Addressing System-Level Gaps in Single-Point Perturbation: How Can AI4S Virtual Organs Find Practical Application Scenarios? | Exclusive Interview with Li Lei, Founder of CANCHEN TECH

The low clinical translation rate of AI-driven drug development stems from the core issues of "single-point prediction" and "single-loop empowerment."

— Only the responses of targets and single cells are observed, while the cascade changes at the organ system level remain unseen. Dr. Li Lei, founder of CANCHEN TECH, stated, "Single-point perturbation predictions fail to capture system-level side effects, which is precisely why many drugs and AI-driven drug development efforts only reveal issues during clinical trials."

CANCHEN TECH is developing virtual organs, and will leverage system-level digital twin modeling and spatiotemporal dynamic causal inference to bridge this gap.

In 2025, CANCHEN TECH was registered and established in Shenzhen, growing into the only AI4S life sciences enterprise in China with full-scale spatiotemporal dynamic simulation capabilities spanning "molecule-cell-tissue-organ." Li Lei and his founding team are among the earliest research forces globally to conduct three-dimensional dynamic modeling of mammalian organs, having successively developed multiple pioneering models with global first-of-their-kind significance. Their series of breakthrough achievements have been published in top-tier international academic journals such as Cell.

Against this scientific research backdrop, CANCHEN TECH has assembled an interdisciplinary team of top-tier talent from prestigious domestic and international institutions — including Tsinghua University, Peking University, Stanford University, the University of Cambridge, and BGI — encompassing end-to-end practical expertise in industrial-scale large model deployment, clinical medicine, and regulatory translation for drug approval.

In 2026, CANCHEN TECH completed a series of seed, angel, Angel-Plus, and Angel-Plus-Plus financing rounds totaling tens of millions of RMB. Investors include Oriental Fortune Capital, Green Pine Capital Partners, Zero Venture Capital, Tsinghua Alumni Seed Fund, TusStar, Qifan Capital, and Nanshan Strategic Emerging Industries Investment.

In Li Lei's view, virtual organs are not merely another AI-driven drug discovery tool, but rather a correction and extension of the current "single-point targeting and black-box prediction" paradigm, enabling the stepwise realization of high-precision digital twin modeling for individual organs and dynamic, multi-system whole-body simulation.

A more realistic focus lies in standardized, replicable, and commercially viable application scenarios and solutions. CANCHEN TECH's system-level virtual organ simulation technology system has delivered multiple industry-first collaborations:

From single-point perturbations and cellular simulations to system-level virtual organs and full-scale spatiotemporal dynamic models, this is how CANCHEN TECH follows technological threads to identify practical implementation pathways. The interview content has been edited for our readers.

VCBeat: AI-driven drug discovery has seen years of heavy investment, so why have clinical outcomes failed to meet expectations? What key challenges do virtual organs primarily address?

Li Lei: Currently, the clinical translation rate of many AI-driven drug development pipelines remains low, and their actual efficiency falls short of expectations. The core reason lies in our inability to observe system-level changes and to model system-level relationships, such as cell-cell interactions and alterations in the tissue microenvironment. This is precisely what we aim to change.

Single-point perturbations can alter the dynamic evolutionary trajectory of the entire cellular network, causing all related factors and cell populations to evolve in directions that deviate from normal developmental pathways. However, most existing models rely solely on black-box, single-point predictions based on an "input perturbation-output result" framework, which means many adverse effects remain undetected until the clinical stage.

VCBeat: In the industry, it is easy to perceive virtual organs as an "upgraded alternative" to organoids, organ-on-a-chip systems, and virtual cells. How do you define the relationship between virtual organs and these technologies?

Li Lei: Biomimetic technology has actually emerged gradually across different timelines — initially limited to animal experiments and human clinical trials, and later advancing to organoids and organ-on-a-chip systems, where animal cells or small amounts of human stem cells are used to induce the formation of cell spheroids, tissues, and organs.

