensitivity to ralimetinib both as monotherapy and in combination with venetoclax. Differential sensitivity of DNMT3A wild-type versus mutant cells to BMS-754807 and elesclomol remained only mild in vitro.

A specificity check strengthened the loop-to-drug link: CRISPR knockout of the CTCF anchor sites for IGF1R and TBC1D16 reverted not only gene expression but also the corresponding drug response. That experiment ties the candidate vulnerability to the specific anchor rather than to the mutation alone, although it was performed in the cell-line models.

Where the conclusions hold and where they thin out

The patient dataset is small and uneven: eight samples, only one DNMT3A-mutant and one double-mutant, and the authors describe it as exploratory and for validation. Patient samples also varied substantially in gene expression and 3D-DNA alterations, consistent with AML's complex landscape. The authors name the reliance on KG1a as the primary cellular model as a limitation, while pointing to the shared alterations and overlapping transcriptional patterns between KG1a and patients as evidence that the mechanism is not model-specific.

The drug-correlation step is also indirect. Because mutation-specific drug-response data were limited, correlations were computed across all 543 samples rather than within genotype, so a loop gene's correlation with a drug's AUC does not by itself show that the drug acts through that loop in mutant cells. The in vitro synergy and anchor-knockout results are what move BMS-754807 and ralimetinib from ranked candidates toward testable hypotheses.

Why the IGF1R loop matters most for IDH1-mutant AML

The paper's most distinctive claim is not that 3D architecture differs between AML subtypes, but that a specific structural change points to a specific drug pairing. In IDH1-mutant AML, the IGF1R locus gains interaction intensity, loses CpG methylation at its downstream CTCF anchor, and shows upregulated expression; the same cells are sensitive to BMS-754807, and that sensitivity increases when the IGF1R inhibitor is combined with ivosidenib. The authors note that IGF1R upregulation had not previously been reported in IDH1-mutant AML, and they connect it to MAPK signaling activation as a possible survival mechanism.

For DNMT3A-mutant AML, the parallel candidate is ralimetinib, supported by TBC1D16 ranking and by monotherapy and venetoclax-combination sensitivity in the mutant cell line. The authors frame this as particularly worth testing because venetoclax response in DNMT3A-mutant patients is variable. The practical priority the data support is therefore a genotype split: IGF1R inhibition plus ivosidenib for IDH1-mutant disease, and p38/MAPK inhibition for DNMT3A-mutant disease, with both still requiring clinical assessment rather than standing as established treatments.

Sources used in this answer