"Machine learning in chemistry and beyond" (ChE-651) seminar by Maria Rodriguez Martinez "AI-driven modelling of T cell-based therapies"

T cells are major components of the adaptive immune system with primary roles in killing infected abnormal cells. Because of their cytotoxic role, T cells are being actively investigated in several cancer immunotherapy approaches. T cells selectively recognise foreign antigens through their T cell receptors (TCRs). TCR sequencing technologies have generated millions of TCR sequences and have fuelled the development of AI models to predict the binding properties of TCRs from their sequences. However, while some models achieve good accuracy at predicting TCR binding to a small set of epitopes, the binding prediction to unseen epitopes, remains a challenge. In parallel, the black-box nature of many of these models has resulted in a limited understanding of the rules that govern TCR binding, preventing us from learning how to engineer receptors of improved affinity.
In this talk, I will present recent efforts done at IBM Research - Zurich to develop AI models to understand TCR - epitope binding. First, I will introduce TITAN, a multimodal neural network that allows to study independently the generalisation capabilities to unseen TCRs and epitopes. TITAN exhibits significantly improved performance on unseen epitopes and, importantly, utilises an interpretable context attention mechanism that selectively highlights relevant amino acids. Second, I will describe DECODE, an easy-to-use computational pipeline to extract the binding rules from any black-box model designed to predict the TCR-epitope binding. DECODE offers a range of analytical and visualisation tools to guide the user in the extraction of such rules. Finally, i will discuss how AI models can facilitate the investigation of current immunotherapeutic challenges, such as cross-reactive events due to off-target TCR binding.