Broske, B.; Binder, S. C.; McEnroe, B. A.; Kempchen, T. N.; Fandrey, C. I.; Messmer, J. M.; Tan, E.; Konopka, P.; Ferber, D.; Yong, M. C. R.; Kleinert, M.; Hoch, A.; Blumenstock, K.; Tödtmann, J. M. P.; Oldenburg, J.; Rühl, H.; Semaan, A.; Toma, M. I.; Markova, K.; Kobold, S.; Rollenske, T.; Geyer, M.; Menzel, S.; Bald, T.; Schmid-Burgk, J. L.; Hagelueken, G.; Hölzel, M.

Artificial intelligence (AI)-based protein design opens new avenues for the rapid generation of new research tools and therapeutics, but experimental validation lags behind the computational design throughput. Here, we present a scalable workflow for the discovery and validation of AI-designed protein binders. Leveraging the RFdiffusion protein design pipeline with a custom filter for stable alpha-helical bundle folds, we construct libraries of thousands of AI-binders against cancer-associated surface proteins. Mammalian cell-surface and phage display screening yield multiple high-affinity PD-L1 binders but fewer hits for CD276 (B7-H3) and VTCN1 (B7-H4), reflecting the target-dependent efficiency of RFdiffusion in generating high-quality designs. Using our experimentally validated AI-designed binder libraries, we benchmark freely available structure prediction models. We find that interface predicted template modelling (ipTM) scores by Chai-1 with ESM embedding correlate well with experimental success and even predict deleterious effects of binding interface mutations. To demonstrate the versatility of AI-binders as research tools, we deploy them in CAR-T cells and also assemble them with fluorophore-labeled streptavidin into tetravalent quattrobinders, which achieve antibody-comparable staining of endogenous PD-L1 by flow cytometry. With high production yields and accessible structural models, AI-designed quattrobinders are versatile and cost-effective research tools amenable to community-driven validation and optimization.