From In-Silico to Wet-Lab: Evaluating AI Protein Design Performance

This tutorial evaluates AI-designed miniprotein binders by comparing computational predictions with real experimental results from two independent labs. The dataset contains over 1,400 AI-designed proteins tested against multiple targets, allowing researchers to assess how well structure predictors identify successful binders and whether computational signals can reliably predict whether designs will work in practice. The analysis examines hit rates across different design models, campaigns, and targets, revealing that target choice has a larger effect on success rates than the choice of design method. Understanding this gap between computational design and wet-lab validation matters because it shows which aspects of AI protein design translate reliably into functional molecules.