Qiming Venture Partners | Weijie Sun: DeepWise Aims to Be the Google of Scientific Discovery
The real moat of a scientific product doesn't come from any single breakthrough, but from the feedback, refinement, and repeated iteration that comes from sustained customer use.
Editor's Note: Amid the AI for Science boom, most companies have focused on using models to accelerate single R&D steps. DeepWise chose a harder path: rebuilding the infrastructure for scientific discovery. From day one, it rejected the limitations of point solutions, making "building scientific research infrastructure" its core mission and connecting the full research chain of "read, compute, do, and intelligize." From the DeePMD method winning the Gordon Bell Prize to replicating the AlphaFold 2 training system, from the Hermite drug design platform to the Bohr Transition Lab, this Chinese startup has consistently balanced technological innovation with commercialization, aiming to become the Dassault of the microscopic world. As "AI scientists" become the industry's new vision, DeepWise hopes to become the platform company powering "humanity's last productive job." This article breaks down its development path and core strategy, showcasing the long-term thinking and innovative force of China's AI for Science enterprises. Reproduced with authorization from the Qiming Venture Partners WeChat official account.
In the narrative of AI for Science (AI4S), the easiest stories to tell are always those dazzling point solutions: a stronger model, a faster prediction, a sufficiently stunning academic result.
But a Chinese startup — DeepWise — has far greater ambitions. Founded before the large model wave erupted, DeepWise was among the earliest teams to bet on the commercial potential of AI for Science. It didn't wager on a single drug pipeline, one material system, or even a star model. Instead, it committed to something more important and slower: reorganizing the research process, originally scattered across literature, computation, experiments, and expert experience, into a sustainably iterative infrastructure.
DeepWise chose to become a platform company — a decision grounded in both commercial logic and value judgment, because they believe that in scientific discovery, the platform approach will deliver more long-term value than point solutions. "How far are we from an AI scientist?" This leading Chinese AI for Science company is answering that question through its own practice.
01/ A Paper and a Judgment
DeepWise's starting point was a paper, but what truly determined its direction was a business judgment.
In 2017, Linfeng Zhang, DeepWise's co-founder and chief scientist, proposed the "Deep Potential Molecular Dynamics" method (DeePMD) during his PhD at Princeton University. Put simply, everything in the world is made of atoms; to understand the properties of a material or drug, one must ultimately calculate the interactions between atoms. The foundational equation of quantum mechanics — the Schrödinger equation — can theoretically solve this, but at prohibitive cost.
DeePMD opened a more efficient path for atomic-scale computation: using deep learning to approximate interatomic interactions, thereby pushing simulations — originally too expensive for large-scale use — to larger systems and longer timescales while maintaining accuracy as much as possible. In 2020, this work received the Gordon Bell Prize, the highest honor in global high-performance computing.
At the time, Weijie Sun, the company's co-founder and CEO, was still completing his master's at Peking University while interning at an investment firm. He and Zhang were classmates at Yuanpei College, Peking University, and knew each other well. Weinan E, one of Zhang's PhD advisors and an academician of the Chinese Academy of Sciences, said this was an opportunity he hadn't seen in over 30 years. It was Academician E and Academician Chao Tang who first proposed that AI for Science should be vigorously developed.
In the summer of 2018, the three gathered to discuss how to advance this vision. Their consensus: to push faster and deeper, the most effective form would not be remaining in the academic system, but founding a company. DeepWise was registered in Beijing at the end of 2018 and began formal operations in mid-2019.
Sun's explanation is that the core judgment was "marketization is more effective." In his view, the real moat for scientific products comes not from any single breakthrough, but from customer feedback, refinement, and re-iteration after sustained use. This is why DeepWise was never a "do research first, figure out commercialization later" company from day one, but one attempting to advance research, engineering, and commercialization simultaneously.
