2026.10.08 | TB | Life Science | AI-Driven Scientific Research Expert

Client-led • 23 Sep 2026

Companies: SDGR US, 2228 HK

Expert Researcher at Beijing Shenshi Technology Co Ltd

Agenda

• Over five years' experience in the AI drug R&D industry, specialising in AI-driven scientific research and the integration of computational chemistry with simulation software
• Familiar with the evolution of AI-driven scientific research, the balance between model accuracy and cost, plus tool and algorithm bottlenecks in the AI drug R&D process
• Able to discuss the differences in business models between offering AI-driven scientific research as a service and selling software, the feasibility of large models directly invoking simulation software, the progress of automated laboratories taking over experiments, plus the barriers to entrepreneurship against the backdrop of large models for scientific research

Questions

00:01:52] What is your understanding and definition of AI for science and the popular RSI [research-oriented scientific intelligence]?

Expert [00:01:52]: AI for science itself is still doing science, dealing with problems related to specific scientific fields, no matter what algorithms are used. For a drug R&D company, its pipeline requires specific molecules. It does not care where the molecule comes from. It can be produced using AI algorithms or through experiments, that is, the traditional medicinal chemistry approach. Either way is acceptable. The key is that it must be active, have corresponding druggability and meet various indicators in the pharmaceutical chemical industry chain.

Therefore, AI or algorithms are just a tool for designing molecules. Whether it is using various pharmaceutical chemistry-related algorithms or deep learning-related algorithms, work related to molecular design and activity evaluation is required. The same is true for other scientific fields. They all need to enter this field to solve the scientific problems in corresponding industries. Computing is already an existing field. For example, there are corresponding theories for material calculation or semiconductor-related physical methods such as solid-state physics. You can write some calculation and simulation methods after calculating the theory. This is the field itself. In terms of materials, there is some R&D related to batteries or electrolytes. A large part of its work also involves molecular simulation methods in computational chemistry.

ry day we resBefore AI came along, it had nothing to do with deep learning; it was purely a first-principles approach. Later, it became possible to introduce this deep learning approach, using highly non-linear neural network functions to describe the system, which might offer better efficiency in certain cases. Of course, this can be elaborated on later, but it generally works this way. So, when doing science, AI is just a relatively new set of computational methods for tackling scientific problems. It might be useful or it might not. It depends on the specific scenario and specific situations require specific analysis.

Like the RSI you just mentioned, it is just an AI term. Using new data to feed back into the model and modifying the corresponding models within to achieve better results is something that has always been done in scientific computing, just perhaps without such a fancy name before. For example, when evaluating the activity in a drug R&D pipeline, the initial algorithms and models might not evaluate the binding free energy very well or accurately. Once real experimental data is obtained, we can look at this data, realise our calculations are inaccurate and then figure out where they are inaccurate, analysing which parts of the algorithm require correction. Perhaps the model's analysis of chemical observations isn't quite right or maybe some force field parameters were not set properly, requiring adjustments. Previously, this was done manually, but now it is moving toward being more automated. However, from our perspective, AI at its current stage cannot automatically iterate that perfectly; it still requires manual correction and manual iteration.

Because what we research every day are scientific problems, a large part of it involves knowledge that simply does not exist anywhere in the world to be learned. Without an existing corpus to allow the model to innately know these things, you have to rely on constant trial and error. You only realise your approach was correct after you have actually achieved it. Therefore, it embodies the concept and spirit of experimentation. In many cases, algorithms must be gradually built up by repeatedly working backwards from actual experiments. For a certain direction, such as a particular target or a specific series of molecules in R&D, having a certain predictive ability for it is based on having seen many similar cases before, which gives us the ability to generalise. This generalisation can then be handed over to current algorithmic models or AI models to execute well. Thus, the prerequisite is that this class of problems must be thoroughly understood first. Before reaching that point, in most cases, you basically have to iterate the methods manually.

00:07:50] Was the previous wave of AI for science companies using neural network methods to replace humans who previously used simulation software?

