Cell surface display has become one of biotechnology’s most useful strategies for controlling what cells present to their surroundings. By placing a protein, peptide or antigen on the exterior of a living cell, researchers can create systems for immune-cell activation, targeted therapies, vaccine development and synthetic biology. Yet a basic problem has limited the field: scientists have had few reliable rules for predicting how a particular sequence will perform. A new study introduces DeepSCan, a collection of artificial intelligence models designed to connect the amino-acid sequence of a cell surface display element with its ability to move a linked protein to the cell membrane. The work, reported in Nature Biotechnology, combines large-scale experimental measurements with machine learning to identify naturally occurring sequences that work especially well and to design new ones from scratch.
Cell surface display elements, often called CSDs, act as molecular trafficking instructions. When genetically attached to an antigen or another protein, they can influence whether the resulting construct is correctly folded, transported through the secretory pathway and ultimately exposed on the outside of the cell. These processes involve passage through intracellular compartments such as the endoplasmic reticulum and Golgi apparatus, followed by delivery to the plasma membrane. A CSD that performs poorly may cause the antigen to remain inside the cell, become degraded or appear on the surface at levels too low to trigger an effective biological response. Conversely, a potent CSD can substantially increase the amount of antigen available for recognition by antibodies or engineered immune cells. Despite their importance, CSDs have generally been selected through trial and error rather than through a systematic sequence-based design framework.
To build that framework, Fang, Saskin, Lee and colleagues generated a large experimental dataset based on more than 570 chimeric antigens. In these constructs, different CSD sequences were connected to a common antigen or protein cargo, allowing the researchers to compare how strongly each sequence promoted surface localization. The team used these measurements to derive translocation-strength labels for approximately 310 CSD elements. These labels provided the central training signal for DeepSCan: instead of simply asking whether a protein reached the cell surface, the models were trained to estimate relative potency and distinguish weak, moderate and highly effective display elements. The researchers also assembled roughly 45 independent training datasets, creating a broader information base for model development and evaluation.
DeepSCan was developed through three generations of models, reflecting an iterative effort to improve prediction quality and experimental usefulness. Deep learning systems can examine patterns across biological sequences that are difficult to identify by manual inspection, including combinations of residues, spacing relationships and broader sequence features. In this context, the models were not merely searching for one universal motif. They were learning how different sequence architectures could influence membrane trafficking, secretion and surface retention. The goal was to capture conserved functional principles while allowing for diversity among effective CSDs. Such a strategy is particularly valuable because sequences with similar activity may not be obviously related at the level of simple alignments, while small changes in a functional trafficking signal can have large effects on the final amount of surface-displayed protein.
The researchers then used the models as design tools rather than as passive predictors. DeepSCan computationally generated approximately 3,700 candidate CSD sequences, expanding the search beyond the limited set of elements found in nature. Around 120 of these generative designs were experimentally tested. This experimental step was essential because computational scores cannot fully capture the complexity of a living cell. The performance of a CSD can depend on the identity of the attached antigen, the expression system, the cell type and the interaction between the engineered sequence and the host’s own trafficking machinery. By repeatedly moving between prediction and laboratory validation, the researchers were able to identify designs that retained strong activity under real biological conditions.
Seven generative CSDs ultimately matched or exceeded the cell surface translocation strength of the most potent naturally occurring elements examined in the study. The result suggests that sequence-based artificial intelligence can do more than classify existing biological parts: it can help create new modules with useful functions. The strongest designs are important not only because they increase display levels, but also because they may offer alternatives to naturally derived sequences that have limitations in size, compatibility or performance. In an engineering setting, having a larger library of potent CSDs could allow researchers to select elements optimized for different antigens and host cells instead of relying on a small collection of standard signals.
The study also examined whether the enhanced display effects were restricted to a single experimental context. According to the researchers, improved surface presentation was observed across multiple cell types, indicating that the best-performing elements were not entirely dependent on one cellular background. This kind of portability is a major consideration for therapeutic development. A trafficking signal that functions in one laboratory cell line may behave differently in primary immune cells or in cells used for manufacturing. Demonstrating activity across distinct cell types provides an early indication that the engineered CSDs could be useful in broader applications, although further testing would be required to establish their behavior in clinically relevant systems and in living organisms.
The researchers additionally tested the biological consequence of improved display in an antigen-specific CAR-T cytotoxicity assay. CAR-T cells are genetically engineered immune cells whose chimeric antigen receptors recognize a selected target and activate killing responses when that target is present. In this experiment, cells displaying antigens through enhanced CSDs were evaluated for their ability to stimulate antigen-specific CAR-T activity. Stronger surface presentation can increase the number of antigen molecules available for receptor engagement, potentially improving the clarity and strength of the immune-cell response. The assay therefore connected the molecular performance of the CSDs with a functional immune outcome, although it does not by itself establish therapeutic efficacy or safety.
DeepSCan could become a general platform for engineering cell-surface interfaces in vaccines, immunotherapies and synthetic biology. For mRNA antigen display, for example, a potent CSD could help direct newly produced antigen to the outer membrane of cells after delivery of the mRNA, where it could be more readily detected by immune receptors. The same principle could support screening systems, cellular biosensors and engineered tissues that need to present defined proteins externally. The study also illustrates a broader shift in biological design: large experimental datasets can be used to train models, models can propose new sequences, and laboratory testing can feed the results back into the next design cycle. The researchers’ findings suggest that this combination may turn cell surface trafficking from a largely empirical process into a more predictable engineering discipline, while underscoring the need for continued validation across cargos, cell types and disease-relevant settings.
Subject of Research: Artificial intelligence-guided discovery and design of potent cell surface display elements for mRNA antigen display and immune engineering.
Article Title: Discovery and design of potent cell surface display elements
Article References: Fang, Z., Saskin, J., Lee, S. H. et al. “Discovery and design of potent cell surface display elements.” Nature Biotechnology (2026). https://doi.org/10.1038/s41587-026-03144-x
Image Credits: AI Generated
DOI: https://doi.org/10.1038/s41587-026-03144-x
Keywords: DeepSCan, cell surface display, CSD elements, artificial intelligence, deep learning, mRNA antigen display, protein trafficking, synthetic biology, CAR-T cells, immunotherapy, antigen presentation, generative protein design
Tags: artificial intelligence in biotechnologycell surface display technologyDeepSCan AI model for protein designdevelopment of cell surface display elementsimmune cell activation strategiesmachine learning for protein sequence predictionnatural and synthetic cell surface display sequencesprotein engineering for cell surface displayprotein trafficking and membrane transportsynthetic biology and cell surface proteinstargeted cell therapiesvaccine development using cell display
