Dialogue with Ren Feng from Insilico Medicine: Young AI4S entrepreneurs, I admire your ambition and creativity.
By Hu Xiangyun
Edited by Hai Ruojing
In 2026, around 51-year-old Ren Feng, more and more young peers in their early 30s have begun to appear.
Ren Feng is the Co-CEO of Insilico Medicine (03696.HK). In the first half of this year, Insilico Medicine officially announced profitability. 35 pipeline assets developed based on the generative AI platform Pharma.AI have entered the PCC (Preclinical Candidate Compound) stage. The disclosed amount of BD cooperation based on the AI technology platform and pipeline has reached 11 billion U.S. dollars, ranking first among domestic AI pharmaceutical enterprises.
The pharmaceutical industry where Ren Feng works originally values industrial accumulation and R&D experience the most, which can be described as a bit "seniority-oriented". But now, AI is quietly rewriting this set of rules that have operated for many years. With the capabilities of large language models, young doctoral students are trying to turn traditional empirical science into engineering problems, so at the age of 30, they can step onto the front stage as entrepreneurs and obtain considerable financing and high valuation.
The generational gap between them and the "Ren Fengs" is extremely distinct.
The latter group mostly faced the capital market and the public with the image of cautious and restrained traditional scientists in the past. "Deliver 80 points and present 60 points" was the survival philosophy of that era. It is never wrong to be more conservative, after all, being too radical often leads to failure. Just like Ren Feng, in 2020, when he received the invitation to join Insilico Medicine, this "veteran" from the traditional pharmaceutical system first spent nearly half a year as a consultant, repeatedly verifying whether AI drug discovery was "real or fake". He did not officially join until he confirmed that AI "can indeed improve efficiency and create highly innovative things in specific scenarios".
But today's young entrepreneurs, whether in personal expression or their belief in AI technology, are far more passionate. In the eyes of investors, they are "more open", and can describe the grand vision of using AI to subvert the pharmaceutical industry without any burden, and set goals for their small teams to benchmark against Isomorphic Labs and Chai Discovery. Among them, according to the new round of financing being promoted by Isomorphic Labs, its post-money valuation can reach at least 40 billion U.S. dollars.
To a certain extent, when communicating with Ren Feng, you will feel that he is a rare "intermediate state" existence among the current active players. He comes from the traditional pharmaceutical system, operates the most cutting-edge AI system, can perfectly integrate the two sets of skills, and is sufficiently sober:
Ren Feng believes that AI will greatly change the old new drug R&D system, so he requires himself to "keep learning and evolving, and never let experience overwhelm judgment"; on the other hand, he is alert to the carnival of large model narratives in the current market, always emphasizes the importance of "implementation", and is willing to take the initiative to talk about many shortcomings of AI model capabilities. In his view, in the final analysis, getting the drug developed is the core, which always ranks first.
The two traits of the old and the new converge on Ren Feng. How does he view this generational transition, and how does he measure the distance between ambition and capability?
Dialogue (edited):
(This interview was conducted during the first "Great Power New Drug" Global Conference CPIC 2026. Thanks to the support of the conference organizer "Tongxieyi".)
Veteran, your experience is no longer that valuable
36Kr: In 2026, the AI4S entrepreneurial ecosystem is exceptionally prosperous, with more and more diverse participants pouring into the pharmaceutical track, and the narrative of "AI subverting new drug R&D" is becoming increasingly grand. From the perspective of a "veteran" with strong practical experience, to what extent has the pharmaceutical industry applied AI?
Ren Feng: Based on my own observation, when ChatGPT first became popular, China did not have a strong base model, and our actual perception in China was not strong. Since last year, first, China's own model products have become stronger and stronger, and second, the breakthrough in coding capabilities has allowed the pharmaceutical industry to iterate higher-quality algorithms at lower costs. Now inside Insilico Medicine, 70%-80% of the code is written by AI.
At the same time, AI's say in the entire drug R&D decision-making process is also constantly increasing. About three years ago, drug development was still a traditional trial-and-error process: after medicinal chemists designed molecules, they handed them over to CRO for synthesis and testing, and AI was used as an efficiency improvement tool in different links; now, AI has been integrated into scientists' decision-making process, and its weight is not low. For example, even if the molecule is first designed by humans, we will hand it over to AI to run again to check whether its properties are good and its activity is strong. Then, scientists will combine the results predicted by AI to decide whether to proceed with synthetic testing. This was completely decided by humans in the past.
The benefit of doing so is that it can save a lot of manpower, time and trial-and-error costs. Once AI calculates and finds that it does not work, it will pass the project and will not push it to the synthesis stage.
36Kr: In the past, everyone positioned AI as a "tool", but now it can participate in "decision-making". What does this mean?
Ren Feng: A very obvious trend is that the pharmaceutical industry's dependence on the experience of senior scientists is declining, but more opportunities are given to young scientists who actively embrace advanced technologies.
