Hype versus reality: What you can't do with DeepMind's AlphaFold in drug discovery DeepMind's AlphaFold model has predicted nearly all known protein structures discovered yet, though its ability to help scientists discover new drugs remains unproven. Proteins are complex molecules created by organisms to carry out the biological functions necessary for life. Generally made up of a string of 20 amino acids, …
So AlphaFold doesn't do docking, which it wasn't designed to do, and neither it does any of the many, many difficult tasks involved in designing a new drug, beyond protein folding prediction.
This sounds a bit like, you've been driving screws by hand since forever, and then someone comes and gives you an electric screwdriver, and you say, okay, but it's not building the house by itself.
Although I blame the press for overhyping and oversimplifying stuff. A bit like when the Human Genome Project was completed, and someone immediately started claiming that every disease would be fixed shortly.
"Utilizing these standard molecular docking simulations, we obtained an auROC value of roughly 0.5, which basically says you're doing no better than if you were randomly guessing,"
...so just like docking scores for crystal structure derived models then.
To be fair, the authors actually say this:
"These findings show that molecular docking using AlphaFold2-predicted structures is similar to using experimentally determined structures"
...which they then go on to improve through a rescoring model, and get an auROC of about 0.6. Still not great, but better. Again, this is worth doing with experimental structures too.
Basically the MIT press release, and consequently this article, don't really reflect the conclusions of the paper.
Basically, if everything sticks to everything else, you're doing the experiment wrong.
My guess is that the authors expected some dead-cert hits (e.g. beta-lactams vs peptidases), some less clear indications and a background of nulls. Instead they seem to have a rather unhelpful mess of random numbers.
I'm sure the method will improve, but it needs feeding with more pertinent information from laboratories. The actual experiments were somewhat on the "high throughput" side for my taste.
When just predicitng how a string of amino acids would fold was viewed as combinatorically impossible.
Which it no longer is. This is great.
If you want to make a protein shape on demand (which you might want to)
Turns out that's the wrong question.
What you wanted to ask is "Given this existing, natural protein, what shapes do I need to make (possibly with an amino acid string, possibly by straight organic chemistry) that lock onto it"
There is a phrase, "The inverse problem" that kind of covers this but doesn't quite. The classic example is where you want to design a wing shape (say for a fluid pump) that delivers certain specific performance parameters, rather than try shapes and analyse them and check what parameters they give. Software to solve this problem does exist but the optimal wing may demand tolerances, or razor edges, that cannot be made by any existing tools.
Still good to know the SoA has substantially advanced in this corner of the problem.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
