#33 · Structural bio & protein design

AI-designed peptides and engineered nucleases assessed by biological readouts

The dominant thread is coupling learned models to experimental evaluation of their outputs. A mammalian high-throughput screen tests protein-language-model-designed peptide guides fused to a CHIP E3 ligase domain and selects degraders of beta-catenin and GFAP. EvoMax applies Gaussian process regression, protein language models and inverse folding in an adaptive loop to engineer compact Fz2 nucleases from sparse data. The CRISPRi study is primarily a dataset and scoring effort, generating large tiling datasets, comparing KRAB systems, and building an on-target scoring scheme with seed-based off-target analysis. What is new relative to generation alone is that candidates are assessed through screening or model-guided evolution, so design performance is measured against biological readouts rather than predicted alone.