David Baker: Design of New Protein Functions Using Deep Learning

Deep Learning for Protein Design

• 💡 Dr. David Baker's group utilizes deep learning methods to engineer novel proteins, initiating with problem specification and generating protein backbones via RF-diffusion.
• 🔬 Protein MPNN is employed to assign sequences to these backbones, while RoseTTAFold or AlphaFold validate whether the sequences accurately encode the designed structures.
• 🚀 The overall process involves creating synthetic genes, expressing them in bacteria or yeast, and subsequently testing the proteins for their intended function.
• 🧠 RF-diffusion functions similarly to image generation models, having been trained on the Protein Data Bank to denoise structures and produce entirely new protein forms.

Therapeutic Binder Applications

• 🎯 The technology enables the creation of high-affinity binders for various therapeutic targets, including the insulin receptor and TNF receptor, achieving picomolar binding affinities.
• ✅ Designed binders have demonstrated complete protection against LPS-induced inflammation in animal models, often proving more effective than existing treatments like ENBREL.
• 🐍 Novel binders effectively neutralize snake venom toxins and provide animal protection, highlighting their potential for stable, long-shelf-life antivenoms.
• 🦠 This approach also extends to designing binders against pandemic families (e.g., influenza, coronavirus) and toxins such as TCSL and C. diff, with some compounds advancing toward clinical trials.

Advanced Protein Function Design

• 🛠️ Proteins can be engineered with conditional activity, for instance, being caged and only becoming active when a specific target like PDL1 is present.
• 🔄 The platform facilitates targeted degradation by designing proteins that bind both a target and an endocytosis/transcytosis receptor, leading to the target's lysosomal breakdown.
• 🔥 This targeted degradation mechanism can significantly enhance the potency of antagonistic antibodies, transforming them into active degraders rather than mere blockers.
• 💡 The process can be rendered catalytic, allowing a single designed compound to facilitate the turnover of numerous target molecules, such as IGGs.

Designing for Disordered Proteins

• 🧩 The deep learning methodologies are adept at designing binders for completely disordered proteins, a class traditionally challenging for drug discovery efforts.
• 🧠 This includes developing binders for peptides implicated in amyloid diseases like Aβ and tau, demonstrating the capacity to suppress aggregation.
• 🔬 The technology enables the creation of highly specific binders, capable of distinguishing between protein isoforms that vary by only a few residues.

Future Vision and Challenges

• 📈 The overarching vision is to revolutionize medicine with more sophisticated and precise conditional medicines, moving beyond broad immunosuppressants to highly targeted interventions.
• 🧪 Future research priorities encompass developing site-specific proteases to destroy targets and engineering molecular machines, such as synthetic chaperones, for systemic cleanup.
• ⚠️ A significant ongoing challenge is the availability of diverse and meticulously curated training data, particularly for accurate affinity prediction and understanding protein dynamics.