AI for Pharmaceutical Chemical Synthesis: Does today's AI have capabilities comparable to those of chemists with 10 years of experience?

AI for Science is becoming one of the most popular tracks in the primary market.

The meeting of the Political Bureau of the CPC Central Committee held at the end of July emphasized "strengthening long-term and stable support for basic research". Combined with the previous Artificial Intelligence + Initiative, a very specific issue has been brought to the forefront: how exactly AI is applied in scientific research, and how to truly improve efficiency and prediction capabilities.

This week, the industry also received blockbuster news. Moderna and Merck & Co. announced that their first co-developed personalized mRNA cancer vaccine has achieved preliminary positive results in the Phase III clinical trial for melanoma. This vaccine benefited from AI in the neoantigen selection process, allowing researchers to simulate complex molecular interactions through computation instead of conducting endless wet lab experiments.

In this issue of Industry Insights, we invited Dr. Ning Xia, founder of ChemMind Technologies. Dr. Xia holds a doctorate in organic chemistry. Counting from his participation in entrepreneurship in France in 2008, he has been on the "AI + Chemistry" track for nearly 18 years. Over these 18 years, he has witnessed the industry grow from scratch, go through the first wave of highlights of AI pharmaceutical, the three-year cold winter of biomedicine, the second boom of AI pharmaceutical after the GPT Moment, and the new trend of AI4S.

They reviewed how the cross-border capability of "standing on two boats" was developed, discussed the respective roles of black-box and white-box technologies in technical implementation, talked about rejected acquisition offers, and calculated a very practical account: if AI is to replace chemical professionals with a monthly salary of 20,000 RMB, how should the token cost be calculated.

The following is the full text of the conversation, which has been edited and abridged.

The people "standing on two boats"

Feng Li: We have been focusing on interdisciplinary innovation around 2015-2016, especially the integration of computer science with biopharmaceuticals and chemistry, and we have also invested in XtalPi and Moleculeverse. After more than ten years of development in this industry, we have found two problems. First, it is relatively difficult to find people who happen to be "standing on two boats" — people with capabilities spanning two non-overlapping disciplines, such as computer science and chemistry. Second, if someone is "standing on two boats", which boat is more important in the end? Let's look at the first question first: when you were young, how did you connect chemistry and computer science, two seemingly unrelated subjects?

Ning Xia: It's mainly out of interest. Programming is a discipline with strong logical thinking, while chemistry requires a lot of memorization and curiosity about nature. I happen to be interested in both. Later, by coincidence, I won awards in both competitions, so I was very hesitant when choosing my major. The admissions teacher from Tongji University told me a sentence: "Computer science is a technology, while chemistry is a specialized discipline." I chose chemistry, which proved to be correct later. If I had studied computer science first and then tried to cross over to chemistry, the difficulty would have been extremely high.

Feng Li: When you were studying chemistry, how did you maintain your previously accumulated computer skills without fading away?

Ning Xia: I wrote code every day. I did experiments during the day and wrote code at night.

Feng Li: What were the programs you wrote at that time related to?

Ning Xia: They were also related to artificial intelligence. At that time, I was very curious whether the way the human brain thinks could be simulated by programs. I did a lot of explorations, but certainly not very successful — if I had succeeded, I might be doing something else now. Later, when I was pursuing my doctorate abroad, it was the same: I did experiments during the day and wrote code at night, trying to make computer programs simulate human judgment in scenarios such as games and chess playing.

18 years ago, there was no data or investment in this field

Feng Li: Did you struggle between choosing to stay abroad or return to China after graduating with your doctorate?

Ning Xia: At that time, the development abroad was quite good, and Europe was not what it is now, so I stayed in France to participate in entrepreneurship. I wanted to find a startup that could use computer technology to solve chemical problems, but there were very few such positions. I was fortunate to find a company in Strasbourg, France. In fact, the company had not been established yet, and its website had just been created. I talked to the boss and felt that our ideas fit very well. They also wanted to use data to solve the prediction problems of chemistry.

Feng Li: How long did this period last?

Ning Xia: 7 years, from 2008 to 2015. At that time, this field was not hot at all, no investors were paying attention to it at all, and we relied entirely on personal ideals and interests. Moreover, there were no underlying AI technologies as there are now, no data, and all infrastructure had to be built from scratch.

Feng Li: At that time, your foothold was also to explore chemical synthesis paths by combining computers and data, right?