In the past, drug R&D was an empirical science, so the more experienced "old veterans" were highly respected, and the executive teams of traditional large pharmaceutical companies were mostly in their 60s and 70s. But in the context of AI, we are committed to shifting the new drug R&D paradigm from experience-driven to data-driven. At this time, the experience of mature scientists often becomes a constraint. They have succeeded before, so they will subconsciously believe and adhere to their own judgments first, and remain skeptical of the results obtained by AI, especially when the conclusions of the two are opposite.
On the contrary, young scientists with scarce experience are fearless. As long as they think the AI method is worth trying, they will take action, and often achieve impressive results. There are many such stories in our company, which makes everyone very surprised, and also makes me more and more cautious about my own experience. I must never say things like "according to my experience, you can't do it" easily.
36Kr: In the industry that values experience the most, experience is no longer that useful, and the "old veterans" also have a sense of crisis.
Ren Feng: Actually, I think this is a good thing, and it is the only way to break through the constraints of seniority and experience and make enterprises more creative and efficient.
In the past, large pharmaceutical companies were sometimes not so determined and inefficient when deploying AI. This is because if the decision-makers in the pharmaceutical industry are always people who only advocate experience, the application of AI will be difficult to develop. Even enterprises that are positioned as AI pharmaceutical companies must continuously iterate their AI capabilities. Some overseas AI pharmaceutical companies that started businesses at the same time as Insilico Medicine have relatively slow project progress, perhaps because they ignored this point.
36Kr: The generational gap among practitioners is becoming more and more obvious, and so is it among enterprises. Has Insilico Medicine ever been asked if it is an "older generation AI pharmaceutical company"?
Ren Feng: Enterprises develop in waves. When the AI pharmaceutical companies of Insilico Medicine's generation started their businesses, there were no large language models, and no one talked about AGI. Everyone mainly built their business based on vertical models. The advantage is that the number of parameters is small and the computing power demand is not high. The disadvantage is that a single modality can only process one or several types of data and solve one problem, while drug R&D requires optimizing many parameters at the same time, so multiple small models can only be called. But after the emergence of large language models such as ChatGPT, the new generation of AI pharmaceutical companies began to focus on multi-modality, integrating different small models relying on agents, which may be more efficient.
This technical difference determines that there is a generation gap between AI pharmaceutical companies. But the good news is that after years of accumulation, for example, Insilico Medicine has some internal models, and we can integrate these independent modules. Our experience, data and test sets for training these independent modules can help us enhance the reasoning ability of open-source LLM in the vertical field of drug R&D. For example, we also released the life science foundational model training framework Science MMAI Gym this year.
If you do not take the initiative to evolve, you will be eliminated soon.
AI for drug discovery, where are we now
36Kr: Model capabilities are evolving. Now, what things is AI good at in drug discovery, and what things are still not working well?
Ren Feng: Overall, AI pharmaceutical is still in a relatively early stage of development, and there is still a considerable distance from "super artificial intelligence", that is, GPT for drug discovery, and the maturity of the full-chain technology of AI pharmaceutical.
In the most upstream of new drug R&D, AI can already well carry out underlying problems that were difficult to break through in the past, such as the discovery of novel targets. For example, it can quickly summarize the existing research conclusions in public literature, identify the associations between targets, help researchers find targets that were easily ignored in the past, or deepen the understanding of the biological mechanism of targets. In the past, building a target-related knowledge graph required huge manpower, but now AI can do it directly.
However, the shortcomings of current large models are also prominent. Since all the learning materials come from published and reported studies, large models cannot discover out of thin air when facing completely unknown biological associations that have no literature records. For example, our PandaOmics (one of the modules of Pharma.AI, used for multi-omics data analysis) can directly analyze patient multi-omics data, compare the omics differences between patients and healthy people, and locate potential targets with abnormally high expression and excessive activation in lesions from raw data, many of which have no literature reports. And this ability to mine novel targets from raw clinical omics data is not possessed by large language models.
36Kr: Going one step further from the underlying biological problems, this year everyone is also talking about automatically putting forward scientific hypotheses through Agent, completing the experimental closed loop, and verifying different hypotheses in batches. How do you view this direction and have you made relevant attempts?
Ren Feng: This is essentially using Agent and cloud tools to reconstruct and optimize the scientific research workflow, taking over the links that originally required manual work, such as using cloud tools to assist experimental design. This is a clear general trend in the industry, and we are also making relevant explorations to involve Agent in our currently verified and effective specialized algorithms and automated laboratories.
However, there is an unavoidable prerequisite at this stage: the automated laboratory cannot be directly used to run brand-new experiments. If an experimental protocol wants to be launched on the automated platform, it must first complete a round of verification in a manual laboratory, calibrate the results, and ensure that the data output by automation is aligned with the data of manual experiments.
