Rational design of signal peptides (SPs), crucial for efficient therapeutic protein secretion in Chinese hamster ovary (CHO) cells, remains challenging due to their context-dependency activity. To overcome this limitation and enable the discovery of novel high-performance SPs, we developed a high-throughput computational screening pipeline. This approach leverages the deep learning model SignalP 6.0 to screen millions of SP variants derived from diverse mouse/human wild-type libraries and C-region mutants. As a demonstration of its broad applicability, we applied this pipeline to optimize SPs for human serum albumin (HSA) expression. Ranking candidates based on predicted translocation efficiency and cleavage accuracy identified thirty promising SPs outperforming native HSA SP. Experimental validation in CHO cells confirmed multiple novel SPs that significantly enhanced HSA yields, both transiently (e.g., M1_MATN2, 1.93-fold; H5_CXL14, 1.63-fold) and stably (e.g., H5_CXL14, 2.89-fold; M1_MATN2, 1.86-fold). Crucially, our analysis revealed novel insights: hydropathicity profiling uncovered a distinctive and highly effective signature in the top high-performing H5_CXL14 SP, characterized by rapid hydrophobic onset, a continuous highly hydrophobic core, and peak hydrophobicity. Solubility predictions suggested wild-type SPs enhanced secreted protein solubility, while C-region mutants had neutral or negative effects. Furthermore, a novel correlation was observed between high-expression levels and more stable mRNA secondary structures (lower minimum free energy, MFE). This integrated computational-experimental pipeline represents a significant advance, enabling the rational design of protein-specific SP with high efficiency. It drastically reduces the experimental screening burden and holds substantial promise for broadly optimizing therapeutic protein production platforms, as demonstrated here for HSA in CHO cells.