Virtual organs and virtual cells are not substitutes for one another; rather, they address different problems at distinct hierarchical levels. Many complex diseases do not stem from changes in a single cell, but rather from alterations in intercellular interactions, the tissue microenvironment, and organ-system dynamics. For instance, the progression of neurodegenerative diseases involves factors such as interactions among different types of neurons, mechanical stress within tissues, and interactions between different regions of the heart. Therefore, virtual organs aim to go further and address what will happen to the entire organ following interventions such as drug administration or gene editing.

From a broader perspective, with the advancement of single-cell and spatiotemporal omics technologies alongside AI, the construction of a digital human in silico is gradually becoming an engineerable problem.

CANCHEN TECH is undertaking a critical step: integrating cell dynamics, spatial relationships, and organ development with aging and disease progression to construct system-level virtual organs capable of temporal simulation. In the future, these models can serve either as standalone computational frameworks or be combined with experimental systems such as organoids and organ-on-a-chip platforms, thereby becoming digital infrastructure for drug discovery and life sciences research.

VCBeat: Molecule-Cell-Tissue-Organ: How to Achieve Higher-Dimensional Modeling of "System-Level Changes"?

Li Lei: We believe that the mechanism of life evolution is the result of interactions among numerous cells. At the tissue level, intercellular interactions constitute the microstructure of tissues, including gene co-expression patterns and cell interaction networks resulting from these interactions. At the organ scale, an organ is a morphological and functional unit composed of multiple tissues organized in a specific structure, interconnected and working synergistically to perform specific functions. In real-world scenarios, the organ scale is primarily assessed through phenotypic functional readouts, such as physiological parameters or key functional indicators during the progression of disease.

CANCHEN TECH's "full-scale" approach aims to establish a dynamic modeling framework that progressively propagates from microscopic mechanisms to macroscopic phenotypes. The key challenge lies in building cross-scale associative mappings from molecules to cells, tissues, and organs. For instance, can changes at the most fundamental level of molecular structure, such as upward or downward trends in transcription factor expression, be correlated with specific functional phenotypic readouts at the organ level, thereby enabling the assessment of disease progression and drug response?

A complete system for virtual organs would encompass the representation and modeling of virtual cells. We learn feature vectors for cells (e.g., based on cell type, gene expression, and other information) in a manner analogous to text embeddings. Building on this foundation, we incorporate two additional dimensions:

First, the cell interaction network brought about by spatial location. A hierarchical graph structure is used for modeling, and then cellular and tissue-level features are integrated as multimodal information into the interaction network. Cells in close spatial proximity are more likely to exhibit stronger interactions; therefore, we focus on spatial transcriptomics sequencing samples. Spatial transcriptomics enables the simultaneous detection of transcriptomic expression profiles of multiple cells (or cell spots) within a specific region of a tissue section, while preserving spatial location information. Based on spatial distribution, the model can construct a cell co-expression network to characterize tissue-scale microenvironmental readouts, including the causal relationships underlying interactions within the network.

Second, the dynamic spatiotemporal evolution model. Cells are not static; by incorporating temporal features, the entire graph structure transforms into a temporal dynamic graph, outputting the trends of information changes across various scales for cells at subsequent continuous time points — this is also the underlying driving logic behind our organ-level dynamic inference.

With the enhancement of spatiotemporal resolution, the internal structures and mechanisms of models will become clearer; black-box models will be transformed into interpretable models, thereby supporting research and development decision-making and aligning with a broader range of application scenarios and needs, such as the selection of timing for clinical interventions and personalized medication regimens.

VCBeat: In the realm of perturbation prediction, how does CANCHEN TECH's technological paradigm enhance interpretability?

Li Lei: Interpretability stems from the spatiotemporal dynamic evolution itself — we can observe how the system evolves step by step.

Current virtual models typically frame tasks as classification problems: given a drug perturbation, the model outputs whether there is efficacy or toxicity. Our approach is causality-driven — injecting the life mechanisms of organ evolution into the model as priors, rather than immediately relying on Transformers to scale up parameters and fit scaling laws.

