Clinical implications of epigenetic age acceleration as a novel biomarker in Acute Myeloid Leukemia

Abstract Background: Building on our first-generation epigenetic clock, we now present an enhanced approach to capture biological aging processes in acute myeloid leukemia (AML). While our initial clock demonstrated proof-of-concept for epigenetic age acceleration (EAA) as a prognostic biomarker, limitations in accuracy and robustness highlighted the need for advanced computational architectures. The clinical utility of EAA in larger cohorts towards predicting treatment response to targeted therapies, particularly IDH inhibitors in AML, remains unexplored despite the critical need for biomarkers for precision medicine approaches in AML. Methods: We developed a robust epigenetic clock using the Kolmogorov-Arnold Network (KAN) architecture, which employs learnable activation functions to better capture the complex, non-linear relationships underlying biological aging, compared to traditional neural networks with static weights. Methylation datasets calibrated for Illumina's 450k array were obtained from NCBI's Gene Expression Omnibus (GEO) and pre-processed using the minifi package. After quality control, the filtered dataset was partitioned into training and testing sets using an 80:20 split, with an additional independent validation dataset reserved for final model evaluation. Model performance was evaluated using three independent approaches: A dataset from our internal GEO collection, datasets accessed through Biolearn, and 21 AML patient-derived samples that we sequenced independently. Performance was assessed using mean absolute error (MAE) and correlation coefficients (R²), with metrics compared against established epigenetic clocks to assess relative accuracy and clinical utility. Results: 12,400 patient samples were screened, and we excluded those with >30% missing CpG site data. 10,368 patients were included in the final analysis and split into a training dataset (n=5,816), an internal test set (n=1,454), and a validation set (n=3,116). Our second-generation KAN-based epigenetic clock achieved strong performance with R² of 0.91 and MAE of 4.22 yrs on the internal test set, and R² of 0.89 and MAE of 4.58 yrs on the validation dataset. For external validation, we benchmarked our clock against four established clocks (Horvath, Hannum, PhenoAge, and DunedinPace) using seven independent datasets (n=2,669) obtained through Biolearn. Our KAN-based clock demonstrated superior performance with a median R² of 0.86 and MAE of 3.61 yrs, outperforming competing clocks in 5 of 7 comparisons, with the Horvath clock achieving best performance in the remaining 2 datasets. We applied our clock to identify patients unlikely to achieve IDH-clearance following IDH inhibitor therapy. Among 60 AML patients, 22 had available IDH clearance data. Patients who failed to clear IDH clones (n=15) demonstrated significantly higher EAA compared to those who achieved clearance (n=7) [11.3 vs 8.8 years; p=0.023], enabling us to establish an EAA threshold of 11 years for predicting IDH non-clearance. To validate this threshold, we sequenced 20 IDH-mutated AML patient samples with known clinical outcomes. Of these, 17 patients failed to achieve IDH clone clearance, and our EAA threshold correctly captured 11 of these non-responders (65% sensitivity). Among the 3 patients who achieved IDH clearance, one maintained a variant allele frequency of 3%, one exhibited minimal EAA, and one showed no acceleration. Conclusion: Our KAN-based epigenetic clock outperformed existing models and identified a clinically actionable EAA threshold for predicting IDH inhibitor non-response in AML patients. Ongoing work is focused on investigating the use of EAA as a predictive biomarker for clinical outcomes across epigenetic therapies in AML.