An unbiased machine-learning model identifies KMT2A::AFF1 complex binding as a driver of differential enhancer usage. (A) Proportion of enhancers (common = no change in activity, RS4;11 = increased activity in RS4;11 cells, SEM = increased activity in SEM cells) containing either an SNV (blue), an indel (yellow), or both (green) in RS4;11 cells (left) or SEM cells (right). (B) Proportion of enhancers of each enhancer type (common, RS4;11, SEM) containing no heterozygous SNVs (dark gray), SNVs removed due to intrinsic bias (eg, problematic genomic regions or mapping bias; gray), SNVs without allele-specific bias in accessibility (light gray), or those exhibiting allele specific bias in accessibility as measured by ATAC-seq (red). (C) Schematic of the strategy used to determine key predictive features of differential enhancer activity. ChIP-seq signal for 56 factors was extracted over enhancers with increased activity in RS4;11 cells (red; 1522), or SEM cells (blue; 1677) or those common to both (gray; 4232). A gradient boosted decision tree was trained from these data, and predictive features were extracted using SHAP. (D) The relative feature importance for each enhancer category (increased activity in SEM cells [blue], RS4;11 cells [red], or common enhancers [gray]) of the top 20 most important features for differential enhancer prediction. Features that correspond to binding of the KMT2A::AFF1 complex are highlighted, and a schematic of the complex is provided for reference. (E) Tornado plot of AFF1-C and KMT2A-N ChIP-seq signal at enhancers displaying increased activity in RS4;11 cells (top) or SEM cells (bottom) in RS4;11 (left) or SEM (right) cells. (F) Pearson correlation between H3K27ac and KMT2A signal at the 290 blast-specific enhancers identified, for each patient sample. (G) H3K27ac ChIP-seq signal at enhancers with increased activity in SEM cells (blue), RS4;11 cells (red), or common enhancers (gray) upon KMT2A::AFF1 knockdown by small-interfering RNA (siRNA; dashed line). (H) Enhancer-promoter interaction frequency at enhancer regions with increased activity in SEM cells (i-iv) or RS4;11 cells (v) upon treatment of SEM (i) or RS4;11 (v) cells with 2 μM EPZ5676 for 1 week or SEM PAF1-FKBP12^F36V (iii)/SEM SSRP1- FKBP12^F36V (iv) cells treated with dTag13 for 24 hours, together with SEM cells treated with an siRNA against KMT2A::AFF1 (i). Interaction frequency for each enhancer-promoter pair is shown relative to the mean interaction frequency of the control; n = 3 biological replicates per condition. ∗P < .05; ∗∗P < .01; ∗∗∗P < .001. (I) Example of a loss of enhancer-promoter interactions at the ARID1B locus in SEM cells as assessed by Capture-C in control (gray) or KMT2A::AFF1 knockdown conditions (red) in 3 biological replicates. Enhancers with increased activity in SEM cells are highlighted in blue. ChIP-seq for H3K27ac in control (gray) or KMT2A::AFF1 knockdown conditions (red) in addition to the N terminus of KMT2A and the C terminus of AFF1 are provided for reference. (J) Model for the role of the KMT2A::AFF1 complex in promoting transcription heterogeneity between patients. bp, base pair; dTag, dTAG-13; EPZ, EPZ5676; indel, insertion-deletion; KD, knockdown; SHAP, SHapley Additive exPlanations.