Integration of pre-trained GRU and molecular docking for virtual screening of quinoa seed derived ACE inhibitory peptides: An innovative prediction strategy

Recent developments in AI, particularly deep learning, are sparking a revolution in bioactive peptide discovery. This study introduced a pretrained Gated Recurrent Unit model (Pre-GRU) to predict the IC₅₀ values of ACE inhibitory peptides derived from germinated quinoa seed protein. Transfer learning was used to pretrain on ACE inhibitory peptides of adjacent lengths, followed by fine-tuning on the target dataset to address data scarcity. To mitigate label imbalance, the balanced MSE loss was employed for improving prediction accuracy, achieving R²_train: 0.86-0.88 and R²_test: 0.53-0.55. Combining the model-predicted IC₅₀ value with the docking score of identified candidate peptides, four out of seven synthesized peptides showed ACE inhibition activity, particularly NLFRP (IC₅₀ = 3.33 ± 0.19 μM). Molecular docking revealed that NLFRP forms a tetrahedral coordination with ACE's Zn²⁺-binding HEXXH motif, enhancing binding stability. This approach integrated pre-GRU predictions, molecular docking, and physicochemical properties to ensure efficient bioactive peptide screening.