Alphafold 3 Extends Modeling Capacity To More Biological Targets

Forbes

21 hours ago

Doctor working on digital tablet with medical interface and digital healthcare and network concept.
The people behind the original protein modeling tool Alphafold have now developed a newer version, Alphafold 3, which is changing the way that this fundamental technology works.
Looking at the changes in the newest version, you find that Alphafold 3 extends to a broader spectrum of molecular structures, including ligands (ions or molecules with certain binding properties) and other ions, as well as something called 'post-translational modifications' – (here's the Wikipedia entry.)
Additionally, Alphafold 3 uses a reformed 'Pairformer' architecture to process pairwise relationships (more on that later) - it has better prediction accuracy, and improved performance in making some types of predictions. (Here's more from NIH).
The original Alphafold technology earned its makers, John Jumper and Demis Hassabis, a Nobel prize, and now these tools are still redefining what it means to do drug discovery.
So how does Alphafold work for big drug companies?
In a TED talk explaining some of this commercial success, Lauren Davis, someone with MIT ties, shows us a bit of how this works, helping companies to come up with life-saving medicines.
Davis points to a 'transformative' process where new tools enable rapid development, on a more efficient framework. One aspect of this, she points out, is target identification – predicting the structure of a given target. That way, companies can sidestep some of the human and animal testing that's expensive and labor-intensive, not to mention sensitive.
She compares the system to a dating app, where you narrow down potential matches before proceeding with actual testing, which would be analogous to meeting someone for a date (read: investing time and effort.)
She talks about the process of scoring potential inhibitors, which she says she's excited about because she used to be on the MIT soccer team.
In general, Davis paints a picture of how Alphafold actually applies in the commercial world.
That's a little bit about the way is that Alphafold is contributing to the medical community.
But there's another feature of this new model that I was interested in, and I ended up getting different results from different LLM models, so let's look at that a bit:
If you ask ChatGPT whether Alphafold 3 is open source, you get this - at least, I did:
'(Alphqafold 3 is) not open-sourced; instead, it is accessible through a cloud-based platform provided by DeepMind for non-commercial research purposes. This approach has elicited some concerns within the scientific community regarding transparency and accessibility.'
As a source, the model lists this Wired article.
However, when I looked at a post from Dario Amodei from November of last year, I found this, suggesting, albeit in a terse way, that Alphafold 3 has become open source:
'AI protein prediction tool, AlphaFold3, is open source.'
Now, when I asked Copilot the same thing, as a result of typing a search into Bing, it gave me this:
'Not fully open source - AlphaFold 3 is not fully open source. While the source code and model weights are available for academic use under specific non-commercial restrictions, access to the model weights is limited to those with academic affiliations. This means that while researchers can use the software for non-commercial applications, they cannot freely access the training weights for commercial use.'
And there were a number of sources listed.
So the most likely answer, based on all of that input, is that Alphafold 3 is 'sort of' open source – that is, as Copilot said, that some of the weights and other aspects are public, but other aspects of the technology are not.
ChatGPT contends, above, that the cloud-based platform that Alphafold 3 is on is 'maintained by DeepMind for non-commercial research purposes.'
Anyway, in this case, we don't really have to guess: just check the GitHub for the new version.
But this shows how you can get different information from different models, something we are going to have to learn to navigate.
In any case, this new version of Alphafold continues the tradition of giving us new tools for drug discovery. Davis, in her talk, used the example of lisinopril as an ACE inhibitor, and explained how the Alphafold process can apply to predicting and modeling how an inhibitor will work.
Practically, we have millions and millions of Americans on these drugs, and they apply to a wide scope of health conditions, so it's abundantly useful to take advantage of AI in these ways.
Will it lower the cost of drugs? We'll see.