The cells of the human body are full of molecular machines: pumps and copy machines, monitors, regulators and signal transmitters, and small power plants are also included. Each cell type has its own makeup, which is determined by the genetic program that is active in the respective tissue. Anyone who knows what the machinery of life looks like and how it works can also repair things if something goes wrong, as is the case with some diseases. In the journal Nature A new version of the AI Alphafold has now been presented, which should be able to simulate the interaction of several biomolecules. Alexander Schug, an expert in biomolecular simulation at the Jülich Research Center, knows the software’s predecessors. A conversation about the capabilities of AI, how far the results can be trusted, and why it is a problem if the manufacturer keeps the program code secret.
SZ: Why do you fold proteins on a computer screen?
At the molecular level, life is dominated by proteins. The best-known example is probably hemoglobin, the blood pigment that transports oxygen throughout the body in molecular pockets. Hair is also made of proteins and enzymes are proteins. Proteins perform a variety of tasks, but are always constructed similarly from a few basic building blocks. Their three-dimensional structure is directly linked to their function. We therefore want to understand how the basic building blocks fold into functional units. Computers have become an indispensable aid in recent years.
How did people research the shape and function of proteins in the past? These biomachines are too small for a normal microscope.
It used to be said that if a doctoral student managed to solve the structure of a protein in his doctoral thesis, then that would be a success. Such work takes several years. To do this, you first have to produce the protein, purify it and then prepare it in such a way that the biomolecular structure can be elucidated, for example using X-rays. This preparation is incredibly complex for some proteins and often fails. Today, many of these steps can be automated, but there is no guarantee of success.
And now the Google company Deepmind comes along and says that with their AI Alphafold 3 it can all be done much easier, in minutes with a few clicks of the mouse.
Alphafold is not the first program to tackle this problem, but it is the first to do it with this quality using artificial intelligence. There have been many approaches to protein structure prediction before, but the quality of the predictions has been quite a bit of a gamble. There are now different software versions. Alphafold 2 was able to predict the structure of a protein from the given amino acid sequence. Version 3 is now able to determine the structural interaction of multiple molecules. This is really a big step.
Because you can, for example, check how drug molecules bind to proteins?
Also, but not only. You can confront different variants of the molecule with binding partners and see how the situations differ. In Alphafold we can now add additional molecules such as DNA or RNA, which was previously not easily possible. For example, we work a lot with RNA and investigate how these molecules fold into three-dimensional structures and acquire different functions. I’m curious to see how well Alphafold 3 performs in these tasks.
Then there is no longer any need for laboratory work; you can investigate all molecular questions on the computer?
But to be sure, we have to test the models in the laboratory with real molecules. AI only makes predictions that reliably reflect reality in very different ways. Luckily, the AI indicates which areas of the molecules it is very confident about and which regions are more questionable.
How does the AI know this?
The proteins emerged over long evolutionary processes and are similar between organisms. However, the same protein in different organisms always has very similar sequences and therefore a similar structure. The AI compares the protein it is currently working with with all the sequences in the database and can recognize similar patterns. Using AI, we can quickly predict hypotheses of structure and only go back to the lab with the best candidates.
Because only the laboratory results are one hundred percent reliable?
Unfortunately, laboratory tests often leave many questions unanswered. Some areas of large proteins are simply difficult to image. You should also not forget that the proteins are not rigid. There is a lot of movement in the cells; the machines change shape as they work. All of this can now be put together much better into a more complete picture with the help of AI. In this respect, laboratory and computer-aided analyzes complement each other.
Accurate to an atom?
There are different methods to determine the accuracy of the models and you will always get very different values depending on which aspect you look at. But the goal is to depict nature as accurately as possible.
If you can test drugs in silico with Alphafold, then you could also develop bioweapons, right?
Unfortunately, it is relatively easy to develop something dangerous, but fortunately it is often very difficult to produce these substances.
Can filters be built into the software to prevent such misuse?
I am skeptical as to whether this can be easily regulated. Filters can often be avoided by simply changing the question slightly.
Deepmind does not publish the code behind Alphafold 3. Is this a problem for you as a user?
Yes, that was already a fear among scientists in the first version. In order to understand how the AI works, we not only need the source code, but also the AI’s training data and the instructions on how the AI was trained with it. I want to understand which rules the AI follows, after all, I also want to learn what the principles of protein folding are, maybe even the basic physics, what kind of forces play a role and so on. I don’t think it’s a good solution that you have to run your analyzes on third-party servers. I hope that changes.
Will you still try Alphafold 3?
I’m very excited and looking forward to it. It is important that public research also has the resources and expertise to realize such developments and is not dependent on the goodwill of a corporation.
Are you now expecting a boost in drug development?
The AI certainly helps to make suggestions for new active ingredients and perhaps undesirable effects can already be suspected, then these substances can be dropped straight away. But in the end, all active ingredients have to go through clinical testing, which is complex and expensive. Of course I believe it will lead to new medications, but it still can’t be done with the snap of a finger.