With its self-developed basic model for biological structure prediction, "Tanxu Zhiyuan" aims to break the "10-year, $1 billion" rule in new drug R & D | Early-stage project

For a long time, the paradigm of scientific research has been based on mathematical principles and experimental observations. From Newton's "Philosophiæ Naturalis Principia Mathematica" to Einstein's mass - energy equation, scientists have used formulas, laws, experiments, and precise calculations to understand and predict nature.

Now, with the development of generative artificial intelligence, especially the remarkable achievements of Transformer and Diffusion models in fields such as text and multi - modality, "generative science" may change the previous scientific research paradigm. That is, it no longer adheres to the precise mathematical description and experimental verification of each intermediate step. Instead, it relies on massive scientific data (such as gene sequences and structures) to train foundation models and obtain the ability to directly generate results, achieving better effects of "relatively accurate, absolutely fast, and absolutely broad".

The real beginning of this "generative science" was the birth of AlphaFold 2, which revolutionarily solved the problem of protein "from sequence to structure" prediction. In 2023, AlphaFold 3 was released, expanding its capabilities to the interactions between proteins and complex biomolecules such as nucleic acids, small molecules, and antibodies, giving AF3 the potential to guide drug R & D.

At the beginning of 2024, Isomorphic Labs, a subsidiary of Deepmind responsible for the industrialization of AF3, received huge orders from two MNCs, Eli Lilly and Novartis, to jointly develop multi - target small - molecule therapies. The upfront payments were as high as $45 million and $37.5 million respectively. It is worth mentioning that at the beginning of 2025, Novartis announced an expansion of its cooperation with Isomorphic Labs, further increasing the number of collaborative research projects. This also means that Novartis recognizes its ability to explore drugs for "undisclosed targets".

36Kr learned that a new startup, Tanxu Zhiyuan (hereinafter referred to as Tanxu), founded in the second half of 2024, recently released its self - developed foundation model, IntelliFold, in the wave of generative science. Currently, its public server is open.

The founder, Sun Peng, is a former tech venture investor. His career started in Accenture management consulting and then he worked in VC institutions such as Mingshi Capital. He has been engaged in frontier technology investment for many years and has rich experience in AI investment and industrial management. The chief scientist, Sun Siqi, is a researcher and doctoral supervisor at Fudan University. During his doctoral studies at TTI - Chicago, University of Chicago, he studied under Professor Xu Jinbo. From 2018 to 2022, Sun Siqi worked at the Microsoft Seattle headquarters, engaging in cutting - edge large - language model technology research. After returning to China, he has been committed to the innovative application of AI in interdisciplinary fields, achieving systematic frontier breakthroughs around the precision and efficiency bottlenecks in the field of intelligent computing in structural biology. His achievements have been published in top journals such as Science and Nature sub - journals, with a total citation of over 7,000 times.

Sun Peng introduced to 36Kr that most of the scientific research team members of Tanxu Zhiyuan have a dual R & D background in "structural biology + large - language models", so they have the ability to independently develop a foundation model for structure prediction.

"A few years ago, the value of AI in the binding prediction and design tasks faced by structural biology was mainly to improve efficiency. But now, new AI is needed to solve problems that are difficult to solve with traditional technologies, such as efficiently exploring undisclosed targets or designing completely new products that medicinal chemists have never touched and even beyond human intuition. In this process, large AI models are indispensable. However, what we do does not consume as much and bottomless computing power as training general large - language models, and the path for industrial value transformation is shorter and clearer. Using generative science models to directly participate in scientific research exploration is one of the three primary entry points for AGI to develop into intelligent productivity."

It is understood that IntelliFold is positioned as a "controllable foundation model". This means that on the one hand, it can perform high - precision three - dimensional structure prediction of the interactions between various biomolecules (proteins, nucleic acids, small molecules, ions, modified residues, etc.). At the same time, through the application of lightweight trainable adapters, it can be guided and controlled to achieve specific targeting capabilities such as allosteric prediction and binding prediction for given pockets, helping to complete complex tasks required for specific downstream applications such as drug discovery.

Image source: "IntelliFold: A Controllable Foundation Model for General and Specialized Biomolecular Structure Prediction"

According to the technical test report provided by the company, IntelliFold performs similarly to AF3 in several key indicators of protein structure prediction tests, such as protein monomer structure prediction, protein - protein interface prediction, and protein - DNA/RNA interface prediction. In terms of antibody - antigen interface and protein - ligand interaction prediction, its success rate is slightly lower than that of AF3. Notably, in RNA monomer prediction, it even surpasses AF3, demonstrating its advantage in nucleic acid structure prediction.

"Given a specific protein sequence, the IntelliFold model can predict its binding conformation and mode with small molecules. This is one of the current characteristics of our technology and also a direction with clear market demand," Sun Peng said. "In addition to predicting specific binding modes such as allostery, the IntelliFold model can also predict Affinity (binding affinity, one of the core indicators for measuring drug efficacy) values, enhancing the efficiency and accuracy of virtual drug screening."

In drug design, proteins may undergo conformational changes according to the bound molecules, forming different functional states. For example, the activity of CDK2 in the kinase family may be affected by allosteric changes induced by inhibitors. This is crucial for drug design but difficult to effectively train in large models. However, through target - specific adapters, IntelliFold can correctly predict allosteric conformations, "identifying rare conformational states without affecting the accuracy of the model in the orthosteric state." This is quite important for the flexibility of precise drug design.

At the same time, Sun Peng also mentioned that generative science models are also bringing rapid changes to protein design. Different from the classic expert - led drug design paradigm, generative models can completely design de novo the position and possibility of each amino acid, and even explore results that do not exist in nature but may be better. "Although the difficult parts of de novo protein design are different from binding prediction, the foundation models used by the two are similar in origin and have the ability to expand horizontally. Undoubtedly, the ability of the foundation model is the key prerequisite for obtaining leadership in specific scenarios and industrial usability in the future."

In Tanxu Zhiyuan's future plan, it hopes to build IntelliFold into a general intelligent scientific foundation model, playing an engine role in different specific tasks and improving the R & D efficiency of the entire industry. Next, Tanxu will commercialize through various means such as joint development with large pharmaceutical companies and providing valuable early - stage assets for pharmaceutical companies/research institutions. It hopes that through the application and continuous upgrading of IntelliFold, the success rate of early - stage new drug R & D can be systematically improved, thus changing the dilemma of "10 - year cycle, $1 billion, 10% success rate" in new drug R & D.

"Designing proteins and drugs like designing chips through AI." This concept of Jensen Huang has been widely accepted in the United States. With the application of AI, the value of pre - clinical and phase - I clinical drug assets is currently being re - evaluated. Due to the intervention of AI, the probability of these early - stage assets becoming drugs has greatly increased. Vas Narasimhan, the global CEO of Novartis, also said that he hopes to see new technologies such as AI increase the success rate of drug R & D from the current 1/10 to 2/10 or even 3/10.