Can AI Recreate "Semaglutide"? Billions of Dollars Pour into AI Drug Discovery

The critical point for AI-driven drug discovery has finally arrived.

This year, the overseas transactions of innovative drugs in China have been booming. Among them, AI-driven drug discovery companies are becoming a force that cannot be underestimated. From March to August 2025, companies such as Yuansi Biotech, Huashen Intelligence Medicine, and Jingtai Technology successively completed BD transactions with a total value of billions of US dollars.

Among them, Yuansi and Huashen, which have only been established for four or five years, have achieved success in the strict BD selection of multinational pharmaceutical companies. They have rewritten the R & D paradigm of macromolecular drugs and improved the R & D success rate through AI. These transactions have also given confidence to the capital market. After three or four years, several AI-driven drug discovery companies have recently received new financing.

BD transactions of some AI-driven drug discovery companies in 2025

The underlying R & D paradigm of biomedicine is shifting from the previous massive screening and experience-based approach to rational design and de novo creation.

If in the previous decade, AI, which made early drug discovery faster and more accurate, was just "icing on the cake"; now, AI with the ability of "de novo design" of drugs can be said to be "creating something out of nothing". Designing proteins that do not exist in nature and designing drugs for difficult-to-drug targets are the keys to conquering stubborn diseases.

Recently, after seeing the data of protein design and generation models released by Chai Discovery (invested by OpenAI), ByteDance, etc., Wang Chengzhi, the founder of Zhiyuan Shenlan, believes that a "qualitative leap" is coming. He has been working in the life science field for more than 20 years and was the former chief scientist of Mega Robot. "I wouldn't have made such a judgment last year. But the capabilities of AI this year have made me realize that there may be disruptive events in the biomedicine field soon, so soon that people won't have much time to prepare mentally."

Ma Rui, a partner at Fengrui Capital, also made a similar judgment. "The current momentum is 'innovative drugs multiplied by AI'. Although there is no final conclusion on the ability boundary of AI-driven drug discovery in the industry, I feel that the situation is surging, and there should be earth - shattering changes in the next one to three years." In the previous wave, leading the Series A investment in Jingtai Technology and the angel - round investment in Jitai Medicine brought huge returns to his institution. The cash - on - cash return multiple (DPI) from exiting Jingtai Technology could reach dozens of times.

So, how exactly does generative AI rewrite the underlying logic of new drug R & D? What changes might it bring to the innovative drug industry? What paths are current practitioners taking to approach "disruptive innovation"?

AI transforms drug discovery from "searching for a needle in a haystack" to "precise design"

The story of "AI-driven drug discovery" has not been without bubbles. Between 2018 and 2021, the AI boom represented by small - molecule drug development attracted tens of billions of dollars in investment but failed to fulfill the initial vision. At that time, under the path of deep learning and virtual screening, AI could accelerate the early drug discovery process, but the screened molecules were difficult to surpass existing drugs in efficacy, or the newly generated molecular structures were difficult to synthesize. This was essentially a limitation of the "inductive ability" of early models.

The new wave of AI is the result of the significant progress of mainstream AGI technology, with its capabilities spilling over into life science fields such as drug R & D and enzyme design.

First, the emergence of AlphaFold 2 verified that the Transformer architecture is also effective in understanding the "language of life" and solved the protein folding problem that had puzzled the biological community for many years. In the 60 years before the emergence of AF2, humans learned about the structures of about 200,000 proteins. But in the two years after AF2's appearance, AI has predicted the structures of more than 200 million proteins, covering most proteins of all known living organisms on Earth with a high degree of credibility.

Second, the David Baker team introduced the Diffusion model from the image generation field into biology. Using the principle of "iterative denoising", the success rate of de novo protein design has made an order - of - magnitude leap.

At the same time, AlphaFold 3 has evolved from only predicting the structure of a single protein to being able to handle the complex interactions between proteins, nucleic acids, and small molecules. This "full - atomic - level" modeling method gives it stronger generalization ability when the data is insufficient.

Currently, several domestic and foreign teams are replicating and improving the model prediction ability of AF3. In the middle of this year, the emergence of new models such as Chai - 2 released by Chai Discovery, ESM3 released by Evolutionary Scale, and Protenix announced by ByteDance verified the creativity of "generating new functional molecules from scratch".

"The latest data released by Chai - 2 shows that for specific targets, the hit rate of the candidate antibodies it generates is significantly higher than that of traditional methods. In the past, only a few positive molecules could be screened out from a library of millions to billions, but now hits may appear in dozens of sequences. This was unimaginable before." Wang Chengzhi told 36Kr. This means that the problem of "generating antibodies for specified target epitopes" is close to being solved.

Traditional antibody drug R & D, from determining the target, immunizing animals to screening out effective antibodies with high affinity, is a long - term "searching for a needle in a haystack" process. In the past, it might take three years and $5 million to solve the problem of antibody molecule discovery. AI models such as Chai - 2 can complete it within a few hours and get verified through biological wet experiments within two weeks.

