Yale University researchers developed DeepSCan, a suite of deep learning models designed to map sequence-function relationships and engineer de novo cell surface display elements for mRNA antigen display, according to a study published in Nature Biotechnology.
The research team measured surface expression across more than 570 chimeric antigens to establish functional data. From these experiments, scientists generated cell surface translocation strength labels for roughly 310 cell surface display elements and assembled approximately 45 independent datasets to train three generations of DeepSCan neural network models.
Guided by the models, the authors computationally designed 3,700 generative display elements and experimentally evaluated around 120 candidates. Seven of these synthetic modules matched or exceeded the cell surface translocation strength of the most potent natural counterparts. The engineered elements sustained activity across multiple cell types and functioned successfully in an antigen-specific CAR-T cytotoxicity assay.
Model benchmarking revealed that high translocation potency correlates with specific structural sequence features, including higher basic residue frequencies at the interface between the transmembrane and intracellular domains. The computational pipeline integrates ESM2 protein language models, XGBoost, and classical machine learning models trained on graphics processing units using Python 3.11.5 and PyTorch 2.5.1.
Yale University filed a patent application based on an invention disclosure related to the platform. The authors made the DeepSCan code, Docker images, and model weights accessible through GitHub, Zenodo, and Hugging Face.
