"BioMap" schließt strategische Finanzierung in Höhe von mehreren hundert Millionen Yuan ab und entwickelt "Mikrowelt-Modelle" für die Biowissenschaften | Exklusiv von 36Kr
Text | Hu Xiangyun
Editor | Hai Ruojing
36Kr learned that recently, the AI-native biotech company BioGeometry has completed a strategic financing of hundreds of millions of yuan. The leading investors are Shanghai Biomedical Innovation Transformation Fund, Guoke Investment, Fortune Capital, and Starlink Capital, with follow-on investments from Gao Rong Capital and Index Artificial Intelligence Industry Innovation Fund. Index Capital serves as the exclusive financial advisor. It is also reported that the funds raised in this round will be mainly used for the continuous iteration of the life science micro-world model GeoFlow and the advancement of self-developed drug pipelines.
Currently, artificial intelligence is evolving rapidly along two main lines: Digital AI represented by large language models and multi-modal models, and Physical AI represented by autonomous driving and humanoid robots. Bio AI may become the next most imaginative frontier, and this judgment is being continuously confirmed by top global capital and the scientific community.
In 2024, the Nobel Prize in Chemistry was awarded to both protein structure prediction and de novo protein design. In 2025, the BD transaction value of innovative drugs in China reached 135.7 billion US dollars, accounting for about 49% of the global transaction value, surpassing the United States to become the world's largest market for external licensing of innovative drugs. In May 2026, Isomorphic Labs, a subsidiary of Alphabet, completed a financing of 2.1 billion US dollars, led by Thrive Capital, with sovereign capitals such as Temasek, MGX, and the UK Sovereign AI Fund participating, setting a new record for single financing in the field of AI drug discovery.
The basic unit of life science is the cell composed of molecules. All functions of life originate from the interactions of molecules at the atomic scale. Whether we can truly understand and precisely design the interactions between molecules at the micro level is the most fundamental proposition in life science. GeoFlow, self-developed by BioGeometry, is such a "micro-world model". By precisely modeling the interactions of biomolecules such as proteins, DNA, and RNA at the atomic level, it can create entirely new molecules that have never existed in nature through generative AI, taking us from "understanding life" to "designing life".
Since its first release in 2024, GeoFlow has undergone three upgrades. GeoFlow V1 implemented the core idea of "modeling molecular interactions with atomic-level precision" into an engineerable large model. In the key task of predicting the structure of protein-protein complexes, GeoFlow V1 achieved the same level as AlphaFold 3. When it was iterated to GeoFlow V2 in April 2025, it was no longer limited to simple structure prediction but achieved the unity of protein structure prediction and "de novo design" ability at the atomic level, capable of tasks such as de novo antibody design, vaccine design, and industrial enzyme optimization.
"Objectively speaking, GeoFlow V2 can generate corresponding binding molecules for some targets, but there is still room for improvement in the binding affinity of the obtained molecules. Therefore, in October last year, we iterated GeoFlow V3. The core goal is to increase the success rate of generating binding molecules and obtain as many high-affinity binding molecules as possible, reaching the nanomolar (nM) level," said Tang Jian.
To achieve this, BioGeometry applied the Test-Time Scaling (TTS) technology in the field of large models to protein design. Simply put, this is an idea of exchanging inference time and computational input for design quality. Specifically in protein design, it means generating multiple protein versions at once for the target, then screening out high-quality samples, verifying and optimizing them, and finally obtaining new proteins with stable structures and qualified affinity. Its core advantage is low cost and fast implementation. Without investing a large amount of money to retrain the model, it can improve the success rate and quality of protein design.
Take the actual antibody design application scenario as an example: In the era when scientists still had to do it manually, the R & D team usually built a molecular library of hundreds of millions of molecules in animals or in vitro and carried out multiple rounds of high-throughput screening that took several months, with high costs and a long process. Now, under the "leadership" of AI, higher hit rates can be achieved with less experimental input.