Expert [00:07:50]: That wave did not eliminate them either. In fact, they still coexist. Fundamentally, computational simulation is still used because this is determined by theory and principles. The question of how well a drug molecule binds to its target protein is scientifically clear in terms of binding activity, as is known from textbooks. It is the free energy of binding between the drug molecule and the protein. Free energy is a concept in statistical physics. There is a formula that can be written down explicitly. As long as the formula is evaluated correctly, in principle you can obtain results exactly the same as those from experiments. The premise is how to calculate it. Therefore, the accuracy of this computational method must be made sufficiently high. Whether the method uses AI models or molecular equation simulation methods, in any case, you need to calculate this formula. What follows is how to keep improving the accuracy of computational models so that this theory can be calculated correctly. In drug development, this is what needs to be done in most cases.

In the past, this simulation was done by some classical physical models and functions. Later, there may be some requirements for accuracy. A higher-dimensional nonlinear method called a neural network can be used to do this. That is, AI naturally transforms things into AI. However, if higher-dimensional nonlinear functions are used for calculation, there will be corresponding performance problems. Maybe in industry and pipelines, they don’t want to pay such high costs, so they make some trade-offs. In addition, it is found that in the classic model, as long as corresponding adjustments are made, the accuracy is sufficient. Therefore, in many cases, for pipeline applications, it is not the pursuit of absolute correctness in theory, but a balance between correctness and performance.

00:11:00] You think it is important to improve accuracy at least in the pharmaceutical scenario. However, neural networks have somewhat excessive performance and relatively high costs for drug discovery, requiring a trade-off, whereas using traditional functions or algorithms with some adjustments can also achieve the required precision, right?

Expert [00:11:00]: That’s right. Throughout the entire AIDD or CADD pipeline, many different modules need to correspond to specific scenarios in drug R&D. Generally, drug R&D also has several milestones. The first is to initiate a project and determine which target protein is genuine. First, confirm that the target genuinely has corresponding biological sites and cellular pathways and can inhibit a particular disease. This is the first step in project initiation. This step mainly involves the investigation of protein targets. A three-dimensional structure of the protein may be needed. If you have the corresponding structure from experiments such as X-ray or NMR, you can obtain it. If not, it is possible to use an AlphaFold-like approach to perform structure prediction using deep learning. This is what DeepMind did well in the past. This is the first step.

The first milestone, target validation, is mainly about determining project initiation, which means deciding to work on this protein target. Next, corresponding drug molecules need to be designed based on the target protein's pocket and its active space. This design process involves several steps. First, after designing a large number of molecules, they must undergo screening. There are several key metrics for screening drug molecules. The first and most standard metric or the first step, is activity. The drug molecule, whether a small or large molecule, must be able to inhibit the activity of the protein. Only in this way can it be effective and achieve the desired therapeutic effect against the disease.

In the second step, when molecules enter the human body, they must not be easily metabolised, nor can they be toxic to the body. For example, they cannot attack the heart or liver. Specific metrics include liver microsomal stability, cardiotoxicity and the inability to cross the blood-brain barrier. These represent another set of indicators. In short, screening is a process similar to looking for a needle in a haystack, performing multiobjective optimisation screening across various metrics from a large pool of generated or designed molecules. Therefore, the first step, molecular design and molecular generation, is an area very suitable for deep learning methods, which, simply put, is about generating molecules. It is just that the generated molecules must match the shape of the protein pocket structure so that they can bind structurally. The binding relationship between a protein and a drug molecule is somewhat like using a key to open a lock, forming a lock-and-key relationship. There is a hole in the keyhole and the goal is to design a key to insert into that hole.

The structure of the drug molecule must be able to bind together, which is called a binder. This molecular structure must meet this requirement and there may be some standard requirements. For example, the molecule must be able to be physically synthesised because you need to be able to produce this drug molecule, which is stored in your computer. You can’t have no entity, so you must be able to produce it and have a better synthesis route. Or if it is an antibody, how the antibody is synthesised and how it is expressed in mice are questions that need to be considered at the survival stage.

Next, there are also indicators such as drug-likeness. In short, these are rules that are already relatively well known. Of course, calling them rules is rather general. They can be refined into dozens or perhaps a series of learnable materials and such data can enable deep learning models to learn. Therefore, with these rule-based materials in place, the model is then given a large number of publicly available drug-like molecules or FDA-approved drug molecules. You know what real and correct molecules look like, so you let AI learn from them. It becomes a model that has learned the language of chemistry, not the Chinese or English that we humans speak today, but the language of chemical molecular structure and you can give the model the ability to generate molecules.

