TwinRL-Onco: A World Model-Empowered Digital Twin Framework with Hierarchical Reinforcement Learning for Venetoclax Resistance Trajectory Prediction and Adaptive Therapy Optimization in Chronic Lymphocytic Leukemia.

Abstract

Chronic lymphocytic leukaemia (CLL) presents considerable therapeutic obstacles due to the development of treatment-resistant disease, especially with BCL-2 inhibitors like venetoclax, and is among the most challenging haematological malignancies to treat. A critical unmet computational need is the ability to predict patient-specific treatment resistance trajectories and to develop optimal adaptive therapy schedules based on a patient's history and experience with treatments. To address these issues, we have developed TwinRL-Onco, a novel computational framework that fuses digital twin technology, driven by world model-based simulations, with hierarchical reinforcement learning (HRL), to predict in silico treatment response and optimize the adaptive therapy process. TwinRL-Onco consists of three synergistic components: (1) a Variational Recurrent State-Space Model (VRSSM); (2) a hierarchical policy structure with two levels of control; and (3) a resistance trajectory predictor that uses Monte Carlo rollouts generated from the world model to simulate treatment response. TwinRL-Onco achieved an AUROC of 0.917 for 12-month simulated prediction of progression, an improvement of 4.6 percentage points over the DreamerV3 algorithm and 7.0 percentage points over the LSTM-Traj algorithm. TwinRL-Onco is an emphasis on an artificial intelligence-based online environment for hypothesis generation; all on-site metrics indicate inventory-only consistency and need actual clinical validation before any conclusions regarding clinical prediction.