Where Are the Opportunities for China's AI4S Industry in Trump's "Genesis Project"? | Linear Voice

On November 24, Trump formally signed the "Genesis Mission," described as "the largest federal mobilization of scientific research resources since the Apollo Program," with the goal of doubling American scientific productivity and influence within a decade.

This initiative is less a mirror than a clarion call — a reminder that the global AI for Science (AI4S) competition has entered a new phase of nation-state rivalry. Against this historic backdrop, where exactly does China's AI4S exploration in labs and industry stand? What distinct characteristics define Chinese and American AI4S development? And over the next 5 to 10 years, what exciting breakthroughs might emerge in this field?

Today's discussion draws from Linear Capital partner Yingzhe Zeng's conversations with three AI4S portfolio companies — GV20 founder Xiaole Liu, Yinghua Chenrui founder Zhen Zhou, and Deep Principle founder and CTO Chenru Duan — who share their real voyages and profound insights from piloting AI as an entirely new vessel through uncharted scientific waters, from the perspectives of biomedicine, molecular discovery, and materials.

The course of history is quietly altered during revolutions in tools. When the astrolabes and sailing ships of the Age of Discovery gave way to precise coordinates and surging engines, humanity's cognitive boundaries of the physical world were fundamentally reshaped. Today, scientific exploration stands at a similar inflection point: AI for Science is leading a revolution in scientific paradigm.

How is this transformation occurring? What "impossible" tasks is it actually solving? More importantly, within the grand narrative of global technological development, what future does it portend? At Linear Capital's AGM, three "explorers" from the frontiers of materials, biomedicine, and molecular discovery — Dr. Zhen Zhou of Yinghua Chenrui, Dr. Xiaole Liu of GV20, and Dr. Chenru Duan of Deep Principle — shared their real voyages and profound insights from piloting AI as an entirely new vessel through uncharted scientific waters.

Their observations converge on one core insight: in this global AI4S competition, Chinese companies and Chinese-founded startups possess tremendous potential. Facing the "impossible" problems in materials, life sciences, and chemistry — highly complex, multi-variable challenges — they can transform the lengthy trial-and-error processes once dependent on intuition and chance into predictable, designable, engineerable rational exploration. From an investor's perspective, this may also be one of the most disruptive and commercially valuable investment themes of the next decade. Below are highlights from this in-depth conversation:

---

Zhen Zhou, Yinghua Chenrui:

Our field operates within the very complex system of polymer materials. Traditional R&D cycles are extremely long, trial costs are high, and success rates are very low. Take the plastic water bottle we casually hold today. The material was developed in 1941, yet for a long time it was only used in fibers and films. It wasn't until 30 years later, with the serendipitous discovery of processing techniques and crystallization technologies, that someone developed PE bottles. Therefore, we believe AI will bring transformative changes to polymer materials at three levels:

First, with the invention and iteration of deep neural network technologies represented by Transformers, we use AI through data-driven approaches and algorithmic training to better understand the mechanisms of how polymer materials interact across different scales, developing materials from structure-property relationships — including reverse design of application solutions.

Second is the integration of AI with quantum mechanics and molecular simulation. We can now obtain highly reliable foundational data through computational methods while maintaining excellent precision. This data-level supplementation brings tremendous assistance to our model development and iteration.

Third, AI enables significant advances in multimodal spectroscopic techniques. Through our multimodal technologies applied in materials characterization, we can build extensive databases of different detection spectra. Simultaneously, through combined approaches, we can obtain more comprehensive and direct analysis of material structures across different scales at the same time. This ultimately greatly benefits the development of large models for polymer materials. These large models can both accurately predict performance for new materials and enable reverse design based on performance requirements.

Xiaole Liu, GV20:

We use AI for biomedicine. I spent 20 years doing cancer research at Harvard, always believing that biological insight could greatly reduce dependence on data and computing power. Through years of immunology research, we discovered that B cells next to patient tumor cells are actually producing anti-cancer antibodies — just not in sufficient quantities, as if the tumor had already become a "wartime occupied zone." We examined sequencing data from tens of thousands of tumors, containing hundreds of millions of antibodies, then used AI to calculate which proteins these antibodies were targeting. So what exactly can AI bring us? People usually say "more, faster, better, cheaper" — in our field it should be "accurate, fast, good, cheap."

