Candidate protein molecules designed by deep learning methods represented by AlphaFold and RFdiffusion typically have a hit rate of only 10–15% in wet lab experiments.