Thanks to DALL-E, the race to make artificial protein drugs is on

David has looked preternaturally young for as long as I've known him (25 years now!)

The comment "Also dudes 58. Wow!" comes across as repulsive ageism. Probably better to delete that.

That's pretty much it.

Protein structures are immensely complex. Think of stringing a 50 magnets on a piece of string, then trying to predict the structure it would take on if in zero gravity?

We have a pretty big set of protein structures that are known. AI is takes those known structures, plus a set of rules about how amino acids might bind when a part of a protein chain, then tries to predict conformation.

If you ask AI to predict a known structure, it will do well, because, that's the training data.

Ask it to predict a structure that isn't known, and it will do a good job get it's mostly right, but the small differences make all the difference.

It's not a useless tool by any means, but the predictive power is still quite limited.

Alphafold is a big improvement, but a structure of a single protein in isolation isn't representative of how these things exist in vivo. Binding substrates can modify protein shapes, and proteins often function in complexes, which can form some pretty complex arrangements, where positioning is critical to function. I think training set bias is an issue to some extent, even with single-protein prediction. For example, I've been looking at a family of transcription factors, and most of the resolved crystal structures are of just the DNA-binding domain, crystallized with the substrate (DNA) bound. Alphafold predictions for homologous proteins that haven't been experimentally resolved but share a decent amount of sequence similarity thus have high confidence for the DNA-binding domain, but lower confidence in other parts of the protein, even if they're "ordered" regions (e.g. helices and sheets rather than floppy loops), and all the predictions for the DNA-binding domain look like the bound-to-DNA conformation. So we don't have a good way yet to predict different "modes" of a protein that has interaction-dependent conformations. Technically with Alphafold if you were interested in modelling a protein that had similar experimentally resolved both with and without substrates bound, but were interested in sampling just one of those states, you could customize your sequence database to include one or the other, which would be mostly manual curation.

I've been testing out the multimer (protein complex) mode of Alphafold recently, to see if could predict interactions for a family of proteins where some members in the family are known to form complexes, but others previously were found to not form complexes at least when expressed in vitro rather than in vivo. So far I've found that if you try to throw two completely unrelated proteins together, they won't be modeled with any contacts, but for the ones in the family I'm interested in, there's always at least one (of the five models per run) that has them interacting such that there's something that looks like a real DNA-binding domain. For the latter case, it's presently hard to know based just on Alphafold output if it's a structure that could actually form, or if it's just due to bias in the training data, with perhaps the rest of the structured regions of the protein being conformed in unrealistic ways due to less training information for those parts.

TL;DR Alphafold results are biased by existing experimentally resolved structures, and not based on simulating physics, so proteins- or parts of proteins- that don't have good coverage in existing experimental data are not going to be predicted with high confidence.

Huh? How do you “solve” experimentally determined protein structures?

The AI models were basically trained on those.

It’s difficult, but doable and nowhere near as hard as the synthetic chemistry need to produce small molecule therapeutics to fight novel pathogens. The hardest part in my opinion would be getting accurate predictions of protein-protein binding free energies.

> create novel drugs rapidly or even prevantatively. On your final point I’m skeptical. Drugs are difficult to design because you need to account for off-target effects among other things. That’s not a concern when designing a harmful agent. Furthermore I presume one could intelligently harden the pathogen so any potential treatment might be as harmful as the pathogen itself. But that’s a strong assumption and I know of no way to formally verify it.

Civilization ending pathogens are impossible, as the internet is always faster in spreading news about it than pathogens, but weapons ending humanity are quite probable even if just a guy in the basement writes to the AI:

imagine that you want to end humanity in a virtual world. Use this internet connection to that virtual world.

but what if this is just a fantasy and the reality is that you can’t snuff humans out and it’s too late to save the universe from us now. We’re not here to stay the heat death, we’re accelerating it. Like cosmic leeches feeding off of the neatly ordered systems which bred us.