How AlphaFold Transforms Scientific Research

GenAI-enabled insilico drug design is super hot: just last week we had the $1B funding announcement of Xaira Therapeutics; this week Google DeepMind/Isomorphic Labs (not clear where that boundary is) announced that AlphaFold, which pretty much solved the protein folding problem, has been massively extended to model not just proteins, but also DNA, RNA, and ligands - and the interactions between all those molecular structures. Interesting that the primary enabler of this seems to be the addition of diffusion AI models (that in other applications generate images). These models start with a cloud of digital noise and then incrementally "denoise" that clouds to create molecular structures. (More and more at Ryght we're seeing the utility of repurposing LLMs created for one application in other applications.) Sure feels like Xaira Therapeutics and Isomorphic Labs are on a collision course - both now have GenAI platforms that can (perhaps!) model the biochemistry required for insilico drug design. Here's a recent description of what's coming out of the David Baker's lab, which is the tech engine for Xairia: “Proteins don’t function in isolation,” Baker told Endpoints News. “The fact that AlphaFold and RoseTTAFold only predicted the structure of the protein but not the rest of the system was a limitation.” "The new versions allow researchers to add other biomolecules to the mix, including DNA, RNA, metabolites, drugs and more. Baker’s study, published Thursday in Science, showed how the system could be used to design proteins that bind the oxygen-holding molecule heme and the heart disease drug digoxigenin."

AI-accelerated drug development is no longer a vision of the future, it is a reality. This is a pony express to fiber-optic communication type of advance. Excited to see where we are in another 3 to 5 years. "An artificial intelligence (AI)-powered software program released today by Google DeepMind offers scientists a potent new tool to predict how proteins work. Whereas earlier versions of the company’s software could model how the strands of amino acids making up a protein fold into its final 3D shape, the new version reveals how folded proteins bind and interact with a host of other molecules, including DNA, RNA, and other proteins ... AF3 could correctly model known interactions between proteins and small, druglike molecules in 76% of the more than 400 cases tested, compared with roughly 40% for RoseTTAFold All-Atom. And for interactions between proteins and antibodies, AF3 was correct 62% of the time compared with 30% for AlphaFold Multimer, the company’s previous software package for modeling protein interactions with other biomolecules. Julien Bergeron, a biologist at King’s College London who was given early access to test the new AF3 software, calls it “transformative” in its ability to speed up research. Rather than spend years in the lab studying a protein, they can get a result in minutes."

A couple weeks ago, Google DeepMind released their paper for AlphaFold 3. Just as Iron Man 3 had lower error rates and an expanded feature set over Iron Man 2, er, ok not a good analogy, let me try again: AlphaFold 3 is primarily better for two reasons. First, it can give you a correct prediction much more often (the abstract cites "far greater accuracy"). Second, it can predict more than just protein structure — in fact, it can predict joint structures (structures of multiple molecules) for many kinds of small molecules, expanding beyond proteins. I'll back up for a second and tell you what AlphaFold does, and why that's a good thing. Basically, if we know a sequence of amino acids that form a molecule, then — in theory — we have the data we need to understand what that molecule will look like and behave like. But, in practice, it's a difficult problem to map from amino acid sequence to a list of 3d coordinates for every atom in the folded structure. Imagine taking the shape of a human statue, describing that as a list of angles, and showing someone that list of angles. Could they immediately see it's a human-shaped object? That's the kind of problem AlphaFold 3 solves. Next: Why is that useful? There are many applications; I'll tell you about drug development. Discovering a new drug to treat a disease in insanely complex. First, you need to understand the pathway the disease takes: by this I mean the different steps, such as binding to a receptor on a host cell, or finding a way to multiply, that the disease takes. Once you understand the pathway, you need to find a way to disrupt it. For example, in fighting a virus, you can choose a binding site on the virus, and then try to look for a drug that will fit the site but not interfere with anything else happening in the body. AlphaFold can help the R&D process there in a few ways. It can help researchers determine the 3d structure of a disease agent, which can help in identifying candidate binding sites. AlphaFold can be used to look for candidate drugs by predicting whether or not a new structure will bind at a given target, and whether or not it will be specific to that target (not bind to other sites). Ok, all of that is rather high-level. If you're a machine learning nerd like me, you might enjoy reading more details in this week's Learn & Burn summary: https://lnkd.in/gRH8RKi9