Whether AI is appropriate for many problems depends on the kind of problem. For molecular design problems of this kind, the goal is to obtain a series of better molecules and this is achievable. However, regarding the activity evaluation step we just mentioned, if I now have a molecule, how do I know whether its activity is good? For this, you need a scientific indicator, such as activity, to evaluate it. Activity, as we just mentioned, is physically measurable and has a theoretical basis, so there is no need for AI to reinvent a theory. Because this theory is already known, the goal is simply to calculate it accurately.

There is a method called FEP, which is a free energy perturbation molecular simulation method. It truly captures the physical laws of the binding between proteins and molecules. Since these physical laws are known, this method can be used and quantitatively matched with experimental activity assays. In this way, the method forms a closed loop, which can explain experimental results and help medicinal chemists save a portion of experimental costs. Ideally, you would not need to conduct experiments in the future and could just rely on calculations. In actual practice, however, you might design dozens of molecules and evaluate them using the FEP method. Instead of synthesising all 50 molecules, you might now only need to synthesise the top three to five, because computational methods help you screen out a batch first. This is a relatively standard operation at present.

00:18:35] Drug R&D involves many links such as project initiation, target selection and subsequent multi-indicator screening. Do AI-driven drug R&D use different neural networks at different links or is there a complete large AI model for it or are they all classified?

Expert [00:18:35]: Different links correspond to different principles, indicators and methods. Therefore, they are different vertical models without a general model. This kind of model may involve more than ten years of work in every stage, so it is not at all simple. Every stage has many details and they must be done accurately before the results can really match experiments and only then can computational or AI methods be used to accelerate development and save experimental costs. In drug development, basically every stage must be done accurately. Therefore, vertical models across various aspects are definitely required. In this field, the best performer in AIDD or CADD is Schrödinger, a US company, while the two or three domestic companies are basically still in a catch-up stage. Schrödinger started out in computational chemistry and later also did some AIDD-related work.

Those who are actually involved in algorithm development in this field will realise that, in order to truly make these methods AI-based and use AI methods to replace them, the first prerequisite is that the most fundamental computational chemistry methods must be good enough and sufficiently detailed. Only then can there be sufficient algorithm data, that is, algorithm rules or data generated by your current algorithms, for AI to conduct deep learning and simulation and learn better. That is basically how it is. Therefore, the success of AIDD depends on whether CADD is good enough. As I just mentioned, molecular design is one problem, activity evaluation is another and molecular conformation generation is yet another. The druggability evaluation at the end involves a large series of problems that may be related to biological cell-level pathways and each problem has corresponding methods to solve it.

Some cases may still need to rely on experiments or on approaches such as patent analysis and patent reading. Some may be obtained through theoretical calculations. Some may be solved by generative models or by retrosynthetic analysis. Finally, you may be more interested in how the new agents can help. It is about streamlining and setting up each specific module. There are a great many stages and in the past each step was relatively fragmented. For example, after a molecule is designed in the first step, how does it enter the virtual screening process? After entering the training process, how can the molecules I have disassembled be further subjected to a relatively high-precision FEP calculation for activity evaluation? Basically, every intermediate step and every connecting process requires a professional chemistry or CADD background to carry out code operations. In the past, without an agent, it was basically necessary to rely on some basic CADD skills in order to use some tools well.

In the past, even for people who had only recently graduated and were just starting to work with pipelines, if they were given the best and most clearly defined tools, they might still not be able to use them well. The reason is that this is a very specialised field and the software itself is also professional. To operate it well, you often need to rely on your own understanding of the problem and of the principles behind the tools and methods you are currently using. Without that, you cannot do it well. At present, it is indeed feasible to have an agent do this work, especially with some better models, such as Opus or GPT, which are relatively advanced large language models. They are better able to do part of the work as a connecting link fairly well. Moreover, although the tools in use deal with fairly research-oriented problems in their specific subfields, there is generally a great deal of literature available on this subject. Therefore, current large language models have no problem with relatively simple understanding. That is, in terms of concepts and knowledge, they are only correct on some relatively cutting-edge issues and may not be very in-depth. Therefore, it is about chaining workflows for some tasks. Each specific part, such as the model for each scenario, needs to be developed by ourselves. It is just that the R&D factory will also get involved and use agents to intervene.

