RFDiffusion3: A Brief Exploration

Last December, the Institute for Protein Design dropped RFDiffusion3, a protein design model that operates at the level of individual atoms. Before the AIs figure out how to use it to craft mirror life bacteria and kill everyone, I wanted to understand its architecture and do a mini exploration on protein design models. How does it differ from an LLM? What are proteins?Disclaimer: This is what RFDiffusion3 looks like to a curious bio noob. I am not a computational biologist and corrections are welcome! Nothing in this article is health or financial advice etc.ProteinsProteins are chains of amino acids (aka "residues"). An amino acid is a molecule with both amino (NH2) and carboxyl groups (COOH). They're sort of like tiny magnets. Amino groups stick to carboxyl groups to form peptide bonds (C(=O)-N). They're often represented as having one N-terminus (amino group) and one C-terminus (carboxyl group) along with a side chain R that can be any set of atoms (including other carboxyl and amino groups, making cool rings and graphs!). Pictures of amino acids. On the right is a skeletal diagram, kinks are CsTwo amino acids forming a peptide bond via a condensation reaction (water leaves)There are over 500 amino acids found in nature but only 20-22 in the genetic code that create proteins (proteinogenic), and most protein models are trained on this set of 20. You'll often see proteins from this canonical set represented with three letter or one letter codes, e.g. alanine as Ala or A.Some amino acids. Notice how they all start with A and start hashing into increasingly wild single letter codesSince these amino acids and their sidechains have charges and push and pull each other, proteins have an energetically favorable 3D structure they will "fold" into, kind of like those snake toys that can fold into 3D shapes from a 1D sequence. one of my favorite toys as a kidProteins can be described at four levels of structure: primary, secondary, tertiary, and (optionally) quaternary. L