Machine learning applied to Hepatitis B virtual patients suggests prognostic biomarker signatures for stratifying responders to standard-of-care therapies.

In-silico trials, using mathematical models calibrated with clinical data, offer a transformative method to accelerate drug development. We have developed a virtual trial framework for chronic Hepatitis B that accurately simulates clinical protocols, patient characteristics, and endpoints via a mechanistic mathematical model. These in-silico trials successfully captured functional cures with standard-of-care therapies, and the virtual patients allowed for the creation of extensive synthetic datasets of virology biomarker trajectories under treatment stratified by response. Such comprehensive real-world clinical datasets are rare and often limit the training of response-classification models. The synthetic database thus enabled the application of machine learning to predict treatment response based on virology biomarker dynamics. Using a minimalist set of features with mechanistic justification, the machine learning model predicts functional cure in virtual patients with approximately 95% accuracy, identifying responders and non-responders to Hepatitis B standard-of-care therapies. Although the prognostic in-silico biomarker signature for classifying treatment responders must be validated with real world clinical patient data before application, this example underscores the synergistic potential of integrating various complex modelling methodologies to conduct precise in-silico trials and expedite drug development.