Boosting the Predictive Power of Virus-Specific T-Cell Clinical Trials via Generative Models

The modest number of participants enrolled in phase I/II clinical trials limits the development of accurate models to predict treatment response. This is evidenced by early-phase clinical trials evaluating partially human leukocyte antigen-matched virus-specific T cells (VSTs) to treat virus infections in immunocompromised persons. In silico models for VST response are lacking due to the small number of participants enrolled in these clinical trials. Here, we developed a generative artificial intelligence approach to overcoming the small n and then trained a binary classifier to predict the likelihood of response to third-party VST therapy. We generated synthetic patient-response training data using a variational autoencoder, whose presence improves the robustness of machine learning models. The binary classification model determines patient response with high accuracy, evidenced by cross-validation using data from 3 third-party VST trials. The combined variational autoencoder and binary classifier provide a potential solution when considering patient suitability for third-party VSTs and also offer a broadly applicable framework to enhance the performance of machine learning models using small-scale clinical datasets.