Incubated by Tsinghua AIR, AI4S enterprise AIPharma has secured an angel round of financing at the level of 100 million yuan | 36Kr Exclusive

Text by Hu Xiangyun

Edited by Hai Ruojing

36Kr has learned that AIPharma Intelligence, an AI-native new drug R&D enterprise, has recently completed an angel round of financing of 100 million RMB. This round of financing is led by Xianghe Capital, with participation from institutions including Xinglian Capital and a well-known industrial fund.

It is reported that the raised funds will be mainly used for the R&D and iteration of basic drug models such as DrugCLIP and the AI drug R&D agent engine PharmAgents, continuously explore next-generation AI drug R&D technologies, and promote independent and collaborative innovative drug pipelines. At present, the company's fastest-growing self-developed pipeline has entered the Pre-PCC stage.

In addition, in terms of commercialization, AIPharma Intelligence mainly adopts the Co-development model instead of pure AI technology services. According to the introduction, the company has carried out co-development with more than five pharmaceutical enterprises and biotech companies in innovative drug discovery projects in metabolism, CNS, autoimmunity, drug delivery and other directions, covering different drug modalities such as small molecules and peptides.

AIPharma Intelligence was officially established in July 2026, incubated by the research group of Professor Lan Yanyan and Academician Zhang Yaqin from the Institute for Intelligent Industry Research (AIR) of Tsinghua University. Gao Bowen, a core member of the team, previously focused on interdisciplinary research of AI and basic science, worked on recommendation algorithms at ByteDance, returned to Tsinghua AIR to pursue a doctorate in 2022, published many papers in top academic conferences, and co-authored a paper published in the main issue of *Science*.

Gao Bowen introduced that although AIPharma Intelligence is "very AI-oriented" in essence, the team also attaches great importance to integrating with real new drug R&D scenarios to avoid AI technology being separated from experimental implementation. At present, the team has R&D heads with backgrounds in medicinal chemistry and structural biology, and the head of business cooperation has nearly 20 years of experience in the pharmaceutical industry, having promoted multiple drugs to enter Phase II and Phase III clinical stages.

"The reason why we completed such a team configuration at an early stage is that we hope to abandon the traditional relay working mode of 'AI completes calculation and delivers it to medicinal chemistry, then medicinal chemistry delivers it to biology', so that R&D personnel in different links can enter the closed loop of collaborative work and decision-making from the first day, and build a real AI-native new drug R&D system with AI as the core productivity," Gao Bowen explained.

In the traditional pharmaceutical model, AI is mainly embedded in the R&D chain as a tool plug-in. This is because enterprises usually split business lines according to molecular modalities and disease tracks from the organizational structure level. If necessary, each business line will independently build its own AI capabilities. For example, when encountering molecular design problems, they will develop a molecular generation model, and when it is necessary to predict druggability, they will develop a corresponding prediction model. It is difficult for AI technologies between different business lines to communicate with each other.

But "the logic of AI Native is the opposite". Gao Bowen believes that if you want AI to truly become the underlying core capability of an enterprise, you need to "radiate all businesses outward with AI as the core", redesign the organizational structure, R&D process and human-machine collaboration mode, and let AI participate in answering fundamental questions such as which experiments to carry out, what experimental data to produce, and how to connect calculation and wet experiments.

In AIPharma Intelligence, this system operates relying on three layers of AI capabilities: "exploration", "decision-making" and "learning and evolution":

First of all, the exploration layer allows AI to break through the existing cognitive and computational boundaries of human beings, and tap into the huge drug R&D space that is difficult for traditional methods to reach; the decision-making layer requires AI to continuously judge the next action in the face of a large number of uncertainties in drug R&D, and screen experiments worthy of implementation and data that needs to be produced; the learning and evolution layer needs to continuously receive feedback from real wet experiments and experts, and push the whole system to iterate and optimize continuously. The three layers of capabilities are closely linked, so that AI can jump out of the tool positioning and deeply participate in the whole process of new drug discovery.

