Document resource
Background AI models have been developed to better understand disease through histology or protein structure, however these tools cannot model the impact of therapeutics on disease in a cellular context. AI cell models could simulate cellular responses to therapeutics, enabling exploration of vast combinatorial target spaces too large for experiments. However, their use in drug discovery is limited by a lack of large-scale, relevant biological datasets.We introduce the gx1TM model, a foundation model of T cell biology and demonstrate how it can be used for target discovery at a scale beyond wet-lab experiments.Methods The gx1™ model, a transformer-based encoder-decoder model, 1 was trained using masked expression modeling on a dataset of over 75M cells (30M+ T cells from ArsenalBio). To adapt the gx1TM model for target discovery in exhausted T cells, a genetic perturbation dataset of exhausted T cells was generated. We perturbed 551 genes using CRISPRi targeting in T cells from four donors.2 T cell exhaustion was induced in vitro via repeated antigen stimulation over 14 days.3 We measured T cell killing, growth, and performed single-cell RNA sequencing. These data were used to fine-tune a perturbation model. After training, the model’s performance was evaluated on a held-out donor and 50 held out perturbations. Finally, we conducted a virtual screen and validated novel targets with superior predicted function.Results The perturbation model’s predictions were robust to technical covariates with high correlation between true and predicted values in held out donors and held out perturbations ( figure 1a). Using the fine-tuned perturbation model we performed a virtual screen of 182,000 gene combinations and ranked by predicted cytotoxicity and proliferation. We selected 113 novel perturbations for validation. Experimental measures were correlated with predicted values (figure 1b). Similarly, predicted expression changes induced by perturbations were correlated with experimental data (log-fold change over controls Pearson’s r=0.67). This correlation was comparable to correlations between experimental replicates. Importantly, the model predicted gene perturbations that had superior function in comparison to perturbations in the training set were confirmed experimentally.Conclusions The gx1™ model, fine-tuned with perturbation data, accurately predicts T cell function and identifies novel genetic targets for enhancement. This AI-driven approach may significantly accelerate drug discovery by enabling virtual screens that are physically infeasible. More generally, this work suggests that cell foundation models may enable the discovery of new biology and the identification of new targets beyond the reach of physical experiments.References Vaswani, Ashish, et al. ‘Attention is all you need.’ Advances in Neural Information Processing Systems. 2017;30.Ran F, Hsu P, Wright J, et al. Genome engineering using the CRISPR-Cas9 system. Nat Protoc. 2013;8:2281–2308. https://doi.org/10.1038/nprot.2013.143Wherry EJ, Kurachi M. Molecular and cellular insights into T cell exhaustion. Nature Reviews Immunology. 2015;15(8),486–499.Abstract 1125 Figure 1Performance of gx1 based perturbation model. [a] The left panel demonstrates experimental validation of model predictions of log fold change gene expression for each perturbation [b] experimental validation of functional predictions