Crowdsourcing functional cell therapy binders from generative artificial intelligence

Over thirty years ago, structural biologists established the Critical Assessment of Structure Prediction (CASP) to rigorously assess methods for predicting protein structures from amino acid sequences. Participants submitted predictions before the corresponding experimental structures became publicly available (Moult et al, 1995). This framework established a common standard against which methodological advances could be measured. Its value became particularly evident at CASP14 in 2020, when AlphaFold2 produced predictions approaching experimental accuracy for many targets (Jumper et al, 2021; Pereira et al, 2021). CASP’s independent, prospective assessment enabled the community to recognize the scale of this advance by providing rapid, independent validation. The next generation of CASP-like open competitions is similarly poised to help discern the impact of emerging artificial intelligence (AI)-based methods across several domains of biology, from genomics to protein design to therapeutic discovery. The implementation and results of one such effort, the Bits to Binders competition, are described in this issue of Molecular Systems Biology, where Kosonocky, Abel and colleagues evaluated 12,000 designed proteins as recognition domains for chimeric antigen receptor (CAR) T cells (Kosonocky et al, 2026). Generative AI is advancing protein binder design. Kosonocky, Abel and colleagues report on the Bits to Binders competition, which tested 12,000 designs in engineered T cells and identified promising candidates for future cell therapies.