Cancer vaccines are not a new story.

Since surgeon William Coley injected Coley with streptococci into the tumor in 1893 and some patients recovered, immune cancer has oscillated between hope and disillusionment over the past century; Around 2011, this track had a golden period, but multiple Phase III products were repeatedly stalled due to targeting issues.

A turning point occurred in 2017, when two teams published similar early clinical conclusions in Nature on the same day: sequencing tumor mutations for each patient, tailoring vaccines, and activating specific T cell responses in melanoma patients.

However, in the following decade, the industry progressed slowly. Until August 2026, the bottleneck was finally broken.
Moderna and Merck announced that the personalized mRNA tumor vaccine intismeran combined with K-drug achieved dual endpoints of recurrence free survival and distant metastasis free survival in a phase III trial. This is the world's first personalized neoantigen therapy to obtain positive evidence in the phase III framework.

But just one week later, BioNTech's BNT122 was terminated in a phase II trial for colorectal cancer due to imbalanced survival, resulting in even more deaths in the vaccine group.

At this moment of division, I had a chat with Dr. Shi Yi, founder and CEO of Xieyi and Depushi AI Biotechnology, and a researcher at Shanghai Jiao Tong University, as well as his team. Since 2007, this team has been doing something that no one thought was possible at the time - exploring the combination of AI and cancer immunity.

After more than a decade of being on the cold bench, they waited for the day when the industry crossed the threshold.

01. Bringing the pharmaceutical factory into the human body

This is a paradigm shift, "Shi Yi went straight to the point.

Whether it's chemotherapy, targeted drugs, ADC, or bispecific antibodies, they are essentially "exogenous" treatment methods, designing a molecule from outside the body and delivering it into the body to attack tumors. The mRNA tumor vaccine takes another path: encoding tumor specific mutation information into mRNA, injecting it into the body, and activating the endogenous immune cells already present in the body to better recognize and eliminate cancer cells.
This is a therapeutic logic within the body.

In the interview, Shi Yi used an interesting metaphor: traditional cancer treatment is a "cat and mouse game", where a tumor clone corresponding to a target is knocked out, and another previously suppressed clone may begin to occupy the ecological niche and produce immune escape. The mechanism of personalized tumor vaccines is to cover more high-value mutation sites in patients' tumors under the premise of accurate prediction of new antigens, which is equivalent to sending a "wanted notice" to the immune system, allowing T cells to patrol and eliminate them on their own.

The key concept here is neoantigens - abnormal protein fragments (peptides) produced by tumor cells due to gene mutations (or viral infections) that are not expressed in normal tissues at all.

Each patient's tumor mutation profile is different, so vaccines must also be tailored to each individual. Another member of Shi Yi's team added an important distinction, pointing out that tumor vaccines are more targeted at TSA (often tumor specific antigens containing somatic mutations) rather than just TAA (tumor associated antigens).  

TSA mainly exists in tumor cells, so the probability of triggering autoimmune toxicity is very small; TAA is an overexpressed self protein in cancer cells, and normal tissues may also have basal expression, facing dual challenges of immune tolerance and off target toxicity.

In August, the BNT122-01 trial developed by BioNTech in collaboration with Roche was recommended to be terminated by an independent data safety monitoring committee due to an imbalance in overall survival values between the treatment group and the observation group. Earlier, the experiment had already reached the predetermined threshold of invalidity.

According to Shi Yi's analysis, there may be two key factors contributing to the failure of BNT122 in colorectal cancer. Firstly, there is no combination of PD-1 inhibitors, which were used in the first stage but not in the second stage; The second is that the tumor microenvironment of colorectal cancer itself is extremely complex (especially microsatellite stable MSS), with various immune suppression and tumor microbial issues, making it a significantly more difficult tumor to treat than melanoma. This section explains why Moderna's success in melanoma is inspiring, while BioNTech's colorectal cancer vaccine BNT122 has failed in colorectal cancer.

02. The key to "one person, one seedling"

The significance of Moderna and Merck Phase III data is that it proves that this "sequencing prediction design production" process can operate on a large scale. But what truly determines the success or failure of a tumor vaccine is the most crucial link in the chain: the accuracy of predicting new antigens.

