Exclusive Interview with Pei Hao, Founder of MobiDrop: Letting Real Cells Teach Virtual Cells

"What I cannot create, I do not understand." In February 1988, while cleaning up Richard Feynman's office, people discovered this unfinished chalk writing on his blackboard. The Nobel laureate in Physics spent his life embodying this creed: the ultimate sign of understanding a system is not merely observing or describing it, but being able to recreate it from scratch.

More than three decades later, as life sciences stand on the threshold of the AI revolution, a group of entrepreneurs is rewriting this maxim in a new way: If humans truly understand a cell, they should be able to create a virtual cell in the digital world that breathes, responds, and metabolizes.

Currently, AI virtual cells have become one of the hottest tracks in global technology investment, widely recognized by the industry as the third paradigm in life sciences, following large language models and general world models. However, beneath this boom, the industry is deeply mired in algorithmic involution and structural imbalances. MobiDrop, based in Zhejiang, China, has not followed the trend of engaging in model competitions. Instead, it started with microfluidic chips and single-cell sequencing hardware, gradually building an industrialized data factory and a closed-loop system integrating dry and wet labs.

In July 2026, Dr. Pei Hao, founder of MobiDrop, offered a fundamental insight during an in-depth discussion: "On the surface, virtual cell technology appears to be a competition among models, but at its core, it is truly a contest of data and experimental closed-loop systems." This statement serves as the key to understanding the company's trajectory over the past eight years and the next eight.

Pei Hao's research origins had nothing to do with biology.

After earning his bachelor's degree in Mechanical Engineering and Automation from Tsinghua University, he pursued his PhD in Applied Physics at Harvard University under the supervision of Professor David Weitz, a founding figure in microfluidics and member of three U.S. national academies. Guided by the instincts of a mechanical engineer, he felt confused and uncomfortable upon his first visit to a biology lab: "Every solution still needs manual preparation. Why can't automation handle such repetitive work?"

That was over a decade ago, when biological experiments still had a strong "manual" feel, with pipettes, petri dishes, and microscopes forming a world that relied heavily on experience and physical effort. Pei Hao recalled, "I had previously been working in the field of mechanical engineering and automation, accustomed to standardized and scalable engineering logic. The biggest shock when I first transitioned into biological experimental research was the rough-and-ready nature of traditional biological scientific research."

The stark contrast between intelligent engineering and the extensive model of biological research guided Pei Hao to establish his consistent direction for research and entrepreneurship: to reconstruct the traditional life science experimental system with engineering, automation and standardization thinking, and transform non-standard, hard-to-replicate biological experiments into a modern engineering system featuring precise controllability and scalable implementation.

However, the reality he faced after returning to China to start a business dealt him a severe blow. "We experienced a situation where the supply of foreign reagents was cut off, leaving us unable to produce for three months. The lead time for imported instruments often stretches to several months, accompanied by high costs, prolonged cycles, inconsistent laboratory conditions, and numerous other restrictions."

This experience reinforced his understanding: The localization of medical devices is merely the first step; what truly matters is achieving self-sufficiency in the generation of comprehensive life data.

This has also become the first underlying thread in MobiDrop's development trajectory. From microfluidic chips to instrument design, and further to data factories and virtual cells, the company has consistently been strengthening the scarcest foundational capability in AI-driven R&D—namely, the ability to produce authentic, stable, and scalable data.

In 2026, the AI virtual cell sector is in a typical phase of concept hype.

Ginkgo Bioworks, a leader in synthetic biology in the United States, has launched the Virtual Cell Pharmacology Initiative (VCPI), focusing on open-source public datasets and general-purpose model frameworks. California-based tech innovators Tahoe and Xaira have successively unveiled virtual cell models with billions of parameters, leveraging strong capital support and computational power to lead in benchmarking competitions. Top research institutions such as the Chan Zuckerberg Initiative (CZI) and Arc Institute are also continuously investing in general-purpose inference models, with a focus on enhancing algorithmic reasoning capabilities.

Capital flooded into the sector, algorithm teams crossed boundaries to enter the market, and enthusiasm for this track kept rising.

Yet Pei Hao maintains a clear-headed distance from this frenzy. In his view, beneath the veneer of prosperity, mainstream overseas players generally suffer from inherent path limitations—heavily relying on existing public datasets and third-party open-access sequencing data to train models, and predominantly adhering to an "algorithm-first, data-second" R&D logic.

Although the global repository of single-cell data is claimed to exceed 300 million entries, it is plagued by severe batch effects and technical noise. Data generated by different laboratories and reagent manufacturers lack uniform standards, resembling mutually unintelligible "dialects." Consequently, AI models must consume vast numbers of parameters to perform data alignment and error cancellation, rather than genuinely learning the underlying biological principles.