Ma Rui also expressed a similar view. Chai Discovery may disrupt the R & D paradigm of antibody drugs. Previously, methods such as hybridoma technology, phage display, and animal immunization may be largely replaced by de novo computational design. If AI can also make breakthrough progress in small - molecule design, "almost all drug modalities may be empowered by AI." He said bluntly: Everything is happening so fast that not many people have really seen this picture.

By then, when a medicinal chemist wants to design an antibody drug for a certain target, the first reaction may no longer be to immunize animals. Instead, they will first use AI models for calculation, generation, and scoring, select dozens of the most promising antibody sequences, synthesize them, and conduct in - vitro experimental verification.

This change in the drug R & D paradigm will bring profound changes to the innovative drug industry chain.

Now, the low - hanging fruits of innovative drugs have been picked. The R & D of "difficult-to-drug" targets that could not be advanced due to the lack of lead compounds is expected to be re - activated by AI. From the previous "screening what is available" to the AI era of "creating what is needed", humans may be able to conquer some stubborn diseases for which there are no available drugs, and some drugs with great side effects may be replaced by better - emerging molecules.

Wang Chengzhi believes that AI will significantly shorten the pre - clinical drug R & D cycle, which is beneficial to indications in fields such as oncology, autoimmunity, and metabolism. Among them, the first to see the dawn may be "chronic diseases". In the future, the frequency of the appearance of "miracle drugs" like semaglutide, which only appear once in a century, will be greatly increased. On the contrary, due to the popularization of computational tools such as AI, the commercial value of traditional drug screening platforms that rely on large - scale animal models will be affected to a certain extent.

In the future, biotechs with AI capabilities will become the "molecular design centers" and "computing power centers" of multinational pharmaceutical companies, responsible for high - tech - density and high - frequency drug discovery at the front - end. Multinational pharmaceutical companies will be more responsible for later - stage clinical trials, registration, and commercialization. The two sides will share the market through pipeline BD authorization, cooperative R & D, and other models.

With limited funds, what to do first?

Every technological revolution brings about industry reshuffling, and all parties are predicting and choosing the players with the greatest chance of winning. As the critical point approaches, there is not only one path to the future. Currently, the participants in the AI-driven drug discovery field generally present three forms.

The first type includes technology giants with sufficient capital and computing power, such as Google (DeepMind), Meta, Xaira (with a $1 billion seed - round financing), and ByteDance. They are committed to building basic biological large models, creating their own open - source ecosystems, and defining industry standards.

The second type consists of startup teams led by top AI large - model and bio - computing scientists. They have the ability to explore in the "uncharted territory" of algorithms. After optimizing and transforming the basic models, they provide platform services for pharmaceutical companies and biotech companies or conduct self - developed pipelines. Typical examples include Baitushengke, Huashen Intelligence Medicine, Yingxi Intelligence, Fenzi Zhixin, and Bai'ao Jihe.

The third type is the "traditional regular army" using AI to develop new drugs. They do not pursue self - developing a new basic model. Instead, based on their insights into indications, targets, and pipeline competition, they use AI open - source models and strong wet - experiment capabilities to accelerate the R & D process of drugs for specific diseases.

In Ma Rui's view, the core competitiveness of players in the AI-driven drug discovery track should be their ability to understand, modify, and evolve models. From the results of many people using open - source models such as AF3 for benchmarking, relying on fine - tuned open - source models may achieve an "80 - point" performance. But to solve the complex problems in real R & D, a performance close to "99 points" is often required. Only by achieving the best in the algorithm can leapfrog progress be achieved.

Another expert in the life science field pointed out that most companies cannot afford the high cost of self - developing basic biological models. Compared with large language models, the cost of obtaining data for biological models is higher. He mentioned that a domestic company supported by a giant once invested tens of millions of yuan to synthesize and test tens of thousands of AI - generated antibody sequences and used the experimental data set for model training. But in the end, they found that such a data scale was still far from being sufficient to enter the efficient improvement range described by the scaling law.

Wang Chengzhi expressed a similar view. He believes that in the future, teams that can quickly and massively produce high - quality biological experiment data are more likely to have high - performance AI models." In the past, automated and high - throughput experiments were often understood as improving screening efficiency. But in the AI era, they can not only provide experimental results for R & D personnel but also efficiently generate structured and iterable data, directly serving model training and optimization.

Algorithms and data are both core elements for improving the capabilities of AI-driven drug discovery. Although the backgrounds of founders and team advantages of each company are different, and there are certain differences in short - term resource investment focus, the industrialization path of AI in the pharmaceutical field is getting closer. That is, to develop truly valuable drug molecules and gain the recognition of buyers and real money in the current mature BD system.

"According to the current trend, all new drug R & D companies will use AI in the future, just with different degrees of dependence. For AI-driven drug discovery companies, only emphasizing the model is not enough. In the end, they still need to deliver drug molecules to be given higher value." Ma Rui summarized. "So I've been saying recently: Innovative drugs are AI-driven drugs, and AI-driven drugs are innovative drugs."