Bringing Antigens to the Cell Surface: DeepSCan Uses AI to Design New Display Modules

Whether cells can deliver artificially designed antigens to the correct location often determines whether an mRNA vaccine or cell-engineering system can function effectively. A study published in *Nature Biotechnology* presents the DeepSCan platform, which uses deep learning to design “cell-surface display modules” that help cell-produced antigens reach the cell membrane more efficiently for recognition by immune cells.

These modules typically contain single-pass transmembrane protein sequences and can be viewed as components that connect antigens to the cell membrane. Although many natural sequences exist, there is no reliable set of rules for determining from amino acid arrangements whether a protein will successfully move to the cell surface, adopt the correct orientation, or achieve high expression levels. Developers therefore often have to rely on screening candidates one by one.

The research team measured the surface expression of more than 570 chimeric antigens, established cell-surface translocation strength labels for approximately 310 display modules, and compiled about 45 independent training datasets. Three generations of DeepSCan models used these data to learn associations between sequence and function, and were then applied to predict natural modules and screen newly generated artificial sequences.

Among 3,700 computer-designed candidates, the researchers selected approximately 120 for experimental testing and ultimately identified 7 artificial modules whose cell-surface translocation capacity matched or exceeded that of the strongest-performing natural module in the study. The better candidates were effective not only in a single cell line but also maintained stronger display across multiple cell types, indicating that the results were not entirely confined to one experimental environment.

The team further linked antigen display to recognition by CAR-T cells. In antigen-specific cytotoxicity assays, cells using enhanced display modules elicited the corresponding CAR-T cell killing response. This result demonstrates that the new sequences do more than increase flow cytometry readings: the antigens displayed on the membrane also retain their ability to be recognized by immune cells. However, these remain in vitro functional tests and cannot be equated with efficacy in animals or therapeutic outcomes in humans.

DeepSCan’s training, inference, and analysis code has been released through public repositories. The research team also provides online tools for the Genesis Quant and Omni models, which can accept protein sequences up to 1,023 amino acids long and output information including surface-expression scores, membrane topology, and transmembrane segments. This gives external researchers a way to reproduce and test the platform, but whether the models remain accurate for antigens outside the training distribution, different delivery methods, and human cellular environments still requires independent validation.

The more immediate value of this work is that it narrows the screening of display modules—which previously required extensive experimentation—to a more manageable set of candidates. To advance toward therapeutic applications, further studies will need to confirm the immunogenicity and long-term stability of the artificial sequences, their expression in different tissues, and their safety when paired with specific mRNA vectors. The current evidence supports a design method validated in cell-based experiments, not a product with demonstrated clinical benefit.