AlphaProteo generates novel proteins for biology and health research

i studied molecular biology and i couldn't help contain my excitement when it was able to bind to another protein. I dont think HN realizes how huge this is. With this level of accuracy, not only can we understand the full mysteries of ourselves but literally any biological entity.

With that level of understanding, its easy to fabricate special medicines that target specific biochem pathways, but more exciting is that we can literally "code in 3d world". We'll be able to print and grow organs in mass. We'll be able to design structures that will bind to target proteins responsible for certain traits. The potential boon to human medicine will be enormous.

I got like goosebumps after watching that video because I understood the implications of being able to predict folds and now generate proteins that will bind to any protein we choose!!!!

We just might have discovered a panacea of sorts and Demis and his team should receive the Nobel Prize.

I'm just ecstatic that we'll see so much drastic improvement in human medicine and importantly how accessible they will be with this new discovery.

That channel is just hype machine. Like half the stuff he says is barely comprehensible if you know the actual science.

Cynical take - mostly to appear like a diverse tech company with lots of different products and services so they don't get regulated for their strength in the search advertising market.

IIUC most of the commercialization is done through Isomorphic (https://www.isomorphiclabs.com/). My guess is that Google Research/DM itself wants to stay at the front of the field rather than develop drugs (of which protein design is really just a tiny contribution).

When I worked at Google I made a case for doing protein design/preliminary drug discovery using Google infrastructure and it was well received by the leadership. The leadership at Google is mostly computer scientists who know about, but can't actually do, leading-edge life sciences research, and they want to contribute some amount of Google's resources to advancing the state of the art. That's the only reason exacycle was permitted- because Urs thought we could maybe help save the world with protein design (and it wasn't a good approach because it wasted enormous amounts of power on unbiased sampling of large proteins).

Honestly I don't think Google proper is really a good place for this work to be applied, though. Their attention is easily diverted, they repeatedly fail to commercialize, and most importantly, potential partners are scared Google will steal their data, and replace their business.

I know that you can actually use AlphaFold at least. My wife, microbiologist, told me she's used it a couple of times at work. I don't know what their monetization model is, if her lab had to get a license or anything. But I know scientists are using it.

I play Go recreationally. I don't think I can use AlphaGo (or its successors) directly, but the published research on AlphaGo has inspired other strong Go AIs. Online Go platforms integrate them to offer AI matches as well as analyze games between humans. I also know that professionally ranked players are adopting things learned from AI into their own play, and a lot of traditional joseki (analogous to chess openings) are being rethought based on insights from AI play.

They penned a 3B deal with Novartis and Eli Lily I believe

Alphabet has a medical division - https://en.wikipedia.org/wiki/Verily . My somewhat cynical take is that most extremely wealthy individuals would like to live longer.

But more immediately this is an interesting and relevant problem to solve, so it servers as a tool to benchmark and improve AI... and the current theory is that at some point AI will {generate unlimited amount of wealth | lead humanity into the post-scarcity society | solve all human problems by eliminating humans}.

If the protein is a novel it does not matter, because it has no normal variants in the nature.

It would be essentially impossible to create a new prion disease by accident- generating random-ish new things with methods like this would pale in comparison to the massive number of weird random-ish things natural biology is already creating in the wild.

However, this category of technologies could potentially be used to develop new prion diseases on purpose. As well as to develop cures for prion diseases that disrupt the misfolding.

Oh this would be 100x worse than the covid lab leak

The context here is that prions are misfolded proteins that replicate by causing other proteins to change their configuration into the misfolded form of the prion. Diseases caused by prions include Mad Cow disease, Creutzfeldt-Jakob disease, and Chronic Wasting disease. All prion diseases are incurable and 100% fatal.

For anyone who would like to know more about designing proteins with a certain function, target, or structure in mind, the term to search for is "rational design."

https://en.m.wikipedia.org/wiki/Rational_design

Not an expert, but you could imagine a protein with two receptors that are required for activation. One of them binds to a protein that is only present in the cells of interest, and the other one binds to the actual target.

they can generate proteins that bind to specific structures with high accuracy, achieving true cell-specificity and avoiding unwanted pleiotropic effects involves many more variables beyond just protein-protein interactions. These tools are more about expanding our ability to target previously "undruggable" proteins rather than solving the cell-specificity problem outright. however they could be valuable components in developing more targeted therapies when combined with comprehensive research on pleiotropic effects across multiple omics levels. real breakthrough will come from integrating these protein design capabilities with a deeper understanding of complex biological systems and developing strategies for precise delivery and regulation of these novel proteins in vivo.

This research is focused on modeling individual protein binding sites. Pleiotropic effects and off-target side effects are caused by interactions beyond the individual binding sites. So I don't think this tool by itself will be able to design a protein that acts in the way you describe (and that's putting aside the delivery concerns - how do you get the protein to the right compartment inside the cell?).

But novel binding domain design could be combined with other tools to achieve this effect. You could imagine engineering a lipid nanoparticle coated in antibodies specific to cell types that express particular surface proteins. So you might use this tool to design both the antibody binding domain on the vector and also the protein encoded by the payload mRNA. Not all cell types can be reached and addressed this way, but many can.

