HomeArticle

AI4S is red-hot, how far is it from the "ChatGPT moment"?

定焦One2026-11-17 09:21
AI has entered laboratories, but it is still far from being able to conduct scientific research independently.

How long does it take for a peptide molecule that helps drugs enter cells to go from initial design to experimental verification?

In the past, it could take a year or even longer just to screen out promising candidates from a massive pool of candidate molecules. At AI4S startup Zhiyuan Shenlan, this process has been compressed to two to three months: The AI model is responsible for design and screening, and the automated experimental platform takes over follow-up experiments. The work that used to rely on researchers' rounds of trial and error is now assigned to algorithms and machines.

Similar changes are taking place in more laboratories. Literature review, data analysis, molecular design, and experiment planning — AI is intervening in links that previously relied heavily on scientists' knowledge and experience. Some entrepreneurs have set even more ambitious goals: to let AI independently raise scientific questions and discover new knowledge.

This direction is called AI for Science (AI4S, artificial intelligence-driven scientific research).

Since the beginning of this year, the financing heat in this track has increased significantly. According to media reports, in the first 10 months of this year, there were more than 80 financing deals in the domestic AI4S track, and the total disclosed financing amount exceeded 200 billion yuan. Newly established startups such as Keli Xinxu and Yuancai Xinjie have successively obtained financing, and capital is pouring into fields including life science and material R&D at an accelerated pace. AI4S startup Deep Principle has launched the scientific research large model Hydrogen-1.0, trying to involve AI in experimental design and new material discovery.

Major tech companies are also accelerating their layout. On November 16, Alibaba released Qwen4-Max, which demonstrated the capability of independently designing new coolant molecules and calling the automated experimental platform to complete synthesis and testing. Previously, Anthropic's Claude Fable 5.2 and Mythos 5.2, and Tencent's Hunyuan Hy4 official version have also successively demonstrated the ability of general large models to solve complex problems in theoretical physics, mathematical research and other fields.

Scientific research has long been regarded as one of the fields that rely most on human judgment. Whether a hypothesis is valid needs to be tested by real experiments; whether a discovery is valuable needs to be judged by scientists. So how much work can AI actually do for scientists? How far is it from independently completing scientific discoveries? And how big can this business be?

01. Scientists are starting to train their own "successors"

Wang Chengzhi, founder of Zhiyuan Shenlan, still remembers an experience during his PhD study.

In order to find an intestinal bacterium that can ferment specific compounds, he first mixed about 10,000 strains of bacteria for fermentation. After confirming that the target bacterium was among them, he kept narrowing down the scope for screening using the dichotomy method, which took more than half a year in total.

He once doubted that he had worked so hard to get admitted to Peking University, so why was he doing these repetitive tasks? Later, when he went abroad for exchanges, he found that some researchers at Harvard and Yale were doing the same things.

This is not a personal predicament. A large amount of time in scientific research is consumed by repetitive labor, which squeezes out the time for scientists to make judgments and create. In 2024, Wang Chengzhi founded Zhiyuan Shenlan, combining AI models with automated experimental platforms, hoping that one day AI can evolve from an "experimenter" to a "scientist".

Apart from repetitive labor, scientific research has two long-standing problems: the cycle is too long, and the threshold is too high.

Zhao Chengcheng, founder of Lightyuan Technology (Hangzhou), has long been engaged in material experimental research before. She found that a scientific research project, from raising questions to clarifying the underlying scientific laws, often takes many years. During the research process, there may be multiple hypotheses and contradictory experimental results, which require scientists to continuously analyze evidence and adjust judgments. The more complex the problem is, the harder it is for individuals to sort things out clearly.

This prompted her to turn to AI4S. She hopes that in the future, AI can participate in the process of raising hypotheses, analyzing evidence, and designing experiments, so that existing research can be continuously connected with new experiments, making every exploration the starting point of the next research, and accelerating the process of discovering scientific laws.

The high threshold is reflected in the gaps between different disciplines.

Erik, co-founder and COO of Keli Xinxu, took a cooperation case with a plant breeding scientific research team from Peking University as an example. The team has profound accumulation in plant science research, but protein design involves another set of professional models and calculation processes. In the past, researchers needed to learn relevant tools additionally, or cooperate with computational biology teams. With the help of the scientific research agent developed by Keli Xinxu, they can describe research goals in natural language, and AI will assist in calling protein design and analysis tools, making interdisciplinary exploration more convenient.

Erik calls this change "scientific research egalitarianism". In his view, the value brought by AI is not only to make scientists who are already good at computing more efficient, but also to allow more researchers with non-AI backgrounds to obtain advanced scientific research computing capabilities. Users of Keli Xinxu's agents come from niche directions such as stomatology, organoids, and plant breeding, where there were almost no dedicated AI tools in the past. Once AI is introduced, the efficiency improvement and "sense of shock" will be much greater.