The more critical industry pain point is the lack of causality: over 90% of existing legacy data consists of static observational data, merely capturing snapshots of natural cell growth, with a severe scarcity of perturbation feedback data following gene knockouts or drug interventions. Consequently, models can only learn superficial correlations, failing to capture the true causal laws governing biological processes.

This judgment has been indirectly corroborated by industry research.

A paper released by Microsoft evaluated the performance of multiple complex models versus linear models in gene perturbation prediction tasks, revealing that complex models did not significantly outperform simple linear models; The ARK Virtual Cell Challenge also publicly acknowledged that most models have yet to consistently surpass naive baselines.

In Pei Hao's view, this article, which "pours cold water" on the concept of virtual cells, serves as a wake-up call for the industry: true virtual cells must be measurable, interactive, and practically implementable, requiring more authentic data better aligned with AI training logic as support.

Having seen through the underlying limitations of general-purpose models, MobiDrop has chosen an industry-native path that is distinctly different from the mainstream approach: Starting from the fundamentals of wet-lab biology, we reverse-engineer competitive barriers based on data-native logic.

As a manufacturer of self-developed equipment, MobiDrop naturally has access to the full spectrum of experimental parameters spanning the entire workflow from sample processing to data output. This includes key variables such as temperature, time, and inducer concentration, as well as metadata like quality scores, integrity metrics, and batch effects for each data point. Such process-generated data is scarce and prohibitively expensive to reproduce, providing AI models with more comprehensive and biologically realistic training dimensions, thereby constituting the company's primary competitive moat.

But Pei Hao believes this is far from enough: "We can not only produce data, but also change the way data is generated." Leveraging a self-developed, controllable high-throughput experimental system, the team can reverse-engineer perturbation experiments based on model capability gaps, generating on-demand dynamic micro-perturbation data—such as gene regulation, drug intervention, and environmental perturbations—rather than being constrained by the limitations of existing public datasets.

This precise data generation capability, characterized by "producing exactly the experimental data that the model lacks," directly addresses the long-standing industry challenge of causal deficiency. MobiDrop has thereby established a complete closed-loop system encompassing "data definition–data design–data production–data validation–model iteration." By continuously accumulating massive amounts of multimodal causal perturbation data, MobiDrop ensures that its virtual cell models are firmly grounded in real biological logic that is measurable, interactive, and verifiable.

If data is the first principle of MobiDrop, then the dry-wet closed loop is the engineering expression of this principle.

In Pei Hao's vision, dry lab and wet lab experiments form an inseparable, continuous iterative cycle: AI generates scientific hypotheses, automated experiments perform validation, real-world data feeds back to optimize the model, and the model then outputs new hypotheses. With each completed cycle, the model's predictive accuracy advances further.

The physical carrier of this closed loop is MobiNova®, a self-developed hardware matrix by MobiDrop. "By converting experimental variables into controllable engineering parameters and leveraging AI for Science technology, full-process data traceability and quality control can be realized." This means Mobio Biotech can not only identify model errors on the algorithm side, but also trace the sources of data noise from wet-lab experiments.

Building on this foundation, MobiDrop has independently developed the MobiBrain dry-wet closed-loop intelligent system, establishing a standardized "Design-Build-Test-Learn" (DBTL) cycle. This system thoroughly breaks down the inherent barriers separating dry-lab algorithms from wet-lab experiments, enabling bidirectional empowerment and iterative evolution between AI virtual models and real-world biological experiments.

It is evident that pure algorithm-driven companies possess only dry-end inference capabilities, while traditional biotechnology firms have only wet-lab experimental capabilities. MobiDrop, however, is one of the few companies in the industry equipped with comprehensive capabilities on both ends.

Leveraging its integrated dry-wet closed-loop system, MobiDrop's virtual cell model precisely addresses the four core challenges of the industry: Can it accurately predict unknown cellular perturbations, possess cross-sample and cross-disease generalization capabilities, authentically reproduce experimental deduction results, and effectively reduce R&D costs while improving experimental success rates?

Meanwhile, MobiDrop has taken the lead in addressing the industry's shortcomings in data dimensions by constructing an integrated multi-omics data system covering transcriptomics, proteomics, epigenomics, microbiomics, and spatial omics. This approach breaks through the dimensional limitations of relying solely on RNA data, comprehensively reconstructing the complete logical framework of cellular activities, including cell growth, gene regulation, drug response, and environmental adaptation.

Amid the industry-wide impatience to deploy applications at scale and chase short-term trends, MobiDrop has remained steadfast in its commitment to industrial pragmatism.

