Using AI to "replicate" human cells and predict drug efficacy, Huayuan Gene Secures Tens of Millions of RMB in Seed Round Financing | 36Kr Exclusive

By Hu Xiangyun

Edited by Hai Ruojing

36Kr learned that AIVC (AI Virtual Cell) enterprise Huayuan SmartGene has recently completed a seed round financing of tens of millions of RMB. This round of financing is led by Tsinghua University Shuimu Venture Capital. The raised funds will be mainly used for the iteration of underlying multi-modal sequencing technologies, further expansion of cooperation with top-tier Class A tertiary hospitals, and team expansion. In addition, the Huayuan SmartGene team has already planned to launch a new round of financing.

The founding team of Huayuan SmartGene is composed of senior pharmaceutical industry practitioners and computational biology R&D personnel. It has invited experts from institutions including the Shenzhen National Gene Bank to form a scientific advisory board, creating a team structure that integrates scientific research and industrial collaboration. The company's CEO Du Runshi studied at the University of California, Los Angeles, and is a serial entrepreneur in the AI field; CTO Wang Yixuan is a doctoral student in the Department of Computer Science and Engineering at the Chinese University of Hong Kong, who previously led the development of the xTrimoSC Perturb prediction model for the transcriptional effects of gene perturbations at BioMap. In addition, Li Yu, an assistant professor in the Department of Computer Science and Engineering at the Chinese University of Hong Kong, serves as the Chief AI Scientist of Huayuan SmartGene.

In the field of new drug R&D, the "10 years, 1 billion US dollars" dilemma is widely recognized. Among all resources invested, 90% are allocated to the human clinical trial stage, but only about 10% of candidate drugs can successfully pass all clinical trials and get approved for market launch. One of the core reasons for this extremely high failure rate is the huge difference between the physiological environment in animal experiments and the real human physiological environment.

So, is it possible to simulate the real human response to a drug on a computer before investing huge costs in human clinical trials, to reduce the failure rate of new drug development?

Du Runshi believes that most traditional AI pharmaceutical companies focus on solving the front-end problems of R&D, that is, using AI to more efficiently identify potential disease-causing targets and generate compounds. However, this is only optimizing the first step of "drug discovery", which cannot fully predict whether the drug will be effective or have side effects when applied to the human body after it becomes a finished product. Determining "whether a certain drug is worth advancing to human trials" is the key to solving the problem.

This is also the fundamental reason why Huayuan SmartGene chose to independently develop the Wise-Perturb cellular drug perturbation application model, focusing on human drug efficacy prediction.

In simple terms, human cells contain approximately 20,000 protein-coding genes. The healthy state of cells, disease progression, and drug response are all regulated by the coordinated action of multiple genes. The "cell perturbation" in the industry refers to artificially applying external interventions to cells, which is mainly divided into two major research directions: the first is gene editing perturbation, which knocks out or upregulates specific genes to observe the linked expression changes of downstream genes, so as to quickly locate disease-causing targets; the second is drug molecular perturbation, which inputs candidate drugs into a virtual cell model, and uses AI to fully deduce the dynamic changes of genes, proteins, and signaling pathways in cells under the action of drugs, to evaluate the human efficacy and potential toxicity of drugs. This is also the core R&D focus of Huayuan SmartGene at this stage.

According to the introduction, Wise-Perturb has two main breakthroughs. First, it specifically addresses the shortcomings of traditional models that rely on single RNA sequencing and are disconnected from the pathological characteristics of real patients. The training data for traditional AI cell models mostly comes from artificially modified immortal cell lines in laboratories. These cells are separated from the original pathological microenvironment of the human body and lack the individual genetic characteristics of patients, which easily leads to significant deviations when used to predict human drug efficacy; at the same time, traditional sequencing methods cannot simultaneously obtain three layers of key omics information of DNA, RNA, and proteins, making it difficult to fully restore the real regulatory logic of cells.

To this end, Huayuan SmartGene has built a single-cell multi-omics integrated analysis system, which captures cellular DNA, transcriptome, and proteome data through single-cell detection of the same cell, and then uses self-developed multi-modal fusion algorithms to connect the three layers of omics information, building a three-layer pyramid data base. The bottom layer is a massive general static single-cell sequencing data set, the middle layer contains hundreds of millions of paired experimental data of drug and gene perturbations in in vitro cell lines and PDX models, and the top layer is scarce paired sequencing data before and after drug administration in human clinical tumor cohorts.

"We have reached an innovative cooperation with top-tier Class A tertiary hospitals to build a joint laboratory. In the future, we also plan to target the superior departments of each hospital, collect exclusive data corresponding to specific indications, and train expert models for subdivided diseases. We hope to achieve in-depth cooperation with at least 30 top Class A tertiary hospitals in the next 1-3 years." Du Runshi said.

In addition, another feature of Wise-Perturb is the cell-specific design of the model. Ordinary AI models usually assume that all cells respond to drugs in the same way, but in reality, there are differences in the physiological mechanisms of different cells. For example, a drug that works in the lungs may damage the liver. Therefore, the Huayuan SmartGene team has specifically added a cell type recognition architecture to Wise-Perturb, enabling it to have zero-shot generalization prediction capabilities across different cells, cancer types, and new drugs. Without large-scale retraining for each disease, it can predict the human effects of new diseases or drugs using only a small amount of known patient data.

Huayuan SmartGene provided two real verification cases, one of which is the cross-indication verification of the star ADC drug DS-8201. Du Runshi introduced: "We used clinical data from breast cancer patients to complete the basic training of the model. Without training samples from ovarian cancer patients, we completed the prediction of drug response in ovarian cancer PDX models. The model scores matched the real drug efficacy in animals; compared with the general virtual cell model, the prediction consistency was improved."

The other case is a drug efficacy stratification study of the targeted non-small cell lung cancer drug osimertinib, conducted in cooperation with the Cancer Hospital of the Chinese Academy of Medical Sciences. "The hospital provided the baseline tumor sequencing data of patients before treatment, without releasing any clinical follow-up outcomes. The model independently output the patient's drug efficacy stratification results, which were consistent with the long-term clinically observed treatment effects." Du Runshi said.

From the perspective of Huayuan SmartGene, this prediction capability verified by clinical samples allows the company to break away from the traditional industry path of "investing huge R&D funds first and then expanding customers later". In the early stage of its establishment, it can already connect with medical institutions and pharmaceutical companies to carry out real paid cooperation.

Du Runshi introduced that at present, Huayuan SmartGene has "received cooperation orders from many top Class A tertiary hospitals, multinational pharmaceutical companies, and AI biotechnology enterprises, and the service fees for multiple projects have been fully collected".

At this stage, Huayuan SmartGene has two core commercial services. The first is to provide pharmaceutical companies with customized services for preclinical pipeline value assessment and stratified patient selection for clinical trials, charging basic service fees or adopting other revenue models based on project evaluation; the second is to build joint wet-lab and dry-lab laboratories with Class A tertiary hospitals to jointly carry out scientific research projects such as repurposing old drugs for new uses and screening new disease targets. In the long-term plan, the company also plans to launch a joint R&D model that shares risks with pharmaceutical companies and distributes revenue after drugs are launched on the market.

"We prioritize the implementation of small-scale verification projects, use reproducible and highly consistent prediction data to build customer trust, and then continue to deepen long-term cooperation. At present, human precision drug efficacy prediction is a rigid demand in the industry, and there is a lack of mature alternative technical solutions. Medical institutions and pharmaceutical companies are willing to open clinical sample data to carry out joint research." Du Runshi concluded.