In 2018, the "GPT moment" was still over four years away. AI for Science was far from a widely accepted concept, and "scientific research infrastructure" had few existing templates — meaning industry pioneers would inevitably bear higher market education costs. Sun recalls that before AI4S became consensus, DeepWise did much "beyond the normal scope of corporate work," including hosting summits, publishing industry reports, and continuously explaining to the market what AI for Science actually meant.
For a startup, this sounds like extra burden; but from another angle, it shows DeepWise wasn't betting on an existing market, but on a new paradigm still taking shape.
02/ From Paper to Model Platform
After founding, the first thing DeepWise did was not grand platform narratives, but something more modest: turning methods from papers into callable model capabilities.
Thus, in its early days, DeepWise concentrated resources on building underlying general-purpose platforms centered on atoms, molecules, genes, proteins, and other objects. Sun summarized this phase's framework as "one horizontal, two verticals": the "horizontal" being the AI for Science foundational platform, and the "two verticals" being life sciences and materials science.
This step was critical because it completed the most underestimated first step in technology commercialization: getting models out of paper appendices and into real R&D workflows. A paper can prove a method works; a platform must prove the method can be repeatedly called, engineering-deployed, and used by different users in different scenarios. Many of DeepWise's later capabilities — whether in drug design, materials R&D, or later large models and agents — were built on this foundation.
After AlphaFold 2 shook the global biopharma field in 2021, DeepWise's algorithm team quickly devoted itself to replicating its training system. But DeepMind had not publicly released complete training code at the time; replication wasn't simply running results from the paper, but a comprehensive test of algorithmic understanding, engineering capability, and infrastructure capacity. This work was later recognized as among the earliest global completions of a full AlphaFold 2 training replication.
Its importance lay not merely in "keeping up with the frontier," but more crucially, in proving to the outside world that DeepWise was not a company with only the original DeePMD innovation, but one with sustained ability to absorb cutting-edge methods, complete large-scale training, and engineer methods into production. For a platform company, this capability is almost a necessary admission ticket.
But replicating AlphaFold was only capability verification, not commercial verification. Proving "we can build it" is one thing; proving "we can sell it and sustain it" is another. DeepWise's real hard battles happened at customer sites, not in papers.
03/ Customers Don't Buy Models, They Buy Results
Top-tier technical models don't guarantee efficient commercial conversion. Looking back at DeepWise's founding, Sun says model development moved faster than expected, while commercialization proved harder than anticipated.
In the early days, to find direction, the team surveyed no fewer than 50 industries for technical feasibility and market opportunity matches, ultimately selecting two entry points with the highest probability of success and clearest market demand: small-molecule drug R&D and battery materials design. The logic behind this choice was clear: both fields rely heavily on atomic- and molecular-scale understanding, and both have real customers where commercial loops can more easily form.
DeepWise's first core commercial product was Hermite, a platform for drug design. This direction was chosen both because overseas benchmarks like Schrödinger already existed with relatively mature market education, and because the pharmaceutical industry is highly sensitive to R&D efficiency with stronger willingness to pay. But the real difficulties soon emerged: running a few impressive cases in the lab doesn't mean an industrial-grade product is formed.
Sun used autonomous driving as an analogy in the interview. Take Hermite's FEP function: a senior computational chemistry PhD student running 4–8 cases in the lab isn't particularly hard; but marketization requires covering 20-plus common scenarios, and industrial-grade availability means stable operation in the vast majority of real projects. Whether protein structures are stable, whether metal elements are involved, whether salt bridges and covalent bonds come into play — details that "toy models" can avoid — all become problems that must be solved when facing customers.
DeepWise's first customer taught the team exactly this lesson. Sun recalls that after the product was sent for trial, the client quickly raised a string of issues that mature products shouldn't have. The team first iterated internally, then did closed development at the customer site, gradually getting past the cold start phase. This process lasted several years before stable product and customer feedback loops gradually formed. "To this day I'm deeply grateful to our first partner," Sun said. "After getting past the cold start, more and more customers began using it, forming a positive cycle." This product iterated for nearly four years and now holds the highest domestic market share.