Yes, in principle but there are huge limitations and challenges to using a protein as a drug in living organisms. It has to be injected to avoid digestion, and a protein can't just pass into a cell, it needs to get in somehow. Current peptide drugs like insulin are identical to, or closely mimic natural small peptide hormones that bind to receptors on the outside of a cell. However, there is a possibility of using gene therapy to directly express a novel protein drug inside of the cell. A novel protein is also likely to trigger an immune response- so that type of gene therapy is mostly useful when that is actually desired, e.g. as a vaccine.

Making proteins is nothing like designing life or viruses. It's barely even related.

They must be somewhat novel in that the wet lab work verified up to 10x stronger binding as predicted. I agree it would be interesting to see how they compare to known binding proteins

Might depend on what your measure of 'novelty' is in protein structure. A single residue change (for example) would not normally be considered a novel structure - it's just a mutation.

However, a new fold - that is, the shape that the backbone folds into - would be novel. Potentially also novel would be 'chimeric' structures with parts from other structures, as with chimeric domain swaps.

There was a structure designed by the Baker lab called 'Top7 - https://pubmed.ncbi.nlm.nih.gov/14631033/ that I remember as ground breaking at the time :) (in the ancient days of 2003 it seems ...)

My guess is that this area is much harder to break into–enzymes facilitate challenging chemical transformations by stabilizing high-energy transition states in chemical reactions. These states are usually highly transient and therefore much harder to capture using the structural biology techniques that generate the structural data that AlphaFold and similar methods are trained on. Even though there are many structures of enzymes in the absence of their substrate, I would imagine that the small number of structures for states that represent actual catalytic intermediates would make it difficult for a model to internalize the features that distinguish a good enzyme/catalyst from a bad one.

Another consideration is that most protein structure prediction methods only generate the backbone, and the sidechains are modeled in afterwards. Enzyme efficiency requires sub-A level structural precision in the sidechains that are actually doing the chemistry involved in catalysis, so it could also be the case that the current backbone-centric methods aren't good enough to predict these fine-tuned interactions.

I want to mention an interesting industrial enzyme project. If you ever saw the laundry detergent commercial "Protein gets out protein", this is referring to an industrial enzyme in laundry detergent. Many years ago, Genentech had built up a significant capability in proteases, which are proteins that cut other proteins into pieces. In the course of optimizing proteases, they made a thermostable, thermoactive protease. Although it wasn't super useful for Genentech in a drug discovery context, it was recognized that you could put an inactive enzyme into laundry detergent that would be activated when the hot laundry water hit the detergent, and the resulting protease would be good at cleaning stains (many stains are composed of protein- blood, food, etc).

Genentech set up a subsidiary with Corning (the glass company) that owns the IP for this protease and then licensed it to laundry detergent manufacturers; many billions of dollars in revenue. I think this is one of the original patents: https://patentimages.storage.googleapis.com/d9/ca/6f/2fb89ff...

Interested observer here, not an expert: My understanding is that they are using another model called FermiNet for chemistry research https://deepmind.google/discover/blog/ferminet-quantum-physi...

Terrible take. The article details independent lab verification with researchers listed by name and a quote from them.

Making a comment without reading the article? Who would do such a thing?

Agreed on the hopes that these methods lead to novel biocatalysts (but they aren’t quite there yet).

David Baker’s lab has recently published on using their own diffusion model (RFdiffusion) to design novel biocatalysts that perform hydrolysis using a catalytic triad of serine, aspartic acid, and histidine, as well as an oxyanion hole, which is much more complex than the binders designed by AlphaProteo [1].

It gives me hope that we’ll soon be able to design biocatalysts as good as natural ones, but for any problem we care about.

1. https://alexcarlin.bearblog.dev/novel-enzymes-from-a-diffusi...

That video is saying that AlphaFold isn't good at predicting how proteins bind to each other, but is actually pretty good at predicting the structure of the individual proteins. Which is exactly what you would expect, because AlphaFold wasn't designed to predict antibody/antigen bindings. That's what AlphaProteo is now trying to solve. Also, as the video points out at the end, AlphaFold is well aware that it's unable to do this task and is communicating that fact via its ultra low confidences in the accuracy of the positioning of the proteins relative to each other. So I am not really sure what this video is trying to prove.

new?

it's also unclear how many of these sequences folded into the predicted 3d structure

The whitepaper depicts some successful cases, determined by X-ray crystallography or cryo-EM.

Did you read it?

I wonder if the backlash they received from inventing transformers and then allowing OpenAI to eat their lunch has changed their attitude towards how they'll commercialize future inventions.

It varies, but as the article says it can be used for things like drug discovery. Imagine there's a new virus running rampant. It works by using a very specific protein to latch onto a cells so it can pull itself in. You would like to develop a drug to stop it doing that and one way to do that is to find a protein that wants to strongly latch on to an important part of the virus. If it's holding onto the virus the virus probably won't be able to penetrate cells because you're engineered protein will get in the way. This is part of how antibodies work to stop viral infections naturally