Ning Xia: Not yet. The first thing was to find data first, because we didn't even have data. We tried many methods: we used high-throughput-like experiments to generate data. At that time, there was no automation, and people carried out reaction experiments. Lab technicians could manually perform hundreds or thousands of parallel experiments, but the amount of data was far from enough. We also asked people to extract structured data from doctoral dissertations, but the cost was too high and the efficiency was too low. Later, we also developed electronic lab notebooks, hoping that university teachers would share data with us after using them, but this method later encountered some commercial difficulties.

Feng Li: In the situation where you were blocked in multiple ways, what did this company do in the end?

Ning Xia: It was acquired by its parent company, which is now the largest CRO (a professional institution that provides pharmaceutical R&D outsourcing services for pharmaceutical companies through contracts) in France. At that time, we accumulated a lot of underlying technologies, and many of our current underlying architectures were also accumulated at that time.

Feng Li: You returned to China in 2015. I remember you first had an experience of helping a friend start a business.

Ning Xia: Yes, the boss of Shanghai Wanghua Technology called me over. At that time, the chemical B2B e-commerce platform was very popular. At first, I was responsible for building the platform. Later, we thought about whether we could launch the direction I wanted to do before but didn't — AI retrosynthesis. In 2016, we were the first in the world to apply for the domain name chemical.ai, and then ChemMind Technologies was established later.

Participated in the solution competition at WuXi AppTec as an individual and won the first place

Feng Li: There is a small story here. When you were still at Wanghua, you developed a set of software that predicted the synthesis of compound B from compound A — including what conditions to use, how many intermediate steps to go through, what results to get in the end, and even the yield in a partial sense. The main business of WuXi AppTec is CRO for chemical synthesis. They solicited proposals globally for who can do synthesis path prediction. Many companies participated in the competition. Since Dr. Xia had not yet founded ChemMind Technologies, he participated as an individual and won the first place.

Ning Xia: It was really a coincidence. That was in 2016. We had contacts with some executives of WuXi AppTec, so I said that I had such a set of software, would you like to try it? They said yes, and we made an appointment. When I got there, I found that many people came, the room was full of their executives. I demonstrated on the spot, ran various molecules, and the effect was very good, so we had the first commercial cooperation with WuXi AppTec.

Through this incident, I realized that first, the industry attaches great importance to this matter; second, the technology had reached the first critical point at that time — previously, no software could do this well, and we were among the first to let the industry see that this matter was feasible. Up to now, it has become a very common tool, and it has taken about six or seven years in the middle.

Feng Li: There was no GPT at that time, I guess the Transformer architecture was not used either, at most it was the convolutional neural network that came out earlier. Looking back, how "AI-powered" was the product at that stage?

Ning Xia: The AI at that time was a different concept from what it is now. At first we talked about machine learning, then deep neural networks and Transformer emerged, and later it evolved into large models. It has been evolving all the time. But generally speaking, anything that makes predictions based on data is essentially a mode of AI. At that time, it was more focused on traditional machine learning, plus a lot of chemoinformatics — a discipline that specifically studies how to write the logic of chemical reactions into algorithms.

White-box vs. Black-box: White-box is reassuring, the biggest problem of black-box is not error, but non-interpretability

Feng Li: When using AI to make predictions, there is a frequently discussed question: should it be black-box or white-box? The white-box is a summary of a lot of known human experience, rules and formulas inside; the black-box means data goes in, goes through a series of calculations in the middle, and then the result comes out. Today, autonomous driving also encounters this problem. At that time, in meeting the needs of WuXi AppTec, how much black-box and how much white-box was there in your software that outperformed large companies?

Ning Xia: The proportion of white-box was very large. For such serious scientific software, if it is a black-box, it is difficult to build customer trust. Because it will occasionally have hallucinations or bugs, the consequences may be very serious, and it is non-interpretable. The biggest challenge is that you find a problem but don't know where the problem is, so you have no way to improve it.

Feng Li: Yes, it is difficult to adjust — after you change A, if it inexplicably changes B, C, and D at the same time, and you have no way to suppress B, C, and D from changing in a bad direction, let alone you don't even know who B, C, and D are.

We have seen many AI for Science companies in various fields including materials and chemistry. Biology is actually better, because genomics provides underlying constraint rules. What about chemistry? To what extent do you think its interpretability has reached today?

Ning Xia: In our system, interpretability is very strong. All conclusions given by large models or black-box models must be explained, verified and constrained in our white-box model, so that we can eliminate hallucinations, and at the same time know which of those ambiguous conclusions it gives is right and which is wrong. This is very important in the scientific field.

Feng Li: In terms of procedures, do you go through the black-box first and then the white-box, or are the two integrated together?

Ning Xia: There is a strong integration, and the two proceed at the same time.

