AI-Designed Miniproteins Can Recognize Cancer Cells, but the Real Test Comes After Linking Them to a CAR

AI can design proteins on a computer that tightly engage cancer-cell markers, but an experimental gulf remains between “appearing to fit” and “working on immune cells.” A study published in *Nature Communications* established a scalable design and screening workflow and tested it against three cancer-associated surface proteins: PD-L1, CD276 (B7-H3), and VTCN1 (B7-H4). The results showed that AI-designed binding miniproteins must not only capture their targets but also possess biochemical properties suited to practical applications.

The team first used RFdiffusion to generate protein backbones, then designed amino acid sequences and predicted binding structures, selecting miniproteins comprising 70 to 88 amino acids from tens of thousands of computational candidates. Researchers then displayed these candidates on mammalian cells or phages and progressively confirmed their binding capabilities using flow cytometry, next-generation sequencing, and surface plasmon resonance, rather than treating model scores as definitive answers.

Although the three B7-family targets appeared similar, their design difficulty varied considerably. In the first round of mammalian-cell screening, the PD-L1-positive binding population accounted for approximately 2.4%, while both CD276 and VTCN1 were below 1%. Only after expanding the design library did the team identify more CD276 candidates and a small number of VTCN1 candidates. Of 17 PD-L1 miniproteins individually validated, 15 could bind human PD-L1. The three whose dissociation constants were measured had values of approximately 2 to 230 nanomolar, indicating that some candidates had fairly high affinity.

The study also examined whether computational scores could genuinely predict experimental success or failure. Chai-1 interface prediction scores calculated using ESM embedding information showed the strongest association with actual binding results and could identify several mutations that disrupted the binding interface. However, differences among the models were limited, and there were too few effective VTCN1 samples to regard any single score as a reliable universal screening threshold.

The miniproteins could also be assembled into tetravalent fluorescent reagents. In some cancer cell lines, the PD-L1-targeting version detected endogenous protein, with staining intensity and patterns similar to those of conventional antibodies. However, the CD276 reagent produced weak signals at natural expression levels, while the VTCN1 version also showed substantial background. These comparisons indicate that when the same design workflow is applied across different targets, its sensitivity and specificity cannot be assumed to carry over unchanged.

The more critical test came in CAR-T cells. After some high-affinity miniproteins were incorporated into chimeric antigen receptors, they failed to display well on the T-cell surface, limiting their function. The team therefore preserved the CD276-binding interface while rewriting sequences at other positions in the reference miniprotein HM9. Variants with isoelectric points of approximately 6.5 to 8.5 generally showed better surface expression. When these optimized CAR-T cells encountered A375 melanoma cells naturally expressing CD276, they produced stronger interferon-γ and tumor necrosis factor-α responses and increased killing in vitro, while generally maintaining low responses to CD276-knockout cells.

This remains an early methodological finding, not a therapy ready to proceed directly into clinical use. The study did not conduct in vivo validation in animals and has yet to address the immunogenicity of artificial proteins, CAR-T persistence, or the risk of “hitting the target but harming normal tissue.” CD276 itself is also present in non-cancer-cell environments such as tumor vasculature. The study’s most practical message, therefore, is not that AI can already deliver therapeutic molecules automatically, but that design, experimental screening, and functional optimization must be connected into a single pipeline: tight binding is only the starting point.