Large-scale AI-guided discovery of antifreeze peptides reveals functional tandem-repeat architectures

Abstract

Despite growing demand for biocompatible materials that regulate ice crystallization in cryopreservation, antifreeze peptide discovery remains constrained by limited candidate sequence diversity and reliance on known protein scaffolds. Here, we establish an AI-guided discovery framework integrating multimodal antifreeze protein prediction, generative sequence expansion and local sequence-pattern analysis. AFPFusionLM combines protein sequence and structure-aware representations, achieving an AUROC of 0.937 and an AUPRC of 0.957 on an independent reference set. Antifreeze-domain-adapted protein sequence generation and multimodel consensus screening of UniProt, glacier metagenomes and generated sequences yielded 87,305 candidate protein records. Large-scale sequence analysis revealed recurring periodic sequence architectures across multiple candidate protein clusters, enabling the identification of short-peptide families with shared repeat features and their natural variants. Experimental evaluation revealed ice recrystallization inhibition activity in multiple peptides from repeat-associated families, with inhibitory effects comparable to or greater than those of reference peptides reported in the literature. Cell-based experiments using a representative peptide-containing formulation also supported its cryoprotective potential under reduced-DMSO conditions. These findings show that large-scale sequence exploration can reveal function-associated peptide sequence architectures, providing a research framework for the systematic discovery and subsequent optimization of antifreeze peptides.