In fact, using computers and machines to transform scientific research is not new. Machine learning entered fields such as drug R&D and material calculation many years ago, but in the past, algorithms could often only solve a certain type of problem, machines could only perform fixed operations, and complex research processes still needed to be organized by humans.

In Erik's view, this year's warming up comes from the convergence of three forces: the continuous improvement of general large models' reasoning, code and tool calling capabilities enables AI to understand and break down complex research tasks. Professional models in fields such as proteins, molecules, and cells are gradually maturing, and more and more scientific problems have corresponding AI tools. In addition, laboratory automation equipment has also been fully accumulated, which can perform experimental operations and feed back results to AI.

So how much work can scientists already assign to AI?

Wang Chengzhi roughly divides the development process of AI participating in scientific research into three stages: assisted scientific research, guided scientific research, and autonomous scientific research.

In the assisted stage, AI mainly helps scientists process information. In his early years, Wang Chengzhi led a team to use algorithms to connect information such as genes, diseases, drugs and molecules involved in published papers into a huge knowledge graph.

This solves the problem of scattered scientific knowledge. A scientist may spend many years studying a certain gene and find that it is related to a certain disease; another scientist studies a drug and happens to find that it can act on this gene. The connection between the two studies is difficult to discover just by people reading papers. The knowledge graph strings these scattered clues together.

In the guided stage, AI begins to participate in the organization and execution of research tasks.

For example, a life science researcher wants to optimize a certain protein. In the past, he had to look up literature by himself, select structure prediction models and analysis software, run different computing tasks, and then switch tools continuously according to the results. Now the scientific research agent can string these steps together: the researcher puts forward the goal, AI calls the corresponding models and tools, completes part of the calculation, and arranges the next step according to the intermediate results. Erik described that AI is increasingly like a "very strong scientific research partner". At this stage, the research direction and key decisions still need to be checked by humans.

The autonomous stage requires AI to go one step further and independently complete scientific discoveries. This step has not yet been achieved.

Wang Chengzhi used autonomous driving as an analogy. Current AI-enabled scientific research is similar to assisted driving, which can complete relatively fixed and standardized experimental processes, but still requires scientists to issue instructions, check results, and intervene when encountering complex situations. In the future, AI should be like a car that can achieve full self-driving, which can independently find problems, design experiments, and carry out the next round of research based on the results.

Before that, the steering wheel is still in the hands of scientists.

02. What are the difficulties for AI to do scientific research?

Why can't AI take over the steering wheel yet? If scientific research is split into several links including raising hypotheses, designing schemes, experimental verification, and learning from results, we will find that there are still several hurdles for AI to independently complete the whole process.

The starting point of scientific research is to raise a good question, which is precisely what AI is least good at.

Wang Chengzhi believes that AI's ability to summarize existing knowledge has far surpassed that of humans, but its ability to create new knowledge is still "failing to pass". AI is essentially a probability machine, good at finding general laws from massive amounts of data, but it tends to ignore subtle phenomena that deviate from the laws. And scientific discoveries are sometimes hidden in these unexpected results.

An experimental result that does not meet expectations may be an error, or it may be a key node for a major breakthrough. Distinguishing between the two requires scientists' experience and intuition.

Erik calls this judgment "scientific taste". Which questions are worth investing in, and which directions may lead to new discoveries, require scientists to make judgments. Under the condition of limited computing power and experimental resources, the value of studying one good question is far greater than studying ten or twenty meaningless questions.

Even if a good question can be raised, AI often struggles when designing solutions.

Keli Xinxu once cooperated with a life science research group from Peking University, trying to use a scientific research agent based on a general large language model to complete the complex molecular docking task. After the agent gave the preliminary results, the researchers thought it was not good enough and asked it to continue optimization. The agent continuously carried out five or six rounds of adjustments, and each round fed back what improvements it had completed. When the researchers intervened to check, they found that the optimizations claimed by the agent were not real optimizations, and some results even got worse.

In Erik's view, the fact that the scientific research agent based on the general large language model can call tools does not mean that it truly understands the scientific objects processed by the tools. Molecules and proteins are complex three-dimensional structures, and the position, distance and interaction between atoms may all affect the final properties. The general large language model that the agent relies on mainly relies on natural language for reasoning. When three-dimensional information is converted into text and then handed over to professional tools for execution, a lot of key information may be lost in the "translation process".

When it comes to the experimental verification link, the problem becomes that AI generates results too fast to keep up with verification.