In terms of specific implementation, a hierarchical graph structure is used for modeling: cells serve as nodes, and edges are formed between cells with stronger interactions due to their spatial proximity. Multimodal features from the molecular layer, transcription factor layer, and tissue layer are then superimposed. Finally, the temporal dimension is incorporated to construct a temporal dynamic graph, which outputs cell states, gene expression profiles, and spatial distributions across continuous time points.

For example, a black-box model may indicate that "this molecule could cause liver injury," whereas a virtual organ model aims to answer questions such as "which pathways are altered in which cells, how the effects spread through tissue via interactions, and at what time point inflammation is triggered." The latter provides the information pharmaceutical companies need for decision-making.

VCBeat: Does the continuous dynamic output of a high spatiotemporal, full-network system entail greater complexity in underlying algorithm and platform development?

Li Lei: The approach is more complex, yet it represents the natural culmination of the team's years of accumulated scientific research. Between 2022 and 2023, the team began conducting time-series and spatiotemporal modeling across multiple scales, including molecular and cellular levels, positioning them as one of the first groups globally to transition from single-scale to system-level modeling. Prior to this, they had completed foundational model research at the molecular and cellular layers, initiated organ-level modeling at an early stage, and pioneered the construction of four-dimensional spatiotemporal atlases of mammalian organs, such as those in mice.

Most of the team's self-developed virtual organ models have already been validated through wet-lab experiments, with some works currently under review at top-tier conferences or journals. These efforts represent some of the earliest globally recognized, truly AI-driven organ-level virtual simulation models. Upon introducing perturbations, the models can continuously extrapolate outcomes, predicting the trajectories of cell division, differentiation, and proliferation from early-stage cells to progressively form complete tissue structures. They can also perform forward simulations step-by-step over time, with wet-lab validation conducted at each key node.

VCBeat: With improved interpretability, will the application scenarios of spatiotemporal dynamic models expand to areas such as disease mechanisms and target discovery?

Li Lei: CANCHEN TECH's positioning and original aspiration is to build foundational models in the field of life sciences. Similar to OpenAI's early training of GPT-1 and GPT-2, where semantic emergence was achieved through massive datasets, the scaling law has indeed proven to be a highly successful discovery.

However, life sciences are far more complex than text and images. While scaling laws may yield effective fitting and deliver short-term results, a profound understanding of expert systems is essential to truly comprehend the mechanisms of life. Therefore, guided by a reverence for life, we design innovative and rational AI architectures grounded in the fundamental knowledge and mechanisms of life sciences.

VCBeat: Taking drug perturbation scenarios as an example, what are the differences in value provided by virtual organs and virtual cells?

Li Lei: In an idealized scenario, a virtual cell involves training a classifier tailored to different drug-related scenarios, leveraging the existing drugs and data available to pharmaceutical companies. When serving a specific pharmaceutical company, the model can achieve excellent fit by utilizing its proprietary, targeted data. However, when encountering new drugs or novel scenarios, the generalization capability of non-customized models is often limited, typically necessitating fine-tuning.

However, the concept of virtual organs posits that a single organ model can address a wide range of issues related to drug administration and disease research and development. Any drug's interaction with the model elicits response mechanisms consistent with those observed in the human body. In theory, once the model is established, it can be reused across different drugs and scenarios. This is the fundamental difference between interpretability and generalization capability.

VCBeat: To build a system-level model with high generalizability, does the data foundation need to be richer or broader in scope?

Li Lei: Correct. CANCHEN TECH's original intent is to design model architectures that align with the mechanisms of life. While Transformer will be one of the core components, the key lies in the modeling mechanism itself. System-level modeling requires data in the form of multi-scale, multi-omics, and multimodal information. Although highly generalized models indeed require a richer data foundation, the global volume of human data is limited and unevenly distributed in quality; therefore, it is currently imperative to leverage cross-species data.