This experience left DeepWise with an important methodology: technological breakthrough is only the seed; the real moat comes from the closed loop of "technology — product — user adoption — feedback — re-technologization." What customers want are more reliable results, smoother workflows, and more predictable R&D efficiency.
From this, his definition is: "Be both the first product and the last product." Because, "being the first product gives you a development window; being the last product means that once users have your product, they don't need to use anything else."
It was also during this phase that DeepWise gradually expanded from pharmaceuticals to batteries and materials. The launch of products like Piloteye wasn't simply replicating Hermite's experience, but verifying whether the same underlying capabilities could cross into different industrial scenarios. Currently, DeepWise's intelligent products and solutions have entered multiple directions in life sciences and materials science, with an increasingly impressive client roster including Hansoh Pharmaceutical, Fosun, CATL, and BYD.
In 2020, Sun found a precise commercial benchmark: Dassault Systèmes. This French industrial software company started with aircraft design and simulation, using solid mechanics, fluid mechanics, and electromagnetism methods for industrial R&D at the macro scale. This is exactly what DeepWise aims to do — internalize quantum mechanics into software and do the same thing at the microscopic atomic and molecular scale, becoming the Dassault Systèmes of the microscopic world.
04/ Full-Chain Scientific Research Services
If the first half of DeepWise's journey was mainly about proving that "compute" could work, then the most notable change in its second half is its growing layout around the research process itself.
Sun abstracted scientific research into four links in the interview: read, compute, do, and intelligize.
"Read" is the organization and retrieval of literature, patents, data, and knowledge; "compute" is scientific calculation, simulation, design, and prediction; "do" is experimental verification, bringing models back into the physical world; "intelligize" is giving systems higher-level autonomous calling and coordination capabilities, gradually approaching the form of an "AI scientist."
This abstraction explains why DeepWise's product matrix has grown increasingly broad. Hermite and Piloteye correspond to "compute"; Bohr Scientific Navigation and related knowledge infrastructure correspond to "read"; the Bohr Transition Lab corresponds to "do"; and SciMaster, PharmMaster, MatMaster and other scientific agent systems correspond to "intelligize."
What truly deserves attention is the "do" layer. Many AI for Science companies excel at "read" and "compute" — integrating literature, training models, proposing candidate structures or formulations. But DeepWise recognized early that if experiments don't connect, AI's value is hard to fully deliver to customers; without continuous, standardized experimental data flowing back, models also struggle to form true data flywheels.**
DeepWise's Bohr Transition Lab attempts to turn laboratories into schedulable, reusable, data-producing systems — not merely stacks of automated equipment. This step is important, and therefore closer to infrastructure construction. It means DeepWise is no longer just building smarter scientific software, but trying to close the loop of "computation guiding experiments — experiments feeding back to models."
By the end of 2025, Bohr Scientific Navigation had served over 100 universities and more than 4 million scientist users, cumulatively integrating over 200 million English-language papers and 200 million patents, plus 80 million Chinese-language papers. DeepWise's scientific intelligence products have served hundreds of advanced R&D enterprises. Of course, these numbers indicate coverage, not final victory. The real difficulty remains: whether these tools and systems have embedded deeply enough into scientists' workflows.
In December 2025, DeepWise announced the completion of its Series C round exceeding RMB 800 million. Across multiple funding rounds, DeepWise has been backed by Qiming Venture Partners and other top-tier institutions.
05/
Building the "AI Scientist"
DeepWise's first five-year goal, set in 2020, was to complete building its micro-scale industrial R&D platform by 2025. According to Sun in the interview, this goal has been "basically achieved": technical elements, R&D systems, and product systems are largely formed. The company's next-stage goal is to build an "AI scientist" by 2030.