Regarding the implementation pace within the sector, Pei Hao maintains a clear-headed judgment: AI virtual cells are still in the early stages of industrial development and cannot fully replace traditional wet-lab experiments in biopharmaceutical R&D. Their core short-term value lies in pre-screening low-value experiments, reducing the frequency of high-cost experiments, optimizing R&D priorities, and proactively mitigating risks.

"Virtual cells represent the advanced next-generation form of AIDD," explained Pei Hao. Traditional AIDD focuses on surface-level issues such as drug molecule structure design and affinity prediction, whereas virtual cells can further deduce the full-dimensional changes in the overall cellular state, target pathways, and perturbation responses following drug intervention, thereby bridging the complete link in drug R&D from molecular design to cellular response.

In line with the industry's pattern of gradual development, MobiDrop has established a multi-tiered, stepwise commercialization strategy. In the second half of 2026, the company will roll out a series of AI-driven biotech innovations. Leveraging its proprietary Virtual Cell Core Engine, MobiDrop will deeply empower core R&D processes for pharmaceutical companies, including target discovery, mechanism validation, drug screening, and toxicity warning.

Meanwhile, MobiDrop has established a deep strategic partnership with leading domestic AI large-model enterprises. By leveraging its proprietary data advantages and wet-lab implementation capabilities, and integrating the computational power and algorithmic strengths of its partners, the company aims to jointly refine a localized virtual cell technology system tailored to the Chinese biopharmaceutical landscape.

Leveraging years of deep industry engagement, MobiDrop has directly served over 220 core research and industrial clients, empowering more than 1,000 end-user institutions, with comprehensive coverage of leading innovative pharmaceutical companies, Grade A tertiary hospitals, top-tier universities, and research institutes. The long-established customer base for single-cell technology provides natural application scenarios for the deployment of virtual cell technology, enabling seamless penetration and rapid iteration of new technologies without the need to develop new markets.

At the industrial ecosystem level, MobiDrop continues to deepen its commitment to domestically produced AI-driven biological data infrastructure, spearheading the establishment of a Virtual Cell Data Alliance. By integrating high-quality resources from industry, academia, and research institutions, the company addresses critical industry pain points in China, including fragmented biological data, inconsistent standards, and low levels of data sharing.

At the capital level, MobiDrop has secured backing from mainstream top-tier investors, including Huagai Capital, Source Code Capital, Proxima Ventures, and LYFE Capital. The company is currently launching a new round of financing to expand production capacity for high-quality causal perturbation data, upgrade cutting-edge cell measurement technologies, and continuously strengthen its core technical framework of "real cells continuously training virtual cells."

Looking ahead, when asked "Which three keywords would you like the outside world to remember MobiDrop by?" Pei Hao provided his answer: data-native, dry-wet closed loop, and TechBio.

The first two keywords have permeated the entire course of the company's development, while the third keyword, "TechBio," must be understood within a broader industrial context.

Traditional biopharmaceutical companies are primarily composed of Biotech and BioPharma entities, with "biology" as the core driver. Hao Pei believes that future industry giants may no longer be traditional Pharma companies, but rather “TechBio” enterprises—those that deeply integrate cutting-edge technologies into biology, thereby redefining the traditionally labor-intensive and low-automation life sciences R&D processes through data and algorithms.

This judgment forms a complete loop with Pei Hao's original aspiration from over a decade ago. The mechanical engineer who once walked into a biological laboratory, puzzled by the question "Why can't repetitive tasks be automated?", is now answering that very question step by step through microfluidics, automation, AI, and virtual cells.

Decades ago on a blackboard at Caltech in 1988, Richard Feynman wrote, "What I cannot create, I do not understand." Across decades, this quote serves as the perfect footnote to Mobio Biotech's development path. Rather than merely observing and describing cells, the company strives to truly understand cells by generating data, conducting experiments and building closed-loop systems. Ultimately, it aims to construct verifiable, interactive and deliverable virtual life forms in the digital world — this is the underlying logic that Mobio Biotech has consistently pursued.

Now, in laboratories across Zhejiang, China, MobiNova® equipment operates stably with a throughput of tens of thousands of cells per day. Each droplet generation, each reagent exchange, and each batch of sequencing data output contributes real-world nourishment to the still-evolving "virtual cell."

"We must still commit to generating such data, allowing real cells to train virtual cells."Pei Hao's statement serves not only as the company's current guiding principle but also holds the key to the future of China's TechBio industry.

From mechanical engineering to microfluidics, and from single-cell sequencing to AI virtual cells, Pei Hao has spent over a decade addressing an engineer's most fundamental question in life sciences: Can we make life science research smarter?

Answers are growing within the data.