According to the data disclosed by BioGeometry, in the de novo design tasks for more than 20 targets such as TSLP, IL-33, IL-13, CCR8, PD-1, H3-HA, and IL-4Rα, only no more than 50 candidate molecules designed by GeoFlow V3 need to be synthesized and verified for each target to obtain epitope-specific nM-level binding antibodies, with an average hit rate of nearly 20%, and the discovery time of lead molecules can be shortened to within three weeks.
Currently, BioGeometry is developing the next-generation micro-world model GeoFlow V4, which will expand the modeling scale from molecular interactions to "designing molecular systems" from "designing single molecules".
In the past two years, BD transactions have become an important commercialization outlet for innovative pharmaceutical companies. In previous cases, the value of drug pipelines usually increases as the clinical stage progresses. This can also be seen from the fact that most of the large-scale transactions worth billions of dollars occurred after Phase 2 or 3 of clinical trials. However, the "de novo design" ability of AI may further amplify the value of molecules in the early stage, thus rewriting this logic.
Tang Jian believes that for antibody molecules that can be easily obtained by traditional methods, multinational pharmaceutical companies value later clinical data more because the faster the speed and the more sufficient the clinical data, the greater the probability of approval for marketing. However, for molecules that are difficult to obtain, even in the early stage, as long as they can form differentiation, they can have high value; moreover, high-quality molecules can also improve the success rate in the clinical stage.
It is reported that BioGeometry has reached more than 20 BD cooperation agreements with domestic and foreign pharmaceutical companies and achieved breakthroughs in many aspects such as de novo design of highly specific antibodies, multi-objective optimization of lead molecules, and vaccine design.
Take the field of tumor immunology as an example. The GeoFlow model achieved the "de novo design" of highly specific antibodies. The target of this project is the antigen unique to the surface of tumor cells. The difficulty in R & D lies in the existence of a highly homologous "twin target" - the two structures are highly similar, and it is difficult for traditional methods to accurately distinguish them at the molecular level, easily causing damage to normal cells expressing the "twin target". By leveraging the full-atom modeling ability of GeoFlow, BioGeometry directly wrote "specificity" as a pre-constraint into the molecular generation stage: by designing no more than 100 sequences, it obtained 2 antibodies with both high selectivity and high affinity - accurately binding to the target while not binding to the "twin target", ensuring clinical safety from the source.
"Currently, this is one of the projects that best represents the model ability of BioGeometry. Usually, most of the projects jointly developed with customers are those that cannot be solved by traditional methods, and there are even some cases that GeoFlow V2 could not complete but were successfully broken through after the iteration of GeoFlow V3. In this process, not only can we feel the emergence of capabilities brought by model iteration, but customers themselves can also intuitively feel the rapid evolution of AI technology. Especially after the explosion of AI Agents, the speed of technological iteration in the fields of drug R & D and protein design has significantly increased," Tang Jian mentioned.
In cooperation with a well-known foreign pharmaceutical company, the target project needed to optimize multiple indicators of the lead antibody, such as affinity, physical and chemical properties, thermal stability, and human origin. In the zero-sample scenario (the model was not fine-tuned with the data of this target), GeoFlow delivered the target molecule that met all the preset indicators through only one round of design and verification: the affinity was increased dozens of times, the expression level was increased by 8 times, the human origin was optimized to over 90%, and the thermal stability was also significantly improved. The project delivery cycle was shortened by more than 80% compared with the customer's expectation.
In addition, in the field of synthetic biology, BioGeometry has reserved dozens of self-developed pipelines. Among them, several pipelines including α-ketoglutaric acid and natural borneol have completed pilot-scale expansion. Currently, many of the company's self-developed pipelines have reached licensing cooperation through the "technology transfer + sales share" model, and the commercialization progress is accelerating.