Around this system, the AIPharma Intelligence team first built a number of basic models represented by the high-throughput virtual screening model DrugCLIP as the technical base to support the team's capabilities at the "exploration" level.

Gao Bowen introduced that the core of DrugCLIP is to use AI to build a high-dimensional representation space where proteins and molecules coexist, just like a "chemical universe". Traditional virtual screening requires one-to-one alignment calculation of proteins and candidate molecules. When facing compound libraries of billions or trillions of levels, the calculation cost will expand sharply; therefore, DrugCLIP does not follow this path, but uses a two-tower contrastive learning architecture to map protein pockets and small molecules to this high-dimensional representation space. In this way, when facing a new target, scientists only need to output the coordinates of the protein pocket in the space, and retrieve molecules with adjacent spatial distances to complete the screening, without having to calculate massive molecules one by one.

"Relying on this model, we can achieve virtual screening at the scale of the human genome. A single server can complete 10 trillion protein-molecule pairing calculations a day, supporting second-level retrieval of trillion-level compound libraries. According to our practice, the hit rate of some optimized targets reaches 60%-80%, and high-activity molecules of nanomolar even picomolar level can be directly screened at one time without subsequent transformation."

Based on the spatial exploration capability of DrugCLIP, the AIPharma Intelligence team further built the PharmAgents drug discovery agent engine to help improve the capability of the "decision-making" layer.

Not only the early molecular screening, the whole process of new drug R&D is full of a large number of complex decisions, for example, after a batch of molecules are screened, which ones should be verified first? When the experimental results conflict with expectations, which part of the evidence should be trusted? How to balance activity, selectivity and pharmacokinetics... These decision-making processes need to integrate multi-source information such as literature patents, model outputs, experimental data, and expert experience, and such information is often incomplete or even contradictory.

Different from some agents on the market that can only replicate human scientific research processes, the core capability of PharmAgents is to comprehensively process conflicting and incomplete multi-source evidence, then independently call various professional tools to put forward scientific hypotheses, and iterate and update the R&D plan after obtaining new information.

"However, we do not pursue to let agents completely replace human experts, but to enable the two to collaborate. Because the experience of human experts itself is also an important source of evidence in the decision-making process, the decisions output by agents will be evaluated and corrected by human experts, and these feedbacks can help agents continuously learn and evolve. In other words, our 'learning and evolution' layer capability is reflected in the continuous iteration of the whole system. The success or failure of execution will be precipitated as system knowledge, making the system stronger as it is used more," Gao Bowen said.

This system has been implemented in actual drug R&D scenarios. At present, AIPharma Intelligence's fastest self-developed pipeline focuses on ADHD (Attention Deficit Hyperactivity Disorder), exploring new mechanisms of action with differentiated potential through the newly discovered allosteric pocket; at the same time, the company also cooperates with different types of scientific research institutions such as the School of Life Sciences and School of Pharmaceutical Sciences of Tsinghua University, Chinese Academy of Agricultural Sciences, and Translational Medicine Research Institute of Shanghai Jiao Tong University to carry out extensive experimental verification.

Since the beginning of this year, the AI4S track has continued to heat up, a large number of start-ups focusing on the pharmaceutical direction have emerged, and the industry evaluation standards have become increasingly pragmatic. In addition to papers and demonstration demos, real and reproducible experimental verification has become one of the hard standards for testing technical value.

Gao Bowen believes that domestic AI pharmaceutical enterprises have completed two generations of evolution: the first generation of enterprises mainly use AI to solve single-point tasks in drug R&D; later, the industry began to pursue the use of AI to empower the whole R&D process and the construction of automated laboratories. "Now, the core work of the new generation of AI-native pharmaceutical companies such as AIPharma Intelligence is to let AI take responsibility for R&D decision-making. Traditional pharmaceutical companies are burdened with historical business baggage, and it is difficult to completely restructure their organizations and processes, while we build the system from scratch, which will be more flexible in this regard."