De Novo Protein Design in Practice: What’s Actually Working (with Michael Holden)

Open-source models and a ~$1,000 GPU have put de novo protein design within reach of almost any lab. But what actually works in practice? In this episode of Protein Engineering in Practice, Leo Wan (RANOMICS) sits down with computational protein designer Michael Holden to cut through the hype: how to choose between models like RFdiffusion, BoltzGen and ProteinMPNN, when to screen with phage vs. yeast display, why confidence scores (ipTM, ipSAE) don’t predict binding affinity, the hardware you actually need, how AI agents are starting to run campaigns, and why “binding is not function.” ▸ Work with RANOMICS: https://www.ranomics.com CHAPTERS 0:00 Welcome & what this show is 4:38 What is de novo design — and why now 13:18 Screening: phage vs. yeast display 22:35 Choosing models (RFdiffusion, BoltzGen, MPNN) 24:08 Negative data: the next frontier 26:48 Do confidence scores predict affinity? 32:27 GPU hardware: what you actually need 38:41 AI agents in the workflow 41:22 Do you need to code? 43:07 Closing the computational–wet-lab gap 49:12 Binding is not function 52:39 Designing enzymes vs. binders 57:30 Wrap-up & resources #proteindesign #denovo #proteinengineering #drugdiscovery #AIforscience