Dominant Regulatory Axis Detection for reproducible regulatory-state analysis in paired AML diagnosis-relapse transcriptomes

Dominant Regulatory Axis Detection (DRA) is a low-dimensional framework for analyzing relative dominance among biologically specified transcriptional programs. This report applies DRA to 59 matched acute myeloid leukemia diagnosis-relapse pairs from GSE327386 using a 29-patient discovery cohort and a separately reserved 30-patient locked replication cohort. The principal replicated observation was relapse-associated redistribution away from relative GR/NR3C1 dominance within a prespecified NFκB/GR/p53 competitive field. All four frozen executable directional replication criteria were satisfied in Cohort II. Post-replication exploratory analyses identified heterogeneous GR-origin relapse configurations and pathway-level differences between NFκB- and p53-directed trajectories. The record contains the manuscript/report, a public-safe supplementary and reproducibility archive, and SHA-256 integrity information. Discovery, locked replication, sensitivity analyses and post-replication exploration are explicitly distinguished throughout. This version has not undergone journal peer review. Abstract Acute myeloid leukemia (AML) relapse is accompanied by extensive transcriptional change, complicating distinction between coordinated regulatory-state reorganization and diffuse gene-expression variation. Dominant Regulatory Axis Detection (DRA) is a low-dimensional framework that projects transcriptomic measurements onto a prespecified competitive field of transcription-factor-associated programs and evaluates their relative dominance. We applied DRA to 59 matched diagnosis-relapse AML pairs using a discovery-to-replication design. In a 29-patient discovery cohort, GR/NR3C1 was dominant in 18 samples at diagnosis and 12 at relapse; 10 of 18 initially GR-dominant cases lost GR dominance. The median paired GR-axis change was −0.449, with decreases in 19 of 29 patients. This observation was converted into four frozen directional replication criteria before analysis of a separately reserved 30-patient cohort. In Cohort II, GR dominance decreased from 15 to 12 samples, six patients exited GR dominance while three entered it, and the median paired GR-axis change was −0.864. All four prespecified criteria were satisfied, including preservation of the GR-loss direction in a 21-pair specimen-source-matched sensitivity analysis. Post-replication exploration showed that loss of GR dominance did not converge on a single alternative state. Among cases that were GR-dominant at diagnosis across both cohorts, 17 retained GR dominance, whereas eight transitioned to NFκB and eight to p53 dominance. Exploratory ranked gene-set analysis further distinguished these trajectories at the pathway level despite the absence of individual genes significant after genome-wide multiple-testing correction. These findings show that DRA can identify a low-dimensional regulatory redistribution in paired AML transcriptomes and reproduce its direction in a separately reserved cohort under a locked analytical framework. DRA provides an interpretable approach for generating falsifiable hypotheses about regulatory-state organization while maintaining explicit separation between discovery, replication, and subsequent exploration.