"Accurate" means finding a completely novel target that no one knew about before. Through tens of thousands of tumor datasets and hundreds of millions of antibodies, we used AI to discover why many patients' tumors produced small amounts of antibodies targeting one particular marker. At the time, no one was studying this target — there were fewer than 10 papers worldwide on it. We didn't fully understand it either, but with a spirit of experimentation, we directly extracted the antibody from patient tumors, injected it into animal tumors, and saw tumor growth inhibition. We pursued the mechanism through reverse research while simultaneously advancing the drug to clinical trials.

"Fast" means what? In most cases, we need to thoroughly understand a gene's function before starting drug development. But we already had the drug, knew it worked in animals, and advanced it to clinical trials while doing reverse mechanistic studies. So from first seeing this gene to entering clinical trials took us just three years, and completing Phase I took five years. Compare this to Amgen's PD-1 drug, the current "drug king" — from research start to Phase I took 20 years; we took five. Moreover, our drug is the world's first to use AI for target discovery, AI for antibody design, and reach clinical trials.

Third is "good." Unlike the many unknowns in traditional pharmaceutical processes, ours was built on antibodies that many patients had already produced, so production was very stable, and what entered the human body was a naturally occurring antibody with excellent performance all around. One patient, before using our drug, had tried five different drugs over seven months with no success. After starting our drug, liver metastases disappeared at five months, and tumor shrinkage exceeded 30% at seven months.

Finally, AI also saved us substantial money. For our clinical trials, a typical cancer clinical trial in the US costs about $250,000–300,000 per patient, but the actual hospital, patient, physician, and drug costs don't exceed $100,000 — the remaining $100,000+ goes to CRO companies and various consultants. We later began using computational optimization, reducing patient enrollment costs by over 40%. Recently we've also been conducting clinical trials in China, which is even faster, better, and cheaper.

Chenru Duan, Deep Principle:

Both Dr. Liu and Dr. Zhou mentioned one point — the design space for chemical molecules and materials is extraordinarily large. Taking previous AI-solved problems as examples, AlphaGo's 19×19 board corresponds to 361 choices, already the most complex among board games. But the design space for small-molecule drugs is estimated at 10^60, equivalent to having 10^60 possible moves on the board — something completely impossible to solve with traditional AI.

We believe the significant change actually came around 2022, with the arrival of generative AI. Instead of considering what the entire board looks like when placing a piece, we can place pieces while observing what the surrounding environment looks like, focusing on more critical, more relevant chemical space, allowing us to explore molecular and materials discovery step by step in a focused manner.

A real case we've recently worked on internally involves exploring reaction networks from known substrates, like continuously playing chess. Through exploration, we discovered a rather interesting intermediate that previously required catalyzing a system with expensive raw materials and less-than-mild reaction conditions. Now, under relatively mild conditions, we can achieve this without adding catalysts. I believe discovering such new reaction mechanisms is a crucial means of advancing new materials to market, and also a direction we value highly in the long term.

Chenru Duan, Deep Principle:

People used to pay great attention to "scientist entrepreneurs," but later found that the success rate for scientists directly starting companies was quite low. From the scientist's perspective, the mentality is usually wanting to "save up a big move" from a Science angle, but the real world is like RL — we need to understand the surrounding world and generate more connections with surrounding industries to obtain Reward, advancing things gradually rather than always wanting to succeed in one step and have everyone in the world buy our product.

We officially started our company in June 2024. At the very beginning, we considered algorithms as our most important moat. But many changes occurred over this year and a half. The most important moat shifted from algorithms initially to "how to generate our own model and data iteration methodology" — whether through data from client collaborations or distilling a self-reinforcing data flywheel from this data. This was a moat achieved within one year of starting up.

In the most recent six months, we've done extensive iteration and experimentation with various top-tier clients. Whenever we exert maximum effort and find something still can't be perfectly solved, we can often summarize substantial know-how, and this know-how itself is priceless.

