Design of new protein functions using deep learning–David Baker (University of Washington)

Designing New Protein Functions with Deep Learning

• 💡 David Baker's lab uses generative AI methods, specifically diffusion models, to design novel proteins from scratch.
• 🎯 The core idea is to condition the generative process to create proteins with specific desired functions, moving beyond naturally evolved proteins.
• 🔑 This approach allows for the creation of proteins that address modern-day problems not solved by natural evolution.

Medical and Therapeutic Innovations

• 💊 Proteins are designed to bind specific targets, such as the insulin receptor (creating insulin mimics) and the TNF receptor (for anti-inflammatory effects).
• 🧠 New proteins can target disordered proteins like Tau, involved in neurodegenerative diseases, for destruction.
• ✅ The technology also enables the design of cancer immunotherapies and on-off switches for controlled drug activation.

Advanced Sensing and Material Applications

• 🔬 Designed proteins can form nanopores in membranes with controlled sizes, enabling applications like electronic noses for detecting specific compounds.
• ⚡ Proteins can be inserted into silicon nitride for homogeneous nanopores, useful for DNA and protein sequencing.
• 🏗️ The lab designs proteins to template mineral deposition (e.g., calcium phosphate, zinc oxide), opening doors for new hybrid materials with atomic-level control.

Molecular Machines and Catalysis

• ⚙️ Proteins are engineered as switch-like molecules that can toggle between states (e.g., with light), with potential for ultra-low-power computing in 3D crystal arrays.
• 🌱 Efforts include creating catalysts for green chemistry and designing proteases that efficiently break down chemical bonds, including plastics and disease-related proteins like A-beta and Tau.
• ☀️ Proteins are being designed to mimic photosynthesis, capturing light energy for chemical catalysis.

Methodological Advances and Future Directions

• 🚀 Recent advancements include atom-level diffusion for more precise control in protein generation, particularly for catalysis.
• 📊 Active learning approaches combine computational design with rapid experimental feedback to accelerate the discovery of highly active proteins, such as plastic-degrading enzymes.
• 🧠 The team is developing general models of cell function by using self-distillation on vast metagenome sequence data to improve structure prediction (RosettaFold-3) and understand complex biological systems.