Human understanding is still very important. You have to do scientific research yourself. AI cannot help you with scientific research because it lacks much of the basic knowledge that you have. About 5 years ago, most of the products launched by Schrödinger and XtalPi were professional CADD software. However, over the past 1-2 years, even the tools released by Schrödinger itself have added agent dialogue boxes to the previous software and built in a relatively good language model to help you with part of the vibe coding operations and get the engineering done well. Therefore, I think at present, agents provide greater help in engineering. Because engineering is a relatively mature field, while science is the most uncertain. However, for engineering, no matter how you want it done, there are always corresponding standard operations and standard answers. Once standardised, it can be done.

Client [00:26:31]: Whether it is the US AI for science pharmaceutical companies or XtalPi, the core moat is still based on industry-specific domain knowledge. The papers, drug handling rules and experience accumulated over the past years are written into algorithms and then transformed into neural networks and large models. After moving from a single vertical model to multiple vertical models and with large models in place, these modules are strung together into workflows to improve overall efficiency. Is this logic correct?

Expert [00:26:31]: That’s right.

00:27:15] Many digital software tools for simulation and computing, such as Matlab, Data or Dassault Systèmes, already exist and are good tools. The best business model is to have a large model directly call these software tools and repackage them into new products for customers or for internal use. Is this the most correct approach?

Expert [00:27:15]: Yes, that is exactly what is being done at present. However, there is a problem here or rather a paradox, namely why doing so in many cases today does not lead to very good outcomes and commercially the revenue may not be much better than it would otherwise have been. The root cause is not the AI issue, but that the commercial software sector itself has a relatively limited market space, which may not be as large as people imagine. It is relatively stable. For example, Schrödinger is such a stable company and a benchmark in the industry. As for overseas markets, you can name all the large pharmaceutical companies in Europe and the US. All large MNCs are Schrödinger’s clients and are long-term stable clients. There is no issue with that. They use the software to perform various operations every day, but the revenue amount has always been this share, very stable. It may not be like developing a new pipeline now. Because we still need to distinguish the client from the vendor. For example, in drug R&D pipelines, if you bring a drug to market and the target and drug molecule are assets, the company can achieve very high revenue.

However, this is a different business model from selling software as a tool to a drug R&D enterprise and the revenue is completely different. Therefore, in terms of services, it supports revenue based on the order volume. It is not like a drug molecule becoming a hit as soon as it is launched. Back to China, the domestic market risk is probably based on this logic. It may not be like a new pharmaceutical company developing another drug or target pipeline. How much revenue can you get after bringing it to market? It is more of a tool provider and a vendor. Therefore, its revenue comes more from how many targets pharmaceutical companies have and how many pipelines can be initiated for services. What I mean is that even if the software is AI-based, it can at most make the software more user-friendly and help less professional users get started more easily. This is the role of AI. However, how to make this software more useful, helpful and able to accelerate drug R&D depends on the vertical model in each software. That is to say, you must solve the problems that are not done well now and see whether you can better solve some problems that might have been impossible before.

For example, 2-3 years ago, it might only be able to handle small-molecule modalities well in drug R&D. In recent years, many pharmaceutical companies have been leaning towards cyclic peptides or antibodies. As drug molecules become larger and larger, tools must have corresponding capabilities to conduct relevant research and improve their functions. The corresponding algorithms require some methods that might be scientific breakthroughs in chemistry. It may be able to handle not only small molecules, but also cyclic peptides and increasingly large molecules. It is very difficult to achieve both accurate calculation and consistency with practice. Therefore, I think doing well in this area depends more on whether the method itself is close enough to directly produce preclinical molecules without any medicinal chemistry intervention or experimental intervention. You must understand all previous stages clearly from a scientific perspective and you need to algorithmise all of this. This is the work we do every day now, which is a process of continuous improvement. Only when all these things are genuinely done well can it be said that this computational tool can truly and effectively assist current drug R&D and then it can simply be handed over to clinical trials.

00:33:22] I think XtalPi does drug R&D as a service to pharmaceutical companies and on the materials side, it provides research services for battery packs of car manufacturers for rapid verification. A service business and a software business are essentially very different. Services are relatively heavy and tiring work and revenue depends on client pipelines. It is difficult to attract new clients because the clients served are professional pharmaceutical companies and laboratories, not the general public, which is different from software. However, the capabilities of large models are advancing very rapidly and have penetrated thoroughly into fields that lean towards dry experiments, such as mathematics, programming or AI research. Although researchers have done an excellent job with operators and algorithms and accumulated a lot of industry knowledge, are you worried that large model companies or new RSI companies could also do all scientific research perfectly?

