Discovery of Small-Molecule Cryoprotectants Using Pretrained Language Model Representations

Cryopreservation supports cell therapy, transfusion medicine, and tissue banking, but cryoprotectant selection remains constrained by empirical screening, toxicity, and ice recrystallization. We used pretrained language models as SMILES sequence encoders to prioritize small-molecule cryoprotectants for experimental testing. Hierarchical representations were coupled to descriptor reconstruction, gated fusion, and downstream prediction of normalized mean grain size and thresholded labels. Splat-cooling measurements assessed candidate ice recrystallization inhibition, followed by compatibility and cryopreservation experiments in red blood cells, human umbilical cord-derived mesenchymal stem cells and rat abdominal aortas. Selected additives improved recovery or preservation endpoints within the tested formulations. This representation-based screening workflow connects molecular property prediction with experimental cryobiology across ice, cellular, and tissue endpoints.