Antifreeze peptides break free of known scaffolds: 87,305 candidates yield tandem-repeat families whose short peptides match or beat a literature reference

The candidate pool for antifreeze peptides shifts from truncating and substituting around known antifreeze proteins, conserved ice-binding motifs or structural templates to a unified screen over three complementary sources: public protein databases, polar metagenomes and generated sequences. Conventional routes design precisely within local sequence neighbourhoods and cannot readily cover the broad diversity of natural proteins, and coordinated exploration across multiple sequence resources has been limited. Three-model consensus screening retained 87,305 candidate protein records: 16,746 from UniProt, 38,611 from glacier metagenomes and 31,948 from the generative model, or 19.2%, 44.2% and 36.6% of the collection.

The predictor behind the screen judges whether a sequence carries antifreeze-associated properties by combining protein sequence representations with structure-aware representations. Single-modality models each produce their own call, and the fusion scores higher accuracy than either alone. On the independent gold-standard test set AFPFusionLM reached an AUROC of 0.937 and an AUPRC of 0.957, above AFP-LSE at 0.924 and 0.938 and PLM-ICE at 0.904 and 0.917; fusion raised accuracy from 0.861 for ESM2 and 0.868 for SaProt to 0.907.

The generative model learns the antifreeze-associated sequence distribution while updating only a small parameter set, adding sequences absent from natural databases to the candidate collection. A frozen backbone or soft prompting adapts with larger sequence-modelling error, while full fine-tuning updates the entire backbone. LoRA lowered validation and internal-test negative log-likelihood per residue to 2.17 and 2.12, close to full fine-tuning at 2.18 and 2.12, while training about 11.8 million parameters, roughly 1.5% of the full fine-tuning trainable count.

Large-scale sequence comparison turns local repeats dispersed among individual proteins into periodic architectures recurring across multiple candidate clusters, and organizes them into short-peptide families that can be tested experimentally. Earlier descriptions of structure-function relationships stayed within particular molecular scaffolds and their neighbouring sequence space. In the completed ice recrystallization inhibition screen, 6 of 17 non-control peptide candidates had relative ice-crystal areas at or below the 61.35% measured for a literature reference peptide in the same batch; 5 of 6 candidates from one repeat-associated family reached that level, with a family median of 55.88%, and another representative candidate measured 30.26%, a 69.74% reduction relative to PBS.

The demonstrated scope is computational screening plus experimental validation: the predictor is evaluated on random partitions, homology-disjoint ten-fold splits and an independent gold-standard test set, candidates come from UniProt, glacier metagenomes and generated sequences, and functional evidence comes from ice recrystallization inhibition assays, 50-ns ice-water molecular dynamics simulations and cell experiments with a representative peptide-containing formulation.

Ice recrystallization inhibition and interfacial simulations were compared descriptively without between-group significance tests; water molecules and frames from the same trajectory were treated as subsamples of their measurements; and generated-sequence novelty is defined against a specified reference collection at a sequence-identity threshold.