AHA Scientific Sessions 2026 Abstract · Nov 2, 2026
Interpretable Machine Learning Identifies Predictors of Response to Hepatic Angiotensinogen Suppression in Uncontrolled Hypertension
By Edelberg D, Sarraju A, Costet P et al.
Background: Tonlamarsen (TLA), an investigational liver-targeted antisense oligonucleotide that suppresses hepatic angiotensinogen (AGT), is intended to lower blood pressure (BP) by targeting the most upstream substrate of the renin-angiotensin-aldosterone system (RAAS). In the KARDINAL phase 2 randomized controlled trial, half of trial participants were assigned to a single 90-mg dose of TLA followed by placebo, and half were assigned to 5 doses of TLA administered once every 4 weeks for 20 weeks. Serum AGT and office BP were measured throughout the trial. Interpretable machine learning was applied to identify clinical and pharmacodynamic predictors of BP-lowering magnitude, durability, and stability, capturing nonlinear and interacting predictors. Hypothesis: Baseline BP level, clinical characteristics, and RAAS biomarker dynamics are associated with magnitude and durability of BP response to TLA. Methods: Participants (N=197) were analyzed using machine learning via gradient boosting and support vector models. Outcomes were defined as early-term change in office systolic BP (4 weeks following initial TLA administration), longer-term change in office systolic BP (20 weeks following initial TLA administration), and control of home systolic BP elevations above 150 mmHg. Consensus predictors were identified using model coefficients and SHapley Additive exPlanations (SHAP). Results: Higher baseline office systolic BP was identified as the strongest predictor of BP lowering 4 weeks following TLA treatment (r = –0.438, p<0.001), demonstrating a continuous relationship with BP lowering ( Figure ). Participant diagnosis of diabetes was also associated with greater BP reduction, with an additional 3 mmHg incremental effect compared with participants without diabetes. An early increase in plasma renin activity (at 4 weeks) was the most significant predictor of sustained BP lowering at 20 weeks following TLA administration. An increase of at least 23% in plasma renin activity was associated with greater durability of BP lowering (85% vs 57% with <23% renin increase, p<0.05). BP lowering was consistent across baseline AGT levels between 0 and 30 μg/mL. Baseline home BP excursion burden and baseline systolic BP predicted improved control of BP excursions. Conclusions: Higher baseline office systolic BP was associated with greater office systolic BP lowering among KARDINAL trial participants administered 90 mg of TLA every 4 weeks. A rise in plasma renin activity was associated with sustained BP lowering, suggesting renin may be a pharmacodynamic marker for BP-lowering response to AGT-directed therapies and reinforces the on-target effect of TLA. Machine learning analysis revealed that readily available clinical features and pharmacodynamic markers help explain response, reinforcing TLA’s mechanism-driven, durable disease-control profile.