This is certainly an extraordinarily ambitious goal. "This target isn't about 'aim for the best, settle for the middle,' but established based on the insights and confidence built through the company's development journey," Sun said. In 2022, he predicted that AI4S would complete infrastructure construction and enter an application explosion phase by 2025. At the time, everyone thought he was overestimating. "Turns out I was actually too conservative."
"In 2025, China, the US, and Europe almost simultaneously released AI for Science strategic plans. China listed 'AI + science and technology' as the top of six key actions in its Opinion on Deepening Implementation of the "AI+" Action; the US launched the 'Genesis Project,' describing it as a national initiative comparable to the Manhattan Project in urgency and ambition. Meanwhile, international giants have also entered the fray. Beyond Google's DeepMind, Anthropic announced an innovative AI for Science program in May 2025; in October, OpenAI announced the formation of a new 'Science Division.'
Today's AI is still better at executing human-defined problems than independently posing truly important ones. It can significantly improve efficiency on localized tasks, but cannot yet claim mature scientific judgment. DeepWise is preliminarily establishing infrastructure covering the full scientific research process — not a pure model company, but a systems-level platform with knowledge, computation, experimentation, and tool scheduling capabilities. DeepWise hopes to hand the gradually forming capabilities of "read, compute, and do" to agents for calling and orchestration, so that AI becomes not just an accelerator for certain steps, but a collaborative actor in the research process.
As digital-age infrastructure, search engines build humanity's shortest path to known information, but don't directly provide knowledge; large models build humanity's shortest path to known knowledge, organizing and presenting it. In Sun's view, AI for Science is building humanity's shortest path to unknown knowledge.
Based on this premise, the threshold for human knowledge accumulation will drop dramatically, scientific discovery will become simpler, and more people will have the opportunity to become scientists. Meanwhile, as AI and robotics liberate productivity, humans will also have more energy to devote to scientific production. So Sun is very optimistic: "This will be humanity's last productive job."
Imagining DeepWise's future in this context, Sun's answer is: the Google of scientific discovery.
Previous Articles
Qiming Venture Partners was founded in 2006. Currently, Qiming Venture Partners manages 11 USD funds and 7 RMB funds, with total committed capital reaching $9.5 billion. Since its inception, it has focused on investing in outstanding early- and growth-stage enterprises in Technology and Healthcare innovation.
To date, Qiming Venture Partners has invested in over 580 high-growth innovative companies, of which more than 210 have listed on the New York Stock Exchange, NASDAQ, Hong Kong Exchanges and Clearing Limited, Shanghai Stock Exchange, and Shenzhen Stock Exchange, or exited through M&A and other means. Over 80 portfolio companies have become recognized unicorns or super-unicorns in their industries.
Many Qiming Venture Partners portfolio companies have grown into the most influential companies in their respective fields, including Xiaomi (01810.HK), Meituan (03690.HK), Bilibili (NASDAQ:BILI, 09626.HK), Zhihu (NYSE:ZH, 02390.HK), Roborock (688169.SH), Hesai Technology (NASDAQ:HSAI, 02525.HK), UBTECH (09880.HK), WeRide (NASDAQ:WRD, 0800.HK), HyperStrong (688411.SH), Insta360 (688775.SH), Unisound (09678.HK), Biren Technology (06082.HK), Zhipu (02513.HK), Gan & Lee Pharmaceuticals (603087.SH), Tigermed (300347.SZ, 03347.HK), Zai Lab (NASDAQ:ZLAB, 09688.HK), CanSino Biologics (688185.SH, 06185.HK), Schrödinger (NASDAQ:SDGR), MicroPort EP MedTech (688617.SH), Sanyou Medical (688085.SH), Amoy Diagnostics (300685.SZ), SinoCellTech (688520.SH), Insilico Medicine (03696.HK), Harbinger Health, Yuanxin Technology, MediLink Therapeutics, LaNova Medicines, StepFun, and others.