Expert [00:33:22]: I am not worried. Let me clarify the fundamental difference here. The reason why it can do well in coding, in deriving mathematical formulae or in solving some simple scientific problems is that these things are known and have fixed rules or fixed answers. This is not a particularly cutting-edge problem. For drug R&D or for research into battery structures of materials, we are often dealing with things for which humanity truly has no answers at present. It may require continuous attempts, whether computational, dry-lab or wet-lab, before some results can be obtained. Based on this result, we then adjust our approach and see whether there are other ways to solve it or to modify the existing model. Therefore, much of the time, what is being done is scientific research rather than repetitive science.

This is therefore different from the software we have discussed. Drug R&D software in particular is very special and software related to materials for battery R&D is also very special. It is not solving a known and mature problem. It is solving a problem that is still being studied. Therefore, various tools and software are iterated every day, with their specific details updated daily. Basically, the more time you give it, the more updated things you can produce. You have to do research every day because it is impossible to research this subject thoroughly. Unless, of course, one day medicinal chemistry really can be fully understood. That could be a great scientific victory for humankind. Of course, this is an ideal, so it may eventually turn out that way. However, at this stage, it cannot be done in the same way as AI deriving formulae or doing some coding, because coding is also a relatively fixed skill.

Therefore, I think it is still necessary to distinguish this issue. In some specific scenarios of drug R&D, human involvement is still required. AI cannot replace human understanding. Human experience and know-how are needed to guide AI in how to conduct the next step of research. Therefore, the current agent is a good helper. It is equivalent to a very powerful intern. He writes code well and has no problems with mathematics, formula derivation or theoretical research. But many things are things that only the real world has answers to and you have to feed it after doing experiments.

Client [00:42:33]: If the automation station is well developed, it will definitely improve the level of models, right?

Expert [00:42:33]: That’s right. An advantage of automated stations is that they don’t have to go off duty, provided that the response must be good enough and the degree of automation must be high enough. As long as certain specific reactions can be achieved to this extent, automatic stations can basically achieve the effect more cheaply than manual work. It doesn’t need to stop, so it can do synthesis more efficiently. It can obtain some models related to possible synthesis strategies and provide corresponding data regardless of positive or negative samples. This is a good job at present.

Previously, in the FTE pharmaceutical chemical synthesis business, it often relied on a senior sub-assembly leader to tell clients what reaction routes should be used to synthesise molecules, what reagents, reactor blocks and reaction conditions were needed. For example, whether LC-MS is needed for temperature, pressure and catalyst to do a good job in molecular experiments. In many cases, it is necessary to rely on some professional and senior sub-assembly supervisors to guide manual operations. At present, how to enable some senior personnel with experience to automatically predict the reaction conditions, reaction routes and reaction centres of a molecule is called a synthesis strategy.

In the past, AI could basically conduct experiments because the corpus that can be learned by AI is nothing more than experimental books made in CROs. The reaction records may be based on certain conditions, catalysts, temperatures and pressures, which may enable a molecule to be successfully reacted. Previously, there were only some scattered or positive samples. In the past, only successful experiments were recorded manually and it was impossible to record a failed experiment by yourself. I know it will fail, but I still do it. If I want to collect yield data, I couldn’t do such a thing in the past.

However, there will be such requirements for the automation of robotic arms in synthesis strategies. This is because the model needs to learn not only positive samples but also negative samples. I need to know what kind of conditions, catalysts and temperatures can make the yield good or bad. In the past, if there were only positive samples but no negative samples, it would be impossible to train the model. Therefore, the robotic arm was later asked to do such experiments because it didn’t have to go off duty. It knew that there would be mistakes, but it had to collect experimental data in between. Because it is produced by the robotic arm workstation, its data are not all programmed.

00:46:46] Finally, AI can completely achieve 24-hour uninterrupted operation without human participation in the experiment. The model completes all dry experiments online and the wet experiment is completed by controlling the automatic workstation, right?

Expert [00:46:46]: That’s right. This is the ideal case. It’s just that reality is still a long way from the goal.

