Article · AI translation
A dual drive of models and pipelines
By Luyu, reporting from Aofei Temple
QbitAI | WeChