As artificial intelligence technology continues to improve, many industries have ushered in new breakthroughs. In recent years, AI has accelerated its efforts to develop new drugs, and has participated almost in the entire process from drug target discovery to clinical trials. Even during the COVID-19 pandemic, there is AI behind the emergence of many drugs. For example, in May this year, Ansi Intelligent, an innovative AI drug research and development company headquartered in Hong Kong, China, used its AI drug research and development platform to discover oral inhibitors targeting the main protease (3CL) of the new coronavirus; in addition, the first new coronavirus oral drug approved by the US FDA - the Pfizer Nematvir tablet/ritonavir tablet combination was also discovered with the help of the AI algorithm of the "MareNostrum 4" supercomputer.
We will pay attention to this issue: What role does
□What role does AI play in drug research and development?
□What is the principle of AI pharmaceutical?
□How to ensure the safety and effectiveness of AI pharmaceuticals?
□How to win the lead in the field of AI pharmaceuticals?
□AI Accelerate the discovery of new drugs, and can also accelerate the "new use of old drugs".
. What we have brought to us is Ye Sheng, professor of , , Beijing University of Aeronautics and Astronautics, secretary-general of the Science Popularization Working Committee of the Chinese Society of Biophysics, and structural biologist.
1. How to develop drugs in AI
[Question] In the past, pharmaceutical was a long and cost-effective process. The addition of the artificial intelligence algorithm not only improved the speed of drug research and development, but also improved the quality of drug research and development, making the pharmaceutical process more cost-effective and faster. So what role does AI play in drug research and development?
[Answer] The most essential reason why pharmaceuticals is a process that costs a lot and takes a long time is that we still have many problems with the human body itself. Many drug molecules are developed into drugs according to the ideas we designed, and when they are used on humans, we cannot predict whether they can achieve the functions we expect and whether they will cause some toxic side effects in our bodies.
Therefore, drug development requires continuous exploration and modification, and after a long process, we can finally obtain effective drugs. This process is full of uncertain factors. It is not that this problem is a random problem, but that the mechanism behind it may not be revealed by our current scientific level, so this makes us face great uncertainty.
AI is precisely when this uncertainty is relatively strong and similar to a chaotic fuzzy system, it has played a very powerful problem-solving ability, which makes AI and pharmaceuticals naturally have the possibility of being linked together.
Image source Visual China
[Question] When it comes to AI, many people may think it can be used to play chess, or do some game sports, and can beat people. It is easy to understand how to play chess. For example, if you use chess scores to train AI, you will know which step is the best. But even people can’t figure out which medicine or which direction is right. Why can AI do it?
[Answer] In fact, some mathematical principles behind this may be the same. For example, the AI program that defeated humans in the Go competition was called AlphaGo, and a later upgraded version called AlphaGo Zero. Both programs were developed by DeepMind, a company owned by Google . When they deal with Go issues, they regard each chess situation as a picture.
Standard Go board Picture source Network
We know that the board specification of Go is 19×19, so AI treats the board as a 19×19 pixel dot matrix diagram. When it sees this picture, what it needs to calculate is where to add a chess piece, so that the change in this picture has a greater chance of leading to victory.
pic source Network
In the pharmaceutical field, we will also find ways to transform some pharmaceutical problems into pictures-looking problems similar to those in chess games. cannot be said to be literate by looking at pictures, but to identify patterns in them by looking at pictures.
For example, what we are more common in pharmaceuticals is the problem of toxic side effects I mentioned just now. Maybe we collect many different drug molecules, or molecules that have conducted clinical experiments and animal experiments, and then build all the relevant knowledge about toxic side effects into a relationship diagram, and then give the pictures to AI for learning.
So when you see enough of this kind of picture, AI can obtain some knowledgeable content from it. But this is something we cannot describe with formulas or rules. It only exists in the neural network of AI. Then, we apply such a neural network to our new drug molecules. Maybe this neural network can tell us immediately how likely this new drug molecule has toxic side effects. If it is very large, then where we modify it, we can remove the toxic side effects. This is what AI can help us.
If there is no such rapid neural network as AI to help us make quick judgments, our traditional methods may have to change and try them everywhere, and this may be a very long process of screening and exploration. Now AI may be able to calculate a result in a few minutes to tell us.
Of course, this result is not necessarily 100% accurate, but it can often give us a good prompt or direction, allowing us to know which direction to do the next work and which group to make changes, the possibility of success will be greater.
2. How to ensure safety and effectiveness of AI pharmaceuticals
[Question] But pharmaceuticals are different from playing chess. The essence of playing chess is zero-sum game . Either I win or I lose. Pharmaceuticals are explorations in unknown fields. Humans need it to solve the two problems of efficacy and safety of drugs. How does AI solve it?
[Answer] Your question is very good. In fact, pharmaceuticals are concerned about many issues. In addition to the effectiveness and safety you just mentioned, there are also costs and economic benefits that should be paid attention to as a product. Therefore, the pharmaceutical process itself is a very long pipeline . We usually use the word "pipeline", which means that it actually has a long process.
From the initial basic biological principles, the understanding of this disease, to the end this drug can become a drug product that is truly approved for sale, there are many links in between. Now that AI is involved in the pharmaceutical process, it often intervenes in a single link to solve the problem of this single link.
For example, AI first played a role in the discovery of chemical synthesis paths. Because many of our drug molecules are not natural compounds, but drug molecules obtained through many artificial design and transformation, although such a molecule can be synthesized step by step in the laboratory, it is difficult to amplify to the industrial level and carry out large-scale production in such a process. If large-scale production is carried out, it must also consider its efficiency, cost and other issues. After AI intervenes, we let AI learn various chemical reaction processes. Finally, when we give AI a new compound, AI can tell us what raw materials are from, what chemical reactions are used, what kind of finished products are obtained, and how to isolate the effective substances of the drug from the finished products, and then proceed to the next reaction, helping us quickly find a suitable chemical synthesis path.
Image source Network
So in fact, AI is not a wise man who can solve all the problems in drug development. It is just that there are special AIs in one link that can help us, give us some tips, or give us some possible options, and then accelerate the process of drug development.
[Question] So in your opinion, people are still the most critical factor?
[Answer] That must be the case. When combining AI with biology, biomedicine research or technical research, we already have a very obvious understanding, that is, we find that professional knowledge is very, very important.After all, AI is just an algorithm, a program. How we build parameters of this program and how we build a neural network model requires very, very strong professional knowledge. So at this time, if we already have biology experts and pharmaceutical experts in this field, then to what extent their professional knowledge can be integrated with AI often determines whether this AI can successfully develop drugs.
3. In the new era of drug development, overtaking on the curve, you can expect
[Question] There is a voice now that worrying that AI will replace humans in the future, but at least the question we talked about today will actually find that AI has developed well, and the requirements for people are higher. If this person needs to understand AI, computers, algorithms, and biomedicine, then this person is definitely not good. With the help of AI, the drugs he can develop will be better and more results. This is a plus for individuals. How do you judge, Teacher Ye?
[Answer] What you are saying is very correct. In fact, the development of AI is not simply replacing human work, but provides us with a very powerful tool that allows us to have more creations.
For example, in the field of biomedicine, the combination of biology and AI has produced many new achievements. For example, the AlphaGo mentioned earlier, DeepMind, the company developed AlphaGo, has also tried to do many other things with AI. One of the very influential things about
is that in the past two years, they developed a program called AlphaFord, which can predict the three-dimensional structure of protein . The three-dimensional structure of proteins is actually a problem I have been studying for so many years. Each protein has its own unique three-dimensional structure. We used to measure this structure through some traditional experimental methods, but this determination process requires one protein to measure once and one protein once. There are a lot of work content and the speed is very slow. There are often some proteins that we cannot measure.
Now with the AlphaFord program, it can help us predict. Although the prediction results are not necessarily 100% accurate, it is a very beneficial tool for us. It can give us a prompt to help us accelerate the determination of protein structure.
blue is the protein structure predicted by the computer, and green is the experimental verification results. The similarity between the two is very high. Source: DeepMind
And as you said, this industry requires a lot of talents in interdisciplinary , which requires both biology and pharmaceutical industry and AI. So in order to promote the development of this field, our Chinese Biophysics Society has recently established a new branch called the Artificial Intelligence Biology Branch. This branch was established to promote the better integration and development of artificial intelligence and biology, including biomedical technology.
[Question] In the future, in the field of pharmaceuticals, rich experience, talents and core algorithms are core competitiveness. From the perspective of raw materials, our country's pharmaceuticals actually occupy a very important market share in the world. In the future, AI pharmaceuticals will definitely become a very important direction. If we want to continue to maintain a certain market share and even occupy a leading position in this field, should we also need to do more work in the development of algorithms, the design of AI programs, and the cultivation of related talents?
【Answer】That must be the case. The industry and academia see the possibility of fusion of AI and biology in the past three or four years. There are really some good results, but it is the past one or two years. In fact, in my opinion, this is a good opportunity to overtake on the curve, because the whole world is standing on the same starting line in this field.
So on the one hand, talents in related fields are being cultivated, and students have already studied related issues in the entire graduate process, and on the other hand, related companies are also being nurtured. Our country now has dozens of companies of all sizes in the AI + pharmaceutical track, some of which are ahead and have produced some results.This is the case from a global perspective, but companies in this field have just begun to be established in recent years.
I think through the common development of everyone, including the Artificial Intelligence Biology Branch I mentioned just now, gathering relevant talents from the perspective of the society. From the perspective of scientific research institutions, some teachers are taking students to engage in related research, and then to related AI pharmaceutical companies, and gradually making efforts to study related technologies. I think we may have been lagging behind in small molecule drugs for a long time before, so in the face of such a new opportunity, we have the opportunity to catch up.
Picture source Network
4. AI Accelerating the discovery of new drugs can also accelerate the "new use of old drugs"
[Question] In the past, there was a very interesting situation in the pharmaceutical field called "anti-Moore's Law". ( Moore's Law : When the price remains unchanged, the number of transistors that can be accommodated on an integrated circuit will double every 18 to 24 months. In other words, the performance of the processor is about doubled every two years, while the price drops to half of the previous one. ) The so-called anti-Moore's Law is that the number of new drugs listed by pharmaceutical companies investing $1 billion is reduced by half every 9 years. How do you interpret this anti-Moore law by Teacher Ye? Why does this happen objectively? Is it possible for the use of AI to break the anti-Moore law?
[Answer] The causes of anti-Moore's law actually have many influencing factors. For example, the most basic thing is that we have a disease. If there is a relatively mature drug that can cure it, then other companies invest a lot of money and time costs to study drugs that also treat this disease, their motivation may not be so sufficient. This is actually a very commercial consideration.
So as more and more diseases have such mature drugs to treat, the speed of corresponding new drugs will definitely decrease, I think everyone can understand.
But in fact, we need to know that diseases that can be cured only account for a small part of the disease, and a considerable number of diseases are actually uncurable. When everyone studies those diseases that cannot be cured, they often find that the reason why these diseases are cured is too difficult. It may be that we do not have a clear understanding of the pathological mechanism of this disease, especially the molecular pathological mechanism, and it is difficult for us to come up with a drug to treat it.
So the reality is that there are already medicines that can be cured, and we already have medicines, and it will be quite difficult to compete. There are many difficulties in overcoming diseases that cannot be cured, which makes the birth of new drugs slower and slower. This is an objective reality that we really cannot avoid.
Will the emergence of AI speed up the solution to this problem? I think it will definitely do it. But our problem now is that AI can only accelerate in some links, such as some links in the long pipeline of drug research and development, but we still have no new ways to accelerate the overall drug research and development process.
But maybe we have some other ways, such as if we don’t do small molecule drugs, because now we have started to use antibody drugs to inhibit the development of tumors, then antibodies are macromolecular drugs and protein drugs. Protein drugs themselves are an emerging direction, and many links on their pipelines are different from traditional small molecule drugs. Therefore, AI is very useful in accelerating the research and development of protein drugs. This is also some of the research directions that our own research group is already doing.
[Question] I saw something mentioned in the report is very interesting. Before, we had many mature drugs on the market, which targeted the treatment of certain diseases, but in the AI algorithm, it was found that these drugs can even cure other diseases. For example, in 2020, the British Shanxin Artificial Intelligence Company announced that through its AI platform, it discovered that a rheumatoid drug, baritinib, in the United States, could be used to treat , the new crown .Of course, this medicine is definitely not a very mainstream drug for treating COVID-19, because I didn’t hear its name later. But does this incident also remind us to a certain extent that AI can dig out more possibilities in these treasures of old medicines that have been launched? What do you think about this direction?
[Answer] AI has great advantages in accelerating the use of "new use of old medicines". The reason why many people are paying attention to this direction is that the failure of many of our new drugs is not in the early research and development, but in the last step, that is, in human clinical trials. The link after the drug enters the human body is the most prone to problems, and it is also a link that we cannot replace with other experiments. It may be that after being used in humans, it is discovered that this drug has some toxic side effects that were not found in animal experiments before, or that it metabolizes too quickly after it reaches the human body and cannot be retained by the human body, or it may have no effect in the human body.
So one of the benefits of old medicine is that the problem we need to solve is whether it has any effect on new diseases. has been verified because of its many problems in the human body. For example, it may have no toxic side effects, or it can maintain a relatively stable metabolic level in the human body and can remain in the human body for a long time to work.
"New Use of Old Medicine" Another benefit is that AI can help us quickly find an effective drug. This is also reflected in this new crown epidemic, because in the urgent situation of new infectious diseases, there is actually no long cycle, allowing us to develop a new drug step by step. At this time, we can first check whether the old drug is useful.
So if we can pass a virtual screening like AI and let us know first that certain drugs also have certain effects, it will definitely help us accelerate the process of "new use of old drugs".
flow chart of traditional drug discovery and new use of old drugs Source: Internet
The traditional methods in the past cannot be understood, but with the help of AI, it may become a path, which also makes us look forward to the future path of drug research and development. With the help of relevant technologies, more and more drugs we use or more effective drugs will be used. This is also a new era brought about by technological progress, including the integration between different disciplines.