VCBeat: How should we understand the use of cross-species data in simulating human organs?

Li Lei: Historically, the research paradigm in life sciences has primarily relied on animal models, as the ethical barriers to obtaining animal data are lower than those for human subjects. In contexts such as development, aging, and certain diseases, humans and animals share significant similarities, including a large number of homologous genes, which help fill gaps where high-quality human data are difficult to obtain or lacking.

But this window of opportunity is limited. The cost of animal experiments is rising; the price of a single experimental monkey has climbed to match that of a compact car. Moreover, sequencing itself destroys the biological specimen, making it difficult to reuse samples across different scenarios. Meanwhile, ethical constraints on animal experimentation are strengthening globally. In the future, obtaining large volumes of high-quality animal data may become increasingly challenging. Whoever first achieves accurate, stable, and generalizable model performance in a specific scenario, while transforming black-box modeling into a highly interpretable approach, will gain the competitive edge.

VCBeat: In the data input stage of the model, what special processing does CANCHEN TECH employ?

Li Lei: Data in the life sciences sector lacks the clearly defined tasks and ideal standards seen in computer vision (CV) and natural language processing. The data is massive in scale, noisy, and sparse, while technical standards for sequencing equipment vary across countries and platforms. There is no industry-wide alignment in the design pipeline from raw data formats to model development.

The company's early technical team published one of the first domestic studies on virtual cells at NeurIPS last year, proposing a "grid learning" strategy — graph structures cannot be directly treated as sequences and fed into Transformers. For instance, single-cell transcriptomic data is essentially a table, where each row and column reflect cell relationships and gene co-expression relationships, respectively, and can be normalized for tabular input into models. The Stack model, the world's largest single-cell foundation model released by Silicon Valley's Arc Institute in the first half of this year, also adopts this tabular learning strategy.

The core logic is: first, collect data as comprehensively as possible; on this basis, truly understand the data, allow it to enter the model in a more optimal manner, and enable the model to fully unlock the value of the data.

VCBeat: To validate the system-level value of the model, why start with virtual liver and virtual brain?

Li Lei: System-level gaps in high-demand, high-complexity scenarios represent a major pain point for the pharmaceutical and life sciences research industries.

The virtual brain boasts the highest level of complexity and serves as CANCHEN TECH's first-mover advantage. Few teams worldwide are engaged in virtual brain development, given that the brain involves multimodal neural information and encompasses the largest cellular volume. Meanwhile, neurodegenerative diseases represent extreme examples of system-level problems — Alzheimer's disease involves multiple brain regions, diverse cell types, and long-term evolutionary processes, rendering single-target approaches consistently ineffective.

We have progressed from mouse and monkey brains to the human brain. The second version of our brain model is expected to be released between late this year and early next year. It will cover dynamic evolutionary predictions for development, aging, and diseases such as Alzheimer's and Parkinson's. Furthermore, we will collaborate with pharmaceutical companies and key hospitals to conduct drug efficacy simulations and optimize combination treatment regimens.

Virtual liver models represent a faster route to commercialization — hepatotoxicity is a critical unmet need for pharmaceutical companies. Drug-induced liver injury is a major cause of pipeline failures, and traditional animal models have limited predictive accuracy. Moreover, the liver is not a single-target issue; it involves systemic responses encompassing metabolic networks, cell-cell interactions, and inflammatory pathways.

VCBeat: What level of precision has the current dynamic evolution achieved?

Li Lei: We currently do not simply define the capabilities of virtual organs using a single "accuracy" metric. Taking the brain as an example, we have achieved relatively clear results in several key areas: spatial information reconstruction, cell identity understanding, and developmental dynamics inference.

What we actually value more is the biological significance behind these numbers. For example, in the neural crest development task, the model not only predicts the overall distribution of cell populations but also recapitulates authentic developmental trends, such as the late-stage increase in sympathetic cells and the decrease in dorsal root ganglion cells.