Wang Chengzhi observed that some AI pharmaceutical companies can design tens of thousands of candidate molecules a day, but they do not have sufficient experimental capabilities to verify whether they are useful.

In the field of AI-assisted scientific research, the work of predicting, screening and designing molecules with computers is usually called "dry experiment". But to know whether these molecules can play the expected role, we have to carry out synthesis, testing and verification in the laboratory, which is the "wet experiment". The efficiency gap between the two is called the "dry-wet gap".

Erik did the math: AI can generate 1000 candidate schemes quickly, but verifying all 1000 schemes is very slow, tiring and expensive. If the model can improve the accuracy of judgment and find effective results by generating only 10 schemes, the value will be greater. What the industry needs is not only stronger generation capability, but also reduction of invalid verification.

Letting robots take over experiments is one path. At present, some automated laboratories can already complete batch experiments under specific processes, but the equipment, reagents and operating conditions involved in scientific research often change. Engineers need to set up processes and adjust parameters in advance, and manual intervention is also required when abnormal situations occur.

Wang Chengzhi believes that the next step is to let AI learn to control equipment independently. Scientists only give an experimental goal, and it can perceive the environment through information such as vision and touch, and learn to operate instruments in successive failures just like a child learning to walk. He judged that important breakthroughs in this area may emerge in the next few years.

The last link of the scientific research cycle is to learn from experimental results, which is inseparable from scientific data.

Wang Chengzhi told *Focus One* that he once led a team to clean up a batch of public single-cell sequencing data, and finally found that about 95% of the data was not suitable for direct model training. These data come from different scientific research projects, and the experimental conditions, measurement methods and data standards adopted by various laboratories are not consistent. Researchers only care about whether the data can support their own conclusions, and do not consider whether the data can be used by AI.

"Garbage in, garbage out." In Wang Chengzhi's view, scientific models require real and standardized data, preferably data that different laboratories can get consistent results with according to the same process. If the training materials themselves have problems, no matter how accurate the algorithm is, it cannot guarantee the reliability of the results.

Erik pays more attention to the utilization of existing scientific data. He believes that a large number of public papers, patents and scientific research data have not been fully sorted and utilized. The AI tools developed by Keli Xinxu are used to "denoise and clean" these data, converting scattered scientific research information into structured data that can be used for model training and scientific analysis. In contrast, experimental feedback that can verify whether AI predictions are correct is much scarcer.

Around different judgments, startups find entry points from different directions.

Keli Xinxu cuts in from the scientific foundational model, first giving AI a "brain" that understands science better. Keli Xinxu develops a cross-modal life science foundational model starting from the pre-training stage, allowing AI to directly learn and understand different types of scientific information, including protein structures, DNA and RNA sequences, and then use this information for reasoning and analysis to reduce invalid calculations and experiments.

Zhiyuan Shenlan devotes more energy to experimental capabilities, building more "hands". In its laboratory, AI participates in scheme design, automated equipment executes tasks, and the generated data is fed back to the model. According to Wang Chengzhi, the platform can complete about 30,000 experiments a day, and these data are of unified standards and can be directly used for model iteration.

Some other enterprises cut in from professional computing and simulation to reduce dependence on experiments. Ye Yitian, managing partner of Advantage Star Launchpad Scientist Fund, mentioned that in fields such as materials, some companies use professional algorithms to simulate material structures and predict performance, and then turn these capabilities into models or software to provide to industrial customers. Such companies do not need to build their own laboratories, and it is easier for them to access the existing R&D processes of enterprises.

These different technical routes are all trying to make AI take on more scientific research work.

03. More industries are squeezing into laboratories

Although AI is still far from doing scientific research independently, in some industries, enterprises have begun to introduce it into their R&D processes. Which industry AI4S develops faster in roughly depends on two types of conditions.

One is the technical condition: whether the results given by AI can be verified quickly and clearly. The other is the industrial condition: whether the industry is willing to introduce external technologies, how high the threshold for new achievements to enter production is, and whether core data and R&D processes are controlled by the enterprise itself. There are very few industries that have both types of conditions at the same time. Those that are currently leading often only meet one of them.

The pharmaceutical industry is the field with the most mature industrial conditions.

New drug R&D is characterized by high investment, long cycle and high failure rate. It may take more than ten years from drug molecular screening to clinical trials. If the early R&D time can be shortened and invalid experiments can be reduced, there is sufficient reason to try AI4S.

The mature R&D division of labor in the pharmaceutical industry also leaves room for AI companies. Pharmaceutical companies have long purchased professional software, and are also accustomed to outsourcing part of their R&D work to third-party institutions. AI companies can cut in along the existing industrial chain without having to cultivate customers' payment habits from scratch.

But early entry also means more intense competition