I think in the long term, the most important thing is the team. Because our direction is very new, our employees are all very young — though I only recently completed my PhD, I'm already the fourth oldest in the company. This is a highly interdisciplinary field. Our technical staff need both Science backgrounds and AI backgrounds. In our product development or business expansion, we need to understand both clients and certain technical aspects. People need to speak different languages to connect these things and make the business model work during scientific exploration.

Xiaole Liu, GV20:

When we first started, people also wondered — you just have one technology, what qualitative change can you produce compared to others? But the drug development process is very long, so we must know how each step is done to use AI to make every step "accurate, fast, good, and cheap."

Some say with AI biomedicine, can you just sell software? Actually, big pharma won't spend much money buying software — maybe tens of thousands or a few hundred thousand. Others say I can help big pharma design a molecule, but big pharma might only spend a few hundred thousand or a million on that. But think about it — in the past 20 years, which drug was actually developed by big pharma themselves? Never — they generally acquire from small companies after Phase I or II clinical development. Their own R&D teams have never developed any molecules.

In this situation, if an AI drug company is just selling software there, or selling a preclinical molecule, it's really hard to get big money. But once a drug reaches clinical trials and proves effective in patients through Phase I-II, it can do a deal with pharma — potentially hundreds of millions, billions, or even tens of billions. So we hope to use AI to improve every step from early discovery through early clinical development, maximizing value for the company. Doing a complete solution, the solution closest to the customer — this has the greatest commercial value.

Zhen Zhou, Yinghua Chenrui:

Both Dr. Duan and Dr. Liu mentioned many capabilities that AI4S companies should possess, whether from the data level or capability integration level — I very much agree. Specific to polymer materials, I think another more important aspect is understanding the entire industry chain. In reality, the journey from material itself to application is very long. For example, take the clothes we wear — from material it goes through fiber, yarn, weaving, dyeing and finishing, garment making, with industry know-how at every step that clients won't necessarily tell you. If we need to disrupt it or bring something new, we must consider these links.

So for our AI4S field, beyond the interaction or fusion of materials and AI, I think more lies in engineering scale-up, including understanding the industry in various domains. Ultimately, moving from a technology platform company to commercialization requires matching various capabilities — like a puzzle where only when all pieces fit together does the picture show its most beautiful form.

Zhen Zhou, Yinghua Chenrui:

For the polymers our company empowers with AI, our early focused industry tracks are likely 3D printing and bio-based fibers. In the past half year, among products developed through our technology, what most impressed clients and made them willing to pay is probably polylactic acid fiber. The biggest problem with this fiber is actually its long-term mechanical degradation, causing it to fail at certain downstream stages in the textile industry.

Addressing this problem, over the past six months we developed a very interesting hydrolysis-resistant, aging-resistant fiber material. Compared to traditional PLA, under accelerated aging conditions, it can maintain complete mechanical performance stability for one month, corresponding to basically 2–5 years under natural conditions. Essentially, we developed a completely different product targeting an industry pain point.

Xiaole Liu, GV20:

Actually for drug companies like ours, our direct customers aren't cancer patients but big pharma. For big pharma, what they find amazing is that you took a gene no one knew about, brought it to clinical trials so quickly, and it actually worked in patients — they find that very impressive.

We originally conducted all Phase I trials in the US, but over the past year we've also begun clinical trials in China. We won't give our lead program to pharma now, because the next year may yield a value inflection point. Then pharma will ask, do you have any other molecules? If this one works, are there newer molecules behind it? So this year we actually did a deal with pharma, collaborating with them on some of our other molecules.

Chenru Duan, Deep Principle:

We've only had commercialization in the last 6 months. I'll share one case — we're working with a European and American beauty company to solve an "ancient problem," namely the stability of a certain active molecule under sunlight. This molecule was discovered starting in 1970, and its stability has never been solvable.