This signifies that the model is evolving from merely "describing the current state of an organ" to "predicting how this organ will change in the future."

VCBeat: Can the developmental and aging processes of specific organs be applied to the field of anti-aging?

Li Lei: Essentially, modeling the spatiotemporal dynamic evolution of organs hinges on modeling the developmental trajectories of individual cells. Theoretically, controlling the progression of cellular aging and evolution could involve simulating key time points to determine whether single-point or combinatorial interventions can induce cell types from relatively mid-to-late stages back to an earlier, immature state. This is a highly promising avenue for exploration.

Our current virtual liver research already encompasses aging and organ regeneration. On one hand, we are conducting drug simulations; on the other, we are advancing exploration into liver aging.

VCBeat: How did you conceive the commercialization model for CANCHEN TECH?

Li Lei: We aim for customers to purchase not merely software licenses or one-off reports, but a digital life system centered on specific research and development challenges and capable of sustainable iteration: in the short term, it serves as a tool-based service; in the medium term, it evolves into standardized organ models; and in the long term, it enables multi-organ synergistic human digital twins.

From the perspective of commercialization pathways, we will currently prioritize entry into pharmaceutical companies and life sciences research and development scenarios, such as research on drug mechanisms of action, evaluation of candidate drugs, simulation of disease progression, and prediction of intervention responses. In the future, the field will gradually evolve from individual projects to standardized organ models and platform-based products, further developing foundational capabilities that are reusable across multiple organs and diverse disease scenarios.

VCBeat: To what stage has the current collaboration progressed?

Li Lei: The ecosystem integrating life sciences and AI is still in its early stages — unlike embodied intelligence, which has "material factories," current data acquisition strategies have yet to be validated or standardized. This requires obtaining data, running initial model iterations, and continuously optimizing through accumulated experience and clinical data.

Our strategy is to first establish virtual organs as a foundational capability, achieve thorough validation in scenarios such as drug development, and then gradually expand into clinical settings. Large-scale validation will be conducted in the third phase.

We are currently engaging in multifaceted collaborations and scenario validations with pharmaceutical companies, research institutions, and clinical teams. Our aim is for every business case to ultimately consolidate into model capabilities and data assets, rather than simply delivering one-off projects.

VCBeat: The National Medical Products Administration's NAMs Pioneer Program places AI simulation, organoids, and organ-on-a-chip technologies on equal footing within the experimental validation cycle. What does this mean for CANCHEN TECH?

Li Lei: First, it is an affirmation — virtual organs are not merely research tools, but can be verified, evaluated, and incorporated into research and development decision-making, enabling genuine entry into the drug research and development evaluation system.

We also maintain ongoing communication with regulators, industry stakeholders, and experts, focusing on discussions regarding the scenarios, standards, and validation processes required for virtual organs to serve as valid evidence in drug development.

VCBeat: Will AI ultimately replace animal testing?

Li Lei: From the current perspective, this represents a long-term trend; however, we do not believe that AI models will replace all animal experiments in the future. A more likely scenario is that AI-based virtual organs, organoids, organ-on-a-chip systems, and other human-relevant experimental platforms will collectively constitute a new evidence framework for New Approach Methodologies (NAMs), thereby making drug development increasingly reliant on models that more closely mimic human physiology. The FDA also emphasizes that different methods must demonstrate their scientific validity and applicability within specific application contexts.

VCBeat: What AI story does CANCHEN TECH aim to tell?

Li Lei: CANCHEN TECH aims to build not just an AI-CRO focused primarily on project delivery, but a life science technology platform with virtual organs as its core infrastructure and computational experiments as its core service model, making software and intelligent models the new infrastructure driving life sciences research and development.

Life sciences are undergoing a major transformation: in the past, our understanding of the human body relied more on experiments, animal models, and limited human data; in the future, we aim to build digital life systems capable of continuous learning and dynamic simulation, grounded in real-world data and experimental evidence.

The digital human twin is precisely the AI and life sciences story that CANCHEN TECH aims to tell.