AUTHORS
Vikram Dhillon, DO, MBA, Jeff Justin Aguilar, MD, MBA2, Asmita Shukla, MD3, Dakshin Padmanabhan, MD4, Sikander Chohan, MD5, Gregory Dyson6, Suresh Kumar Balasubramanian, MD7
AFFILIATIONS
1 Karmanos Cancer Institute, Detroit, United States
2 Karmanos Cancer Institute, Detroit, United States
3 Karmanos Cancer Institute, Detroit, United States
4 Karmanos Cancer Institute, Detroit, United States
5 Karmanos Cancer Institute, Detroit, United States
6 Karmanos Cancer Institute, Detroit, United States
7 Karmanos Cancer Institute, Detroit, United States
ABSTRACT
Background: Venetoclax (VEN) resistance in acute myeloid leukemia (AML) has been linked to monocytic differentiation, leukemia stem-cell states, and metabolic reprogramming, but published signatures have not been reconciled or translated to large bulk RNA-sequencing cohorts. We asked whether single-cell resistance signatures converge on shared programs measurable in bulk transcriptomes and whether these define AML subgroups with distinct genotypes, VEN responses, and therapeutic vulnerabilities.
Methods: Published AML single-cell RNA-sequencing datasets were harmonized, retaining blast/myeloid cells. Reported stem-cell and VEN-resistance signatures were scored, grouped by correlation, and factorized by nonnegative matrix factorization into consensus programs. Full-transcriptome references were deconvolved into BeatAML2 and TCGA-LAML using BayesPrism. Program abundance was associated with phenotype, morphology, cytogenetic risk, mutations/co-mutation class, ex-vivo VEN response, broader drug sensitivity, and survival. Stability, reconstruction, inter-program correlation, and cross-cohort concordance underwent pre-specified quality control.
Results: The atlas included 129,401 cells from 51 patients, with 56,031 blast/myeloid cells used for derivation. Eighteen published signatures consolidated into 12 correlated modules and 6 stable programs (R²=0.862; stability 0.980±0.039), projected into 707 BeatAML2 transcriptomes (367 with ex vivo VEN response).
The granulocytic/favorable-risk program was VEN-sensitive, enriched for IDH2 and STAG2 and depleted of RUNX1, and RAS-pathway mutations. The monocytic program, the largest and predominant VEN-resistant state, was enriched for NRAS, KRAS, PTPN11, ASXL1, RUNX1, and inv(16)/t(16;16), and depleted of NPM1, CEBPA, GATA2, WT1, and IDH1. The proliferative/Ox-Phos program was enriched for CEBPA (highest biallelic frequency) and TP53, with NPM1, FLT3-ITD, DNMT3A, IDH1/2, and TET2 absent. The erythroid/megakaryocytic-progenitor (E/M) program marked adverse-risk, secondary-type disease with complex karyotype, chromosome 7 abnormalities, and inv(3)/t(3;3), and was enriched for BCOR, RUNX1, SRSF2, IDH2, JAK2, PHF6, and RAD21 but depleted of NPM1, NRAS/KRAS, and DNMT3A. The primitive stem-like program showed the greatest VEN sensitivity, with NPM1/FLT3-ITD co-occurrence, frequent signaling alterations, and enrichment for NPM1, GATA2, WT1, and IDH1 but depletion of RAS, ASXL1, RUNX1, and SRSF2. The immune/cDC-like program was low-abundance, associated with BCORL1 and CCND3, and lacked a VEN-response phenotype.
Resistance programs coexisted in 154/707 patients (22%), and combinations had distinct phenotypes. Granulocytic-with-primitive was most VEN-sensitive (ρ=−0.638; p=2.87×10?4³), whereas granulocytic-with-monocytic conferred resistance (ρ=0.490; p=1.62×10?²³), indicating that VEN response reflects composite lineage-state architecture. Eight resistant samples were unexplained by the six-program taxonomy and instead showed enrichment of the E/M program (p=7.96×10?5). Independent NMF recovered a stable heme/megakaryocytic program (ITGA2B, GATA1, MPL, SLC40A1, ANK1, TFR2; heme-metabolism FDR=2.34×10?4). Cohort-wide, it was enriched in BCOR-, RUNX1-, and SRSF2-mutated AML and more abundant in chromatin/spliceosome-only than monocytic-route genotypes (0.225 vs 0.113; p=1.20×10?6), supporting an alternative to monocytic resistance.
Drug profiling identified 302 program-drug associations across 154 agents. The monocytic program was the most MEK-sensitive state (selumetinib ρ=−0.473, FDR=7.02×10?²4; trametinib ρ=−0.349, FDR=2.49×10?¹³; mean MEK-class ρ=−0.41), with additional sensitivity to panobinostat (ρ=−0.506; FDR=1.35×10?¹7), rapamycin, and MK-2206, whereas the granulocytic program was resistant to both MEK inhibitors, supporting program-specific treatment selection. The E/M route showed distinct vulnerability to IGF1R inhibition and fatty-acid oxidation blockade (NVP-AEW541 ρ=−0.370, FDR=0.043; etomoxir ρ=−0.368, FDR=0.044).
Conclusion: Harmonization of published single-cell resistance signatures produced a clinically interpretable VEN-response taxonomy measurable from bulk RNA sequencing. VEN response reflects composite lineage-state architecture rather than a single resistance program, with a dominant monocytic, RAS/signaling-linked route and a previously unrecognized, genotype-linked E/M alternative associated with chromatin/spliceosome mutations and distinct metabolic vulnerabilities.
DISCLOSURES
Vikram Dhillon, DO, MBA: No relevant conflicts of interest to declare.