Fig. 3: Multitask training promotes degenerate sampling of polyspecific TCRs.
From: Conditional generation of real antigen-specific T cell receptor sequences
a, Diagram showing polyspecific TCRs binding different, unrelated pMHCs juxtaposed against regular TCRs sharing a more conserved cross-reactivity profile. b, Scatter plot of the number of polyspecific generations as a percentage and mean polyspecificity (number of distinct peptides) of the polyspecific TCRs per model is shown. c, Distribution of TCR polyspecificity across the parallel data and model generations. Density plot of cognate peptide counts for polyspecific TCRs aggregated from the combined training and validation set (reference CDR3β sequences) and the model variants per class. d, Venn diagrams of translation overlaps for TCRBART-0 and TCRT5-FT model variants. e, TCRBART-0 and TCRT5-FT sample polyspecific and known binders with higher sequence likelihoods than those of unknown specificities. Discrete heat maps in which rows indicate individual validation pMHCs, columns indicate the translation rank and colour indicates known binding and polyspecificity status shown for TCRBART-0 and TCRT5-FT variants. Colour/intensity reflects the increasing functional utility of a particular translation as the product of its specificity and known binding status. Panel a created with BioRender.com.