The first cooperation of more than 1 million US dollars, and a demand list with thousands of feedbacks

Feng Li: When we met Dr. Xia in 2018, Dr. Xia had just started ChemMind Technologies. The first milestone after the company was established was to provide the first considerable software service to a world-famous MNC (multinational large pharmaceutical company).

Ning Xia: I am very grateful to WuXi AppTec for introducing us to one of their big customers. This customer was building an AI pharmaceutical system, and the AI retrosynthesis part was completely handed over to us. This was a cooperation of more than 1 million US dollars, and the difficulty was not small. Through that cooperation, we truly made the productization meet various requirements of these big customers in terms of security, speed and quality. This is a very important starting point for us.

Feng Li: This is a problem that many scientist entrepreneurs will face: they have developed a very usable scientific research result, but when it is applied in a real industrial closed loop, it is sometimes unusable, not easy to use, or if the user is not the right person, there will be resistance. How did you overcome these problems?

Ning Xia: There will definitely be these problems. When your product first comes out, it is definitely not 90 points, maybe 60 or 70 points. We actually experienced a lot of customer feedback, iteration and optimization — which is also why white-box is so important. If you are a black box, customers give feedback but you can't modify it, they will be desperate.

So we kept improving to let customers see that the product is making continuous progress. I also checked our demand list, there may be thousands of opinions put forward by various customers, each of which means a lot of work to improve and correct.

This process is very important, but now many people, no matter investors or the industry, may ignore this point — they think that as long as the model is good and the technology is good, it can immediately become the best in the world. But I think in many cases, it is achieved through such iterations.

Feng Li: Take this single customer as an example, in the first three years, were the departments and people using it relatively fixed, or more and more cross-departmental people were using it?

Ning Xia: More and more. In the process of application, customers will have a lot of ideas, they will say, hey, wouldn't it be better if you do this function in this way? Wouldn't it be better to connect that for me? In this way, we understand how customers use it in real scenarios, which is usually different from our own imagination.

The same molecule has to go through several thresholds inside the pharmaceutical company

Feng Li: When enterprises use it internally, are the users mostly people focusing on traditional chemical synthesis, or will it expand to people with different backgrounds and in more links on the R&D chain?

Ning Xia: It will be used in several links. First, the people who do molecular design, their strengths are not synthesis, they focus on CADD (Computer-Aided Drug Design) or AIDD (AI-Driven Drug Design), but when designing molecules, they must consider whether this molecule can be synthesized and whether it is easy to synthesize, so our tools can come in handy.

Then to the medicinal chemistry department, they will use it to quickly design routes, their demand is to get molecules quickly, for example, a few milligrams are enough, without considering the cost.

When it comes to the clinical process scaling-up link, the demand is reversed: I only make this one molecule, you don't use the easiest reaction for me, you have to find the one with the lowest cost. Sometimes even if the reaction itself has high risks, they are willing to use it because it can reduce costs. Of course, the reaction can't be too dangerous, which may cause explosion when scaled up, that's not acceptable.

Feng Li: I was "tortured" by chemistry major. When I was in graduate school, I spent all day exploring different synthesis routes. Back more than 20 years ago, the norm for chemistry graduate students was that the supervisor pointed out a direction, for example, from A to B, then the students checked the literature by themselves, tried the conditions by themselves, and traversed various failures. Assuming we had this tool at that time, how many problems could it solve for a graduate student?

Ning Xia: Probably all the work of students can now be done with the help of AI. The first thing AI will replace is probably this kind of primary intellectual work.

Feng Li: Then chemistry graduate students will be like many programmers today, being gradually replaced, or maybe they can do more creative things?

Ning Xia: It's a bit similar, but the chemistry field is more optimistic. There are more tasks in the chemistry field, we have too many innovative materials to develop, and the space is large enough. Even if the proportion of work we do is less, we can still carry out more projects.

Feng Li: It is equivalent to increasing the bandwidth of each person.

Highlights, troughs, and unfinished acquisition offers

Feng Li: Back to 2020-2021, due to the epidemic and other reasons, biomedicine reached its highlight moment of being sought after by the capital market. At that time, Dr. Xia also faced a temptation: a large CRO sent you acquisition offers more than once. At that time, your company was very small, with only a dozen or two dozen people. Looking back today, do you think it would be better to sell it then?

Ning Xia: Not really. I am a relatively idealistic entrepreneur. In the highlight cycle, I will also think about whether these funds really help me push this project and this ideal forward. If not, maybe this fund doesn't have that much effect. In the industry trough, when the financing environment is relatively