Client [00:47:18]: Is XtalPi already going to do this?

Expert [00:47:18]: Yes, it is a consensus in this field. As we all know, large companies in this field basically have such an idea. This is not a particularly difficult idea and it has been incorporated into various synthesis strategies. There are many papers on it. It is just that this field is relatively new and may have only emerged in 2017. Even so, the concept itself is not new. Now I would like to add that what I have just mentioned is the field that is currently progressing well or the one in which we are more capable, such as synthesis reactions, that is, small-molecule synthesis in organic chemistry, which can be done relatively well. For this synthetic route, first of all, humans have carried out enough experiments in the past and organic chemistry is sufficiently well understood, so we can do such work.

However, for some work that needs to be done at the cell level or for biological experiments, a drug molecule needs to be injected into mice or rats to observe its half-life and see whether its activity is good. This type of biological experiment is still quite difficult and automating it is even more difficult. First, raising mice is a very non-standard process. In many cases, each mouse is in a different condition. Some mice may inexplicably develop a disease in the environment. This disease is not the one you intend to deal with but something else and it may affect your experimental results. Therefore, raising mice is often a problem in itself and then how to inject it into the mice is another problem.

The production method of this molecule may be relatively easy, but how to introduce it into mice and how to make it express and exhibit the corresponding activity make this experiment highly non-standard again. It may require observation under a microscope or processes such as extracting mouse blood or something similar. This process is currently difficult to automate, so anything related to biology in particular is a later topic. Of course, ideally we also intend to automate this, but it may not happen very soon.

00:50:37] Where are the opportunities for the latest wave of RSI companies? AI for science companies doing wet experiments such as XtalPi still have operational barriers, which are quite different from dry experiments, right?

Expert [00:50:37]: In this regard, everyone certainly has this concept in mind and hopes to apply such a process in their daily work, but it may not be that easy. In many cases, this is a scientific research problem. In many cases, human participation is required, namely human-in-the-loop. In AI terms, people need to be in this module and continuously iterate our knowledge together with the model. Therefore, in many cases, human intervention is still required in the daily R&D process, especially for CADD, that is, the specific algorithms in drug R&D. These algorithms still need human iteration and AI may simply carry out the specific tasks for you. Therefore, I think the moat here lies not in AI but in chemistry. More precisely, it lies with specific computational chemists or medicinal chemists. A professional understanding of this field is very important. In many cases, AI in this field is developed by ourselves rather than by people with computer backgrounds, but by people with medicinal chemistry or chemical backgrounds.

The most classic example is AlphaFold from DeepMind. Although it feels like an AI company, its team consists of biologists with a biological background to solve this problem because they must understand such problems well enough. This may not be considered science, but AI for games. An earlier example is Go. Of course, Go can also be said to be a competitive science. Anyway, the same is true for Go. The reason AlphaGo was developed so well at that time was, first of all, that the main researcher was a very senior amateur Go player from Taiwan who later joined Google to do this. You must have an understanding of this matter so that you can know how to write what the model needs to learn and solve this problem. This is the key. To develop AI in this field, it is necessary to have people in the field with the background to do AI, so as to find a way to solve this problem to the current level.

AlphaGo itself did a really good job later. In the early days, AlphaGo learned human game records and later it learned to play Go with AlphaGo itself. It learned from game records made by computers playing Go with computers themselves and it kept learning. Because Go just has rules, black plays first and then white takes turns. There are some basic rules for capturing stones. You can only play a certain number of moves at a time, so what must you do? What are the scoring rules, that is to say, how to win and how to lose? Comparing the points of both sides, these are very simple rules that can be written on one piece of paper. Therefore, next it is just a matter of knowing how to make each move through calculation to maximise and optimise the winning rate, which is basically the logic. So it is the same logic for humans playing Go as for AI playing Go. It is nothing more than a difference in computing power.

That is why AlphaGo can now crush human Go players. Because a computer's computing power is definitely higher than a human's. Given known rules, it can certainly compute faster than a human can. This is the advantage of computers and that is exactly the principle behind it. But when it comes to the materials or drug R&D we are now discussing, things are different. Because the problems are different every day. They are all scientific problems concerning specific targets. If such an answer really existed, there would be no need for medicinal chemistry or pharmaceutical companies today. We would simply produce drug molecules directly.