Shi Yi repeatedly emphasized in the interview that a crucial aspect of personalized tumor vaccines is accurate prediction of new antigens. If not accurate, there will be no basis for subsequent delivery of LNP, combination therapy, etc.

This is not an easy problem to solve. Previously, the mainstream prediction strategy in the industry relied more on peptide MHC binding affinity as a screening label, but it has been proven that this is not enough. A more challenging issue is data scarcity, as algorithms typically require a large amount of labeled training data, which is extremely scarce.

Shi Yi has been exploring the combination of AI and cancer immunity since 2007. He received training in artificial intelligence at the University of Alberta in Canada and was a member of the AICML (now Google DeepMind) research group. Later, he worked as a postdoctoral researcher in computational biology at the University of Southern California. This background gives him a unique judgment on the application of AI in biomedical fields.

He mentioned that there is a Turing Award winner named Rich Sutton in the research team, who is a pioneer in the field of reinforcement learning. He has always emphasized that the current direction of large models may not be the correct path - "what can truly make models more effective is not to pile up more data and computing power, but to use continuous learning algorithms with long-term memory to compensate for the lack of data and computing power. This approach directly influenced Depp's technical roadmap.

Depp's AI model is completely developed independently from underlying algorithms, rather than calling general tools. Shi Yi explained that the review of new biomedical technologies is not just about individual molecules, but a complete set of therapies, including new antigen prediction technology, preparation process, and clinical protocol design. When applying, the entire protocol must be fixed: the algorithm must be determined and interpretable. If a universal algorithm is used, it is difficult to trace the weights and parameters, making it easy to be challenged.

Taking immunogenicity as an example, many early algorithms used the binding affinity between antigens and MHC as the most important screening label. However, in a large number of experiments, Depp found that many antigen peptides with low binding affinity actually have high immunogenicity.

It has a correlation, but it is not a very strong positive correlation. At that time, due to limited data, everyone thought those were irrelevant outliers, but we don't think so. ”Shi Yi added.

Based on this understanding, Depp constructed a prediction system that integrates AI algorithms, immune prior knowledge, and biological prior knowledge. Over the past decade, Shi Yi has led a team to conduct extensive research on the laws of somatic cell mutations, tumor evolution, and immune escape mechanisms in pan cancer, and has published over 100 international journal articles.

According to publicly available information, the accuracy of Depp's new antigen prediction has significantly improved compared to Moderna and BioNTech algorithms, and it can compress the number of antigen peptides that need to be designed from Moderna's 34 to around 10-15, thereby reducing direct production costs and more uncertainty caused by long sequences.

03. "Doing everything by yourself" is not feasible

Even if the prediction accuracy is industry-leading, the commercialization of personalized tumor vaccines still faces a structural challenge: it is a personalized pipeline that cannot dilute costs through mass production.

Shi Yi calculated that a domestic company that produces DC vaccines (dendritic cell vaccines) had a revenue of about 200 million yuan from approval in April 2024 to April 2026, corresponding to about 200 patients. If you are an mRNA company, achieving 500 to 1000 cases per year is already quite a burden. And among the new cancer patients in China every year, the number of patients who can only be resected reaches 1.5 million, which is a huge number based on a market penetration rate of 30%.

This means that relying solely on one enterprise to connect the entire chain from sequencing to new antigen prediction, from mRNA design to LNP delivery, and from production to commercialization requires a significant amount of investment, making it almost impossible to achieve optimal efficiency and cost. Shi Yi's judgment is that each company should focus on their own strengths and then piece together the puzzle.

This is precisely the underlying logic behind the strategic cooperation signed between Johnny Depp and CanSino in August 2026. According to the announcement from both parties, Depushi will leverage its AI led platform for tumor neoantigen discovery and design to be responsible for neoantigen prediction, screening, and sequence design optimization; CanSino provides mRNA vaccine delivery, formulation, and research and development production capabilities. Collaboration is global, focusing on the treatment of solid tumors and rare tumors in the digestive tract.

