From CRISPR Perturbation Screens Towards Consilient Model-Informed Principles for CAR T-Cell Design 2309679

Introduction Adoptive cell therapies must combine indefinite self-renewal with mature effector functions. Physiologically, cells are rarely simultaneously endowed with these attributes, therefore cellular therapies must, in response to environmental signals, dynamically transition between self-renewing and mature effector states while avoiding dysfunctional states. Modifying regulation via edits of TF and epigenetic modifiers can help but are pleiotropic, while exhaustive experimental discovery of cis regulatory edits (cCREs) is infeasible. Methods We introduce Consilience, a Bayesian deep generative model trained on paired scRNA-seq/scATAC-seq from human T cells. It encodes cell-state dynamics as an SDE in a low-dimensional latent space, predicts time-evolving transcriptome/regulome trajectories, and simulates perturbations to trans factors and specific cCREs with calibrated uncertainty. We also performed pooled CRISPRa screens in anti-CD20 CAR-T cells under exhaustion-provoking conditions. Results CRISPRa hits implicated TGFβ and TNFα modules in durability and nominated MFNG, NFIB, SMYD2, and TRIAP1 in central-memory—biased states - targets where full knockout may harm proliferation/cytotoxicity. Consilience supports (i) reverse screening via cis-tuning at TF loci: editing TCF7 and TOX promoters/enhancers is predicted to increase central/stem-like memory at baseline yet preserve subset-specific effector recruitment upon antigen; and (ii) forward screening via stage-resolved cCRE scans: silencing late-phase elements is predicted to block terminal exhaustion while retaining early protective programs (survival and restraint). The model proposes minimal multi-edit sets optimizing persistence, effector competence, and safety with credible intervals. Conclusion Consilience reframes CAR-T engineering as dynamical systems optimization and prioritizes tractable, minimal cis edits with predicted effect sizes. Refined training on single-cell multiomics data from CAR-T cells will further tailor designs as data mature.