We did a POC project, recommending over ten molecules to stabilize this active ingredient to the company. Initially their R&D staff were quite candid, saying if even one of your ten-plus molecules works, that would already be impressive, we'd give you an award. Unexpectedly, after lab experiments came back, every single one of our recommendations exceeded their control group, and we subsequently launched deep collaboration.

Xiaole Liu, GV20:

I think developing drugs in China has unique advantages: First is Chinese talent. In the US there are more people with very senior pharmaceutical experience, but there are enormous numbers of PhDs and postdocs in China, and labor costs are cheaper than the US, plus they work really hard, learn very fast, and pick up computing quickly.

Second is the entire foundational infrastructure. For example, CRO companies doing reagents, mouse experiments, toxicology, pharmacology, including clinical CROs — this foundation has been well built over 20 years of industry investment. We tried doing some experiments in the US, and ultimately found orders were sent to China, done there, then shipped back — less efficient than us just going directly to Zhangjiang in Shanghai.

Third is clinical resources. In China, when people get cancer they generally go to major tertiary hospitals — for example, we're collaborating with Shanghai Pulmonary Hospital, which performs 30,000 lung surgeries annually. A US hospital wouldn't do this many across all cancer types combined in a year — it's unimaginable.

Finally, the Chinese government has invested heavily in scientific research. Now the Trump administration is suppressing universities, health departments, and funding agencies. In this situation, Chinese innovation — even grassroots doctors want to develop new drugs, do related research, because they know there's a good reward mechanism once successful.

Chenru Duan, Deep Principle:

Before returning to China, I worked at Microsoft, doing AI4S-related product integration on Azure. One thought I had was why cloud is so expensive — basically a V100 costs three dollars per hour. A major change after returning was that computing power itself is relatively cheap, though there are still certain limitations at the highest performance tier, but for computing power generally usable in AI4S currently, our cost advantage is very significant.

And this computing power can help us generate data faster. Unlike large language models where open internet corpora have already been scraped clean, and quality corpus generation speed depends on population size. But computing power itself is something that can grow quickly and is relatively cheap in China, so advantages in computing power will cascade into data advantages, then feedback into model advantages, getting the entire flywheel spinning.

Additionally, the current moment is very interesting. Several star startups have recently emerged in Europe and America, like CuspAI, LiLa Science, Periodic Labs — they all raised big money in their first round. The environment is different, and people's thinking is very different: they still have very high-profile, ambitious ideas, wanting to save up a "nuke" — for example, doing high-temperature superconductivity, doing nuclear fusion; but I think AI4S itself still needs strong industrial integration, because creating an entirely new industry takes a very long time.

Even with frontier discoveries, integrating into industry is still an essential and lengthy process. Doing these star-gazing things is of course important, but in this process we can certainly do more down-to-earth things, solving real industry needs.

Zhen Zhou, Yinghua Chenrui:

Because AI4S often involves vertical domain practical implementation, moving from the lab inevitably requires industry customers to have high acceptance and tolerance for new materials — this is precisely a very important advantage of China's innovation soil.

I previously worked at foreign enterprises doing global project development. There was one project for both European/American and Chinese customers — when our lab product prototypes came out, during actual scale-up and iteration, we found a huge difference. Foreign customers are extremely cautious about new products, with very long cycles and many restrictive conditions; while Chinese customers are very receptive to new materials and products, even willing to clear equipment time slots and adjust different operating conditions to meet experimental requirements.

Comparing timelines, testing cycles with Chinese customers may be one-third of those with European/American customers, with rapid iteration possible. So I believe China's soil is very suitable for AI4S future implementation across many application layers.

Zhen Zhou, Yinghua Chenrui:

For polymer materials, the biggest revolution in five years may be disrupting the existing entire industry ecosystem, bringing maturity and widespread adoption of polymer materials large models. The industry is currently very long, with very long investment and maturation cycles.

Through large models, we can use models to predict future performance from the design stage of new materials, or identify applicable industry directions. This can transform what the industry currently considers a 20-year cycle into 2–3 years. New materials can quickly enter the market and rapidly scale in suitable application scenarios.

On the other hand, for any new application, such models can quickly find suitable materials and formulations, making the entire industry more efficient and intelligent.

