AI-driven design of photosynthetic systems

Designing a Binding Pocket for the Light-Harvesting Molecule Bilin using RFDiffusion All-Atom

Read our detailed tutorial on RFDiffusion All-Atom here →

It’s been five years since AlphaFold shattered the 50-year-old CASP benchmark, revolutionizing our approach to protein structure prediction and design. In the wake of that breakthrough, researchers have expanded its legacy to create novel proteins that nature never envisioned. Today, we stand on the brink of a new era with the introduction of RoseTTAFold and RFDiffusion All-Atom, cutting-edge AlphaFold derivatives that go far beyond modeling protein structures alone. Despite sparser training data than AlphaFold, these new models can accurately design entire biomolecular assemblies - integrating proteins, nucleic acids, and small molecule ligands - paving the way for advances in de novo enzyme design with far-reaching implications for human and planetary health.

To illustrate its transformative potential, we highlight a compelling use case involving the light-harvesting molecule bilin. Bilin is the chromophore of choice for cyanobacteria, some of nature’s most efficient photosynthetic systems, expanding the spectrum of visible light beyond what chlorophyll can capture. Our design task creates a binding pocket for bilin, marking the first building block in constructing complex light-harvesting systems. This task underscores the versatility of RFDiffusion All-Atom, and lays the foundation for AI systems that could drive future innovations in sustainable energy.

RFDiffusion All-Atom can be prompted to design protein-ligand pockets in two ways: 1) by maintaining residues from previously known binding sites, and 2) generating fully de novo designs. We use the first strategy to generate a new bilin-binding protein, while also explaining how fully de novo design can work on Superbio. In our PDB output above, we indeed see that the chosen residue interactions were conserved during our design task, with a new protein backbone generated.

Note: RFDiffusion All-Atom generates a protein backbone based on 3D coordinates, composed of mostly Ala and Ser residues. Next time, we’ll demonstrate how to populate the amino acid residues and generate a fully-functional protein chain, as well as some in silico strategies for de-risking target ligand binding.

Want to see another model or tutorial form the Superbio team? Let us know by hitting reply →

Until next time 💗,

The Superbio Team