STAMPS: Signal-peptide Transformer for Augmenting Mammalian Protein Secretion
While secretion plays a key role in the diverse applications of cell engineering, to date only a handful of mammalian signal peptides have been characterised in depth and systematic efforts to build novel variants remain sporadic. We present STAMPS (Signal-peptide Transformer for Augmenting Mammalian Protein Secretion), a generative, autoregressive transformer fine-tuned on [~]6,000 mammalian signal peptides to design de novo sequences that modulate and enhance secretion across proteins and hosts. We show that STAMPS can be used to identify candidate signal peptides that outperform the widely used IgG {kappa} light-chain leader (IgKL) when used to secrete EGFP in HEK293T and CHO cells, and hEPO in HEK293T cells. When incorporated in an industrial cell line development framework, STAMPS leads to the generation of signal peptides that yield [~]2.3-fold gain in secretion of a VHH-Fc compared to the internal industrial benchmark in CHO G22 cells, with the same candidates ranking highly in both CHO and HEK293T hosts. Sequence-to-function analysis highlights a longer, strongly hydrophobic core and a tightly positioned cleavage site as drivers of this strong performance. Together, these results establish data-driven, generative design of mammalian signal peptides as a practical route to tune and improve secretion for bioproduction and cell engineering applications.