AI Chemical Synthesis Nears a Decade-Experienced Chemist, but White-Box Is the Commercial Prerequisite
AI chemical synthesis capability is close to that of a chemist with ten years of experience and can propose five to six routes, but serious scientific software must be white-box to build trust; AI for Science penetration is less than one in a thousand, yet the trend is inevitable.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Serious scientific software cannot rely solely on black boxes
Black-box models occasionally hallucinate or have bugs, with severe and inexplicable consequences. The biggest problem is they cannot be tuned—if you change A, B, C, and D change unpredictably, and you don't know which one caused it. White-box models make the system explainable, verifiable, and tunable, enabling customer trust. In ChemAble's system, all conclusions from large or black-box models must be explained, verified, and constrained by white-box models to eliminate hallucinations.
— Xia NingAI synthesis capability has reached the level of a chemist with ten years of experience
Internal and external evaluations at ChemAble show that current AI chemical synthesis capability roughly matches a chemist with ten years of work experience. Routes a chemist can think of, AI can also think of, and AI can typically think of five to six different strategies, while humans usually have only one or two. For more difficult process route design, AI can essentially match human thinking. When customers test with real cases, routes that take one to two months to design can be generated by AI in five minutes, along with alternatives.
— Xia NingThe bottleneck in the DMTA cycle is the synthesis step
Drug discovery is essentially multiple DMTA cycles (design-make-test-analyze), typically seven or eight rounds. In the design phase, hundreds, thousands, or even tens of thousands of molecules can be designed in a day; in the testing phase, high-throughput screening can test hundreds at once. But the synthesis step is stuck—a chemist synthesizes about three to five molecules per month. The entire R&D efficiency is bottlenecked at synthesis. Only by increasing synthesis efficiency two-, three-, or even ten- to twenty-fold can the throughput of new drugs and materials multiply.
— Xia NingThe biggest hurdle for pure AI founders is industry know-how
When collaborating with people from pure AI backgrounds, the biggest issue is their misunderstanding of industry problems—sometimes even getting them wrong—yet they don't voice it, continue down the wrong path, and blame the data when results are poor. In vertical domains, without industry understanding, you can't define problems clearly, let alone handle things that can't be explicitly expressed in language. They cannot build a credible white-box system. Industry know-how is the core moat for AI for Science startups.
— Xia NingAI for Science data volumes are orders of magnitude smaller than the internet's
Large models rely on over forty years of accumulated public web text data, while any scientific field—even with all its papers—has far less, by several orders of magnitude; some fields have only hundreds or thousands of data points. Thus, the biggest shortage in AI for Science is data volume. For retrosynthesis, public literature and patent data contribute the most because they teach retrosynthetic thinking, which high-throughput experiments cannot solve.
— Xia NingAgent calls require intelligent scheduling; cost-saving is an engineering necessity
The biggest difference between AI and previous software is that every interaction and call incurs cost, and customers care deeply about this. ChemAble does significant engineering work to help customers save money: they cannot arbitrarily call foundation models; they only call appropriate models when necessary, using minimal cost. Internally, problems are classified—simple ones are solved with small or locally deployed models, while critical ones call advanced models, reserving the most expensive resources for the most core problems.
— Xia NingChina's biomedicine may become the world's pharmaceutical factory
China's biomedical supply side is growing rapidly. License-out deals drive R&D efficiency competition. Demand is overseas, supply is domestic—similar to the precision manufacturing stage of the Apple supply chain. In the coming years, it's not impossible that, like manufacturing, a 'world pharmaceutical factory' model will emerge, accounting for the vast majority. But India is also growing fast; ChemAble added over thirty Indian customers this year, with growth not slower than China's.
— Xia NingInvestors most often misunderstand two points
First, investors think the market space must be large enough only if you do innovative drugs or new materials. But Xia Ning believes that's a result, not a cause—only by solving tool efficiency can you gain multiplicative advantages in drug and material R&D; otherwise, it's just storytelling. Second, investors see that past tool software markets were small, and they can't imagine that the agent model, by completing the entire workflow, delivers an overall result, so the market space is underestimated.
— Xia NingIn their own words · checked verbatim
The current AI capability in chemical synthesis is roughly equivalent to the level of a chemist with ten years of work experience.
现在的化学合成的 这个AI的能力 大致相当于十年工作经验的化学家的水平
Xia Ning37:29
AI can typically think of five to six different strategies on average—that surpasses humans, who usually have only one or two.
AI能够一般能平均想到五到六种不同的策略 这个是超过人的 人一般就是一到两种策略
Xia Ning38:18
A chemist synthesizes about three to five molecules per month; that's the common efficiency in the industry.
一个化学家 在一个月的时间 大概合成三到五个分子 这是行业内普遍的一个效率
Xia Ning41:19
I can't just arbitrarily call the foundation model; I only call the appropriate model when necessary, using the minimal cost.
我不能说就是随意的调这个基础模型 我只有在必要的时候调合适的模型 用最小的成本去调
Xia Ning52:28
AI for Science is still very early; its industry penetration is even less than one in a thousand, but I think its trend is almost inevitable.
AI for Science 还在非常早期 它的行业渗透率 甚至不到千分之一 但我觉得它的趋势 是几乎是必然的
Xia Ning1:12:43
Figures
| AI chemical synthesis capability equivalent to chemist's years of experience | ten years | 37:29 |
| Number of synthesis strategies AI can typically think of | five to six | 38:18 |
| Number of synthesis strategies a chemist typically thinks of | one to two | 38:18 |
| Molecules a chemist synthesizes per month | three to five | 41:19 |
| AI for Science industry penetration rate | less than one in a thousand | 1:12:43 |
| New customers ChemAble added in India | over thirty | 1:01:40 |
Glossary
- DMTA cycle
- The core drug discovery loop: Design, Make, Test, Analyze, iterated repeatedly.
- White-box model
- An explainable, verifiable, and tunable model, as opposed to a black box, used to build customer trust.
- Retrosynthesis
- The method of working backward from a target molecule to derive a synthesis route; AI can quickly propose multiple viable routes.
How to listen
AI for Science entrepreneurs, R&D managers in biomedicine, investors focused on AI deployment, and engineers curious about the current state of intelligent chemical synthesis.
The opening personal growth story (0:00-6:25) can be skipped without missing core judgments.