This study constructed a salty peptide prediction model using a dual-branch fusion strategy that integrates Evolutionary Scale Modeling Cambrian (ESMC) sequence semantic representations with APAAC global physicochemical features. A cross-attention mechanism was incorporated to identify the amino acid residues contributing most to the model's predictions. Locust proteins were used as the study subject； in silico digestion combined with bioinformatics methods was employed to construct a library of salty peptides， and the developed model was used to identify novel salty peptides. The interaction mechanisms between the salty peptides and the transmembrane channel-like 4 (TMC4) receptor were further elucidated through molecular docking. The model exhibited excellent performance and robustness in the salty peptide prediction task， achieving accuracy， precision， recall， F1 score， and AUC of 0.929， 0.939， 0.925， 0.927， and 0.960， respectively. Arginine， methionine， and glutamic acid were identified as key residues for the recognition of salty taste. A total of 46 salty peptides were identified from locust-derived proteins through in silico digestion， followed by screening for aqueous solubility， toxicity， and salt taste potential. Molecular docking analyses indicated that all screened peptides were successfully docked into an internal binding pocket of the TMC4. Cys475， Gln377， Glu235， Gly242， and Leu618 were identified as key amino acid residues involved in the interactions， and hydrogen bonds and hydrophobic interactions were determined to be the primary forces driving the binding of peptide and TMC4. This study established a high-throughput screening framework for the prediction and mechanistic elucidation of salty peptides by integrating deep learning， in silico enzymatic hydrolysis， and molecular docking. This approach provided a methodological foundation and candidate molecules for the identification of salty peptides derived from locust protein for the application in low-sodium foods.