JURA Bio, Inc., Maryland, United States
Poster stand number: T129
Keywords: cell therapy, cancer, machine learning, personalized medicine, Bayesian inferenceCAR-TCR therapies are a promising class of cell therapies for a diversity of oncologic disease targets. Two major challenges in identifying therapeutically relevant TCR sequences are off-target effects and rejection due to immune intolerance. Current methods for discovering sequences that avoid these effects typically search through experimentally observed patient TCRs only, and are therefore limited in their discovery space; tools for developing semi-personalized candidates, that can avoid off-target effects and rejection in more than one patient, are almost entirely lacking. In this poster we describe a computational algorithm that, based on patient TCR sequencing data as well as any initial binder sequence, designs ultra-large scale combinatorial libraries of personalized or semi-personalized candidate TCR sequences. The designed libraries are both highly diverse and closely matched to patient repertoires. We describe extensive in silico validation of the algorithm, using nonparametric statistical tests for biological sequences and deep learning-based TCR binding predictors, and then outline our future experimental plans. In sum, we demonstrate biologically and statistically-principled machine learning methods for generating personalized and semi-personalized libraries of candidate sequences for CAR-TCR therapy.