There are a large number of adjustable parameters in the experiment itself, such as enzyme concentration, reaction temperature, incubation time, etc., which need to be iteratively optimized, and there is no one-step optimal solution. Skipping manual verification and directly handing it over to the automated system has high risks. Therefore, although the prospect of this direction is clear, the current overall technical capability is not sufficient to realize fully autonomous closed-loop scientific research.
36Kr: Many teams will choose to use open-source models such as AlphaFold2 and ESM as the base. Can innovative pharmaceutical companies get usable prediction results by superimposing their own vertical data for development based on the open-source base?
Ren Feng: Relying on open-source models and superimposing the unique internal R&D data of the enterprise can indeed effectively improve the efficiency of drug R&D, which has been verified. However, the performance of the open-source model itself is limited. The team needs to continuously optimize and fine-tune on the open-source base based on self-developed data to obtain a domain model with better effect than the open-source original version. The parameter tuning is relatively difficult, not everyone can do it.
36Kr: After the "full-chain technology" you mentioned matures, in which target or disease areas may AI pharmaceutical achieve breakthroughs first?
Ren Feng: In fact, AI itself has strong generalization ability. Therefore, AI has no instinctive selectivity for disease fields and targets. As long as there is enough data, the accuracy of drug R&D in any disease field will be improved accordingly.
Fields with large disease types and larger patient cohort data volumes are more likely to find better targets first, so the market expectation is relatively higher. For example, in oncology, many cancer patients have undergone gene sequencing, and the industry has accumulated a large amount of data; in addition, diseases with a large patient base such as autoimmune diseases and neurological diseases are also promising.
36Kr: If AI itself is unrestricted, logically we should see "blossom everywhere" style of novel target research. But at present, the candidate molecules listed by many AI pharmaceutical companies are still relatively concentrated, for example, everyone is working on GPCR. Why is this?
Ren Feng: It is definitely not AI that causes homogenization, but the people who use AI.
In the ideal state, AI pharmaceutical should always solve the most difficult and highest failure rate problems in the industry. However, due to factors such as project delivery pressure and large pharmaceutical companies' BD preferences, AI pharmaceutical companies will choose targets of medium difficulty as the first choice when verifying their technical capabilities, otherwise there will be risks of relatively longer verification cycle, greater investment, and the possibility of failure in the end.
Targets that are relatively easy and supported by medical evidence in traditional drug R&D are also easier for AI. For example, GPCRs and kinases themselves have clearer binding pockets suitable for druggability; on the contrary, if we work on protein-protein interaction, current research has not found a good binding pocket, and there is no way to intervene with small molecules.
Therefore, the common practice is to accelerate R&D and improve creativity through AI, first make relatively easy targets, accumulate capital and experience, improve platform capabilities, and then gradually expand to more difficult fields.
36Kr: Even in the context of AI pharmaceutical, there will be cases of picking "low-hanging fruit" first?
Ren Feng: Yes.
36Kr: If AI is also working on the same target and indication, will there be "involution" in the end?
Ren Feng: I don't think so. Because different algorithms and different people lead to different results. Even inside Insilico Medicine, different R&D personnel use the same AI platform to screen for the same target, and the final results will be different. Because everyone sets different molecular property parameters, and there are a lot of parameters, such as water solubility, membrane permeability, affinity and so on.
In the final analysis, drug development does not pursue a unique solution, and the market will not be dominated by one company. Just like the GLP-1 target, there are various molecules on the market such as semaglutide and tirzepatide.
Survival of the fittest, drugs are the final card
36Kr: Insilico Medicine also cooperates with some AI technology companies, such as Google Cloud and Tencent Health. How are large manufacturers doing in cross-border pharmaceutical?
Ren Feng: We can clearly feel that their model capabilities are very strong, but at first they did not understand what problems the industry really wants to solve, so they will constantly look for AI pharmaceutical companies like us and multinational pharmaceutical companies to get demands.
I remember once the Google team discussed with us, talking about using algorithms to study biological mechanisms, such as gene interference, predicting the impact of genes on cells. To be honest, they did a really good job. But after listening, I found that they kept talking about biology, but missed chemistry, and never mentioned how to design the molecules.
What we really want is to apply the strengths of technology companies to the pain points of drug R&D. For example, Google has a large amount of patent and text information. If a platform can be built to analyze the coverage of compound structures in patents, allowing us to put the designed molecules on the model for evaluation, quickly judge whether there is patent risk, which can save a lot of time. Now the cooperation between the two sides is to "break patents" together.
36Kr: The influx of various new players into AI4S and the pharmaceutical field also makes the industry worry about the re-emergence of "bubbles". What do you think is the most necessary to guard against now?
Ren Feng: I think everyone needs to figure out what the final implementation scenario is. At present, capital is still concentrated in the AI field, and many companies with unclear business models will appear under the technology bubble, which is the most needed to be guarded against.
In the past two years, many practitioners with AI