This "platform+pipeline" business model has its inevitability in the field of personalized treatment. The research and development investment of personalized tumor vaccines is highly concentrated in the early stages - sequencing, algorithm development, and immune validation. Once these capabilities are established, the marginal cost does not increase linearly with the number of pipelines. However, the delivery of LNP, GMP production, registration and application processes require heavy asset investment and long-term experience accumulation. Only by combining the two can a balance be found between time efficiency and cost control.

Shi Yi further elaborated on the vision of this model: "Only when more than ten or twenty personalized tumor vaccine pipelines have all emerged and been launched, can this market truly be scaled up

However, even if the technological path is validated and the production process is streamlined, there is still a thicker wall in front of personalized tumor vaccines: who will foot the bill?

According to investment bank William Blair's estimation, the wholesale purchase price of Intismeran from Moderna/Merck in the United States may reach as high as $475000 per patient, approximately 3.18 million yuan. This number refers to the pricing level of approved CAR-T cell therapy products. Industry estimates show that the production cost of personalized mRNA vaccines for a single patient ranges from 100000 to 300000 US dollars.

Moderna has not yet announced the final pricing, but the $475000 does not correspond to the "one shot" price - according to the Phase III plan, patients can receive up to 9 injections, with a treatment period of about one year.

In the United States, payment for innovative cancer drugs is relatively secure; But in China, the positioning of basic medical insurance is to "protect the basic", and a large number of innovative drugs with outstanding clinical value but high prices are difficult to be included in the reimbursement scope. The terminal pricing of CAR-T products in China is still around one million RMB, mainly relying on commercial insurance and Huimin Insurance to bear payment pressure.

This is an ongoing policy experiment. The first edition of the commercial insurance innovative drug catalog, which will be implemented from January 1, 2026, selects 19 "sky high priced drugs" from 121 applications that have passed formal review, providing the first "map" of commercial insurance covering high-value drugs outside of medical insurance. For personalized tumor vaccines, it is only possible to achieve true accessibility if a reimbursement coverage of 30% -70% is formed through Huimin Insurance and commercial health insurance, and the domestic pricing is reduced to less than 500000 yuan.

04. Pioneer on the cold bench

Between 2014 and 2018, a group of companies focused on personalized tumor vaccines emerged in China and entered early clinical exploration. However, due to technological bottlenecks, capital outflows, and changes in regulatory environments, a considerable number of these companies underwent transformation or exit.

This matter is actually very difficult. It's not because of the company's strategy that I want to open a pipeline for cancer vaccines, which requires a lot of technical accumulation in advance. "Shi Yi joked that his team has been sitting on the" cold bench "of personalized cancer vaccines for ten years. At that time, I chose this direction because I was very optimistic and believed that we could continue to work on it. There were too many problems that could not find a breakthrough point. In recent years, there have been many institutions that specialize in mRNA, but only a few still insist on developing tumor vaccines, very few. ”

This persistence paid off in August of this year. The Phase III data from Intismeran not only validates the feasibility of the technological path, but also prompts the entire industry to re-examine the industrial value of personalized tumor vaccines.

For platform companies like Depussy, the rising industry heat means more opportunities for cooperation and a faster pace of pipeline advancement. But Shi Yi is also aware that the reading of Phase III data is just the beginning. "It proves the feasibility of this path, and there is still a long way to go before it can be truly applied clinically and industrialized

The Boao Lecheng Pilot Zone in Hainan has opened a window for this road. In November 2024, Hainan passed legislation to clarify for the first time that eligible biomedical new technologies within the Lecheng Pilot Zone can be used for clinical research and translational applications, and treatment fees will be charged after completion of registration and price disclosure. Multiple personalized tumor vaccine technologies have been implemented and applied here.

Personalized tumor vaccines are moving from "concept validation" to "path validation". It may not generate a huge general market in the short term like PD-1, but it may establish irreplaceable therapeutic value in specific patient populations like CAR-T, and then gradually expand the boundaries in the long process of technological iteration and cost reduction.

This road is long, but the threshold has already been crossed.