In my humble opinion, this work is not that innovative: de novo protein binders have been done to death, either by AI approaches or otherwise. Check out the work by David Baker’s group, for instance. They have a myriad of examples already.

That being said, as others have commented, my hopes are that all these advancements lead finally to reliable design methods for novel biocatalysts, an area that has been stalling for decades, compared to protein folds and binders.

Maybe this is in the supplement of the whitepaper [0], but I would have loved to see more analysis of how novel the designed proteins really are.

In the whitepaper they mention that they are novel compared to other in silico design techniques, but to my knowledge other binders to VEGF and Covid spike protein exist and would already be found in the PDB database that Deepmind trained the model on.

This is not to minimize the results- if the history of ML is anything to go by, even if AlphaProteo does not currently beat the best affinity found by in vitro screens, I do not doubt that it soon will!

[0] - https://storage.googleapis.com/deepmind-media/DeepMind.com/B...

Two minute papers video on the subject: https://www.youtube.com/watch?v=lI3EoCjWC2E

Interesting work, but there's a huge sector they're missing - industrial enzyme and catalysis design. Most of this field is concerned with small molecule binding - methane, carbon dioxide, ammonia, methanol, acetic acid, etc. Binding is often just the first step, as you're typically trying to do highly specific chemistry, e.g. attaching a single oxygen to methane or a single hydrogen to carbon dioxide, etc.

Working in this area might also be good test of their technological approach, as small-molecule binding can be somewhat challenging, and even evolved biological systems can struggle to achieve high specificity.

What is Google actually doing with these systems? Are they using it to develop new drugs themselves? Or licensing it to the pharmaceutical industry?

I have a question that hopefully a molecular biologist can answer. Can tools like this potentially create protein structures that specifically bind in certain cells? Or is this more about a way of being able to create proteins for genes / structures we haven't been able to before?

I'm very interested in my research at the moment in pleiotropy, namely mapping pleiotropic effects in as many *omics/QTL measurements and complex traits as possible. This is really helpful for determining which genes / proteins to focus on for drug development.

The problem with drugs is in fact pleiotropy! A single protein can do quite a lot of things in your body, either through a causal downstream mechanism (vertical pleiotropy), or seemingly independent processes (horizontal). This limits a lot of possible drug target as the side-effect / detrimental effect may be too large.

So, if these tools can create ultra specific protein structures that somehow only bind in the areas of interest, then that would be a truly massive breakthrough.

Question for bio folks here, and not to steal from the joy of this article but I've been recently curious how far are we from engineering something like a virus that targets a subset of the population (e.g. via specific genetic markers). This sort of tech being commoditized feels much much scary than the LLM safety talk - by a mile.

I saw a recent video about the errors of alphafold 3.

https://www.youtube.com/watch?v=E61wJXlENoE

This is interesting work but I think something has been intentionally overlooked. Creating proteins is difficult and it's also unclear how many of these sequences folded into the predicted 3d structure. Small molecule synthesis is still easier, cheaper, and more scalable than protein therapeutics. I think this would've been more impactful had they focused on improving on the SOTA small molecule - protein interaction models.

> Trained on vast amounts of protein data from the Protein Data Bank (PDB) and more than 100 million predicted structures from AlphaFold, AlphaProteo has learned the myriad ways molecules bind to each other. Given the structure of a target molecule and a set of preferred binding locations on that molecule, AlphaProteo generates a candidate protein that binds to the target at those locations.

I wonder how many prions will be accidentally created by this, or if it can even predict if a particular protein will have prion-like effects

One question is how specific the binding is -- what's the level of off-target effects, etc.

I am sorry for my naivety, but what is the practical benefits of this?

I only comment on hacker news posts about biology because I'm a voice crying in the wilderness about the most important goddamn startup on Earth, I think, maybe. www.molecularReality.com

what kind of model architecture was used for this? is it safe to assume they used a transformer model or a variant of it?

This could go wrong, in many directions

any resources to self learn biotech and how to use ML in biotech ?

yeah yeah whatever another protein discovered oh wow... When are we going to see actual results? Hurry up Deepmind!

(not to be confused with AlphaProto, which is helps with Google's core business of turning protocol buffers into differenter protocol buffers.)

It generates novel candidates doesn’t actually generate proteins, and none of these proteins have actually been generated to validate whether these candidates are shit or not

Safety is the new gatekeeping.

This is equivalent of ChatGPT generates novel code, but we didn’t run it. It probably works though.

It's extremely refreshing that DeepMind is still working on using AI to solve hard problems instead of attempt to put creatives out of work.