Exploring the rules of chimeric antigen receptor phenotypic output using combinatorial signaling motif libraries and machine learning
Chimeric antigen receptor (CAR) costimulatory domains steer the phenotypic output of therapeutic T cells. In most cases these domains are derived from native immune receptors, composed of signaling motif combinations selected by evolution. To explore if non-natural combinations of signaling motifs could drive novel cell fates of interest, we constructed a library of CARs containing ~2,300 synthetic costimulatory domains, built from combinations of 13 peptide signaling motifs. The library produced CARs driving diverse fate outputs, which were sensitive motif combinations and configurations. Neural networks trained to decode the combinatorial grammar of CAR signaling motifs allowed extraction of key design rules. For example, the non-native combination of TRAF- and PLCg1-binding motifs was found to simultaneously enhance cytotoxicity and stemness, a clinically desirable phenotype associated with effective and durable tumor killing. The neural network accurately predicts that addition of PLCg1-binding motifs improves this phenotype when combined with TRAF-binding motifs, but not when combined with other immune signaling motifs (e.g. PI3K- or Grb2- binding motifs). This work shows how libraries built from the minimal building blocks of signaling, combined with machine learning, can efficiently guide engineering of receptors with desired phenotypes.