Fig. 1: Overview of TCR-TRANSLATE.

a, Casting antigen-specific TCR design as a seq2seq task. We make use of an encoder–decoder abstraction to process pMHC sequence information and autoregressively sample target-conditioned CDR3β sequences. b, Specific architecture of TCRBART and TCRT5. Transformer architecture juxtaposing BART and T5 encoder and decoder layers, highlighting key operations to the residual stream, inspired by Vaswani et al. (2017)⁶⁰. c, Dataset creation. Given the severe data sparsity, the top-20 pMHCs from IEDB, VDJdb and McPAS (by known TCRs) were withheld as validation, whereas the remainder were used for training with allele-imputed pMHCs from MIRA. d, In silico benchmark performance of TCRT5 and publicly available methods. Overview of benchmark dataset creation (n = 14) and performance radar plot are shown with the averaged metrics across pMHCs. soNNia model’s unconditional metrics are averaged over 1,000 simulation runs. e, Illustrative diagram of the in vitro validation pipeline of the generated CDR3β sequences using NFAT-associated luciferase expression for T cell activation-induced luminescence. Panels a, c, d and e created with BioRender.com.