Xiaole Liu, GV20:

From a scientific discovery perspective, I believe AI will bring very important breakthroughs in understanding immunity. Actually, people in developed countries now have infectious diseases from birth, then allergies, autoimmunity, various inflammations — all because the immune system has problems, either overactive or underactive.

Genomic data has been increasing in recent years, and with so much data, we need AI to truly figure out the immune causes in different diseases. This is not only interesting scientific exploration, but I believe will continuously produce new drugs in the pharmaceutical industry. Like many traditional Chinese medicines that work because they modulate patients' immune systems — we don't fully understand the scientific mechanisms behind them, but perhaps through AI we can calculate exactly how they activate or suppress abnormal immunity.

Chenru Duan, Deep Principle:

Looking at a 10-year timeline, I believe high-temperature superconductivity will definitely be broken through by AI4S. But this may not be the most impressive outcome. Over these 10 years, AI itself will very likely help us unlock any material we want to synthesize: we can not only design what materials are through reverse approaches, but also combine with forward approaches to realize them in the physical world.

Commercially, this may bring paradigm changes to R&D in the chemical materials field. Chemical materials currently still largely follows integrated R&D — from initial materials design through scale-up to production and distribution, all dominated by large manufacturers. But if we have a sufficiently powerful R&D machine, what we call an AI Materials Factory, I believe we can disaggregate the entire industry chain, achieving a collaboration model more like that between biotech and pharma, letting everyone focus on what they're best at, working together to make human society's exploration of the material world extremely rich.

Chenru Duan, Deep Principle:

Europe and America currently have unprecedented attention on AI4S, but it's still mainly driven by foundational models or PaaS, and seems difficult to transition from virtual to real in industry-related applications. This relates to their own economic structure — at this point in time, doing enterprise services or models can quickly generate commercial value at this level.

But from China's perspective, not having established PaaS business models beforehand has actually become a good thing, pushing us to better understand the deeper virtual-to-real transition. Like what we're doing — we won't take pure enterprise service as our sole purpose. It must be integrated into industry, with substantive discoveries that change current industrial structures, for the logic to hold.

AI4S is the first direction in my life where I see China and America working on it simultaneously with no essential technical gap. This field has important significance for future technological格局, and has thus been incorporated into high-level strategic vision on both sides. We hope entrepreneurs in this field can be more confident, and view industry competition and cooperation with a more internationalized perspective. Perhaps the two markets currently have large gaps in capital support, but we have more industrial advantages, and we'll catch up in these areas subsequently.

Xiaole Liu, GV20:

My personal feeling is that the US government currently, in actual operations, is quite suppressive or closed toward higher education and science, while China actually has very large investments in this area. Meanwhile, I think Chinese people doing AI4S have unique advantages, whether in China or in the US.

In the US, many white students who study biology do so because they're bad at math, physics, and chemistry, but Chinese people who do biology have excellent math, physics, and chemistry foundations, and can quickly pick up computing upon exposure. I actually think the ultimate competition will likely be between Chinese people and overseas Chinese — both sides will do very well.

Zhen Zhou, Yinghua Chenrui:

I think the current AI4S wave has historical similarities, much like the previous internet wave: America is better at 0-to-1 breakthroughs and underlying technologies, because it has innate capital catalysis and technology development. But China's great advantage lies in application scenario implementation, because China has massive industrial foundations, including data, resources, and demand across various scenarios. China often excels in细分领域, through learning, including data and industry integration to achieve "corner overtaking."

So I've always believed that if AI4S future moves into various细分领域 and vertical domains, Chinese soil is definitely very suitable. And at the computing power and infrastructure level, much AI4S doesn't necessarily require the most advanced GPUs or cutting-edge computing power. But how to integrate with applications, obtaining needed data through architectures and workflows adapted to it — at this point we can circumvent some of China's temporary lag in highest-end computing power in the short term.

From long-term applications including industry track implementation, Chinese soil is very suitable for doing AI4S. I also believe that over a longer timeframe, say 5–10 years, with continuous iterative development, China can at least surpass America in specific AI4S tracks.