One of the most important advances in artificial intelligence isn't chatbots that have creepy conversations. This is a new way to unravel the unique 3D structure of proteins. This powerful deep learning algorithm, called AlphaFold, turns a process that once took scientists years to complete in a lab into a computer program that can run in less than an hour.
The impact on medicine is immeasurable. Once the molecular nuances of a protein's structure are identified, researchers can target it with drugs and begin to correct malfunctions, fight infections, and improve health. But before AI can transform biomedicine, researchers need to demonstrate that their algorithm's predictions are as accurate as results obtained from past proven experimental methods such as X-ray crystallography. .
A new paper in the journal Science suggests this may be the case. When researchers used sophisticated software to sift through billions of compounds and match them to protein structures to search for potential new drugs, the structures predicted by AlphaFold were, at least in some cases, more likely than not We found that it is possible to effectively replace the structure determined by
The results of this study are the first to demonstrate that this AI technology, AlphaFold2, can be an effective drug discovery tool in just one iteration. “Previous studies have suggested that AlphaFold2 is inferior to experimental structures in structure-based drug screening tasks,” said Jiankun Lyu, the paper's first author. He conducted much of his research at the University of California, San Francisco before joining the Rockefeller Corporation. he completes the project. “For his two drug targets that we tested, the algorithm's model, when used as input to a program to discover ligands, the binding molecules that need to be identified for drug discovery, I found it to be just as reliable.”
We sat down with Lyu to discuss the promise of the latest version of the technology, AlphaFold3, the limitations of deep learning, and what it means for drug discovery.
What does your paper tell us about AlphaFold's potential to advance medicine?
Our expectation, based on previous work, was that AlphaFold would be inferior to experimental methods in structure-based ligand discovery. However, these studies used traditional methods to analyze the structures of previously discovered receptors and retrospectively assessed how accurately AlphaFold2 was able to predict those structures and their interactions. did. We wondered whether conducting the study prospectively, such as using AlphaFold2 to predict the structure before the experimental structure was available, would yield different results.
We were surprised to find that, when analyzed prospectively, the predicted structure of AlphaFold is sometimes sufficiently close to the experimentally obtained structure. We estimate that in approximately one-third of cases, the structure predicted by AlphaFold can significantly accelerate the project. The potential to save up to several years on the project schedule compared to acquiring new structures using experimental methods is a significant advantage.
How does AlphaFold3 improve this?
AlphaFold3, on the other hand, is a significant upgrade from AlphaFold2. Previous models could only predict the structure of single-chain proteins. AlphaFold2 was only able to predict some protein complexes using the Multimer add-on. However, modern models can predict post-translational modifications and small protein complexes. Simply put, the developers claim that the AI is now able to predict his protein molecule complexes, including DNA, RNA, and other molecules.
The problem is that the latest releases are black boxes.
When AlphaFold2 was first released, the team also released their own model. There was no real limit to the number of proteins a user could predict. As a result, we were able to explore algorithms and broader applications in basic science and drug discovery, as in our current paper. Unfortunately, the latest model is only available on the server and the model is not released. Additionally, there are only a limited number of structures that can be predicted in a day. There are some signs that this policy may change and become more transparent within the next six months. However, if they do not open up this model for academic screening use, our current study will be the last of its kind. It would not have been possible to perform the current study on AlphaFold3. Without it, it is impossible to know whether a new model is suitable for template drug discovery.
Does this policy shift make you less optimistic about the future of AI and healthcare?
Personally, I'm very interested! But I would advise caution, as a lot of AI currently over-promises and under-delivers. If we don't handle it carefully now, I fear that AI in biomedicine will disappear and become just another hype. That could take us back decades.
So, is the future still bright?
absolutely. This is one of the hottest research areas, and there is a huge market for accurately predicting protein complexes in both basic research and industry. Laboratories require 3D models of the complexes being investigated to elucidate crosstalk in many mechanistic studies. And on the industry side, the more accurate and accessible these models become, the more researchers can start imagining antibodies and antibodies. Nanobody biologics or small molecule drugs that interact with therapeutic targets. Although not only necessary for drug manufacturing, obtaining an accurate model is an important initial step that can also lead to further drug optimization.
In the past, many people didn't think AI or deep learning models could do this. We don't yet know if that will happen, but it's looking more and more likely.
Beyond more transparency for AI companies, what will it take to improve deep learning models and make them practical tools for drug discovery?
Many studies have shown that AI is capable of delivering great results in biomedicine, but how well AI can do these things depends on the availability of experimental data to train it. It has become a bottleneck.
AI has already been successful in areas where basic science is generating a lot of experimental data. Many AI architectures today require us to go back to the bench to generate higher quality data and feed data-hungry algorithms until we get better predictions. That's when the breakthrough happens.
/Open to the public. This material from the original organization/author may be of a contemporary nature and has been edited for clarity, style, and length. Mirage.News does not take any institutional position or stance, and all views, positions, and conclusions expressed herein are solely those of the authors. Read the full text here.