In terms of the team, BioGeometry was founded by Professor Tang Jian, an AI4S scientist, and Yoshua Bengio, the Turing Award winner and the father of AI, serves as the chief scientific advisor. The team has been exploring AI-driven drug discovery since 2018 and has achieved many results in industry-university-research cooperation. In 2021, it applied the diffusion generative model to the generation of molecular three-dimensional structures (representative works ConfGF, GeoDiff). In 2022, it jointly released the open-source machine learning drug discovery platforms TorchDrug and TorchProtein with NVIDIA, Intel, and IBM. Recently, as a core contributor, it participated in the R & D of NVIDIA's open-source large protein model La Proteina and independently developed the cutting-edge AI virtual cell model PerturbDiff.
Investors' views:
Guo Qiushan, the president of Shanghai Biomedical Innovation Transformation Fund, said that the development of macromolecular drugs has long been restricted by the long process of traditional screening and the cumulative errors of the previous step-by-step AI toolchain. BioGeometry has achieved atomic-level precision modeling of biomolecular interactions and integrated structure prediction, sequence generation, druggability evaluation, and wet experiment feedback into a closed loop, representing an AI Native path that is closer to the underlying scientific logic and real industrial needs. This full-atom de novo design concept allows the company to show unparalleled generational advantages in high-difficulty pipelines such as traditional undruggable targets, complex antibodies, and multi-specific macromolecules, and has achieved the delivery of PCC-level molecules. We look forward to BioGeometry fully promoting the clinical development of self-developed pipelines and global cooperation with its rapidly iterating and self-controllable GeoFlow algorithm base.
Zhang Kun, the leader of the intelligent medical group of Guoke Investment, said that AI-driven drug R & D is expected to break the R & D dilemma of the "anti-Moore's law" in the biomedical industry and promote macromolecular drugs to gradually establish a new R & D paradigm of "structural understanding, targeted design, and closed-loop verification of wet and dry experiments". We highly recognize the technical ability and global influence of the team led by Professor Tang Jian in the field of AI4S. The self-developed GeoFlow model has shown differentiated technical advantages in the scenarios of antibody drug and industrial enzyme de novo design. We believe that in the context of AI reshaping drug R & D, BioGeometry will deeply empower innovative pharmaceutical companies and the biomanufacturing industry, accelerate pipeline transformation and commercialization, and continuously bring industrial value and realize industrial dividends.
Dr. Wang Dakui, the managing director of Fortune Capital, said that the "intelligent emergence" moment of AI in the biomedical field came faster than the industry expected. AI can design and screen candidate molecules almost infinitely, and the innovation in drug R & D is shifting from traditional experimental trial and error to computation-driven. BioGeometry is a top AI4S team led by Professor Tang Jian, which has been deeply involved in biological computing for a long time. It not only has solid academic accumulation but also has the ability to transform cutting-edge algorithms into engineering solutions. The self-developed GeoFlow micro-world model can accurately predict the structures and interactions of biological macromolecules such as proteins, and its technical ability is in the global first echelon. It is also a key breakthrough to break the monopoly of overseas closed-source models. Coupled with the advantages of low cost and fast iteration in the wet experiment link in China, Chinese AI pharmaceutical companies represented by BioGeometry are fully capable of catching up.
Li Wenjue, a partner of Starlink Capital, said that life science is entering a new era: from relying on experience and accidental discoveries to precise innovation driven by computation and design. BioGeometry uses generative AI as the engine to explore the programmable design of proteins, the underlying language of life, and continuously accelerates model iteration and experimental verification through the closed loop of wet and dry experiments, improving the efficiency and success rate of new molecule discovery and functional design. We are optimistic about the systematic accumulation of BioGeometry in AI basic models, protein design ability, and experimental verification systems, as well as the long-term innovation potential shown by its global and interdisciplinary team. We look forward to BioGeometry continuously promoting the in-depth integration of AI and life science and opening up a more efficient, predictable, and engineerable new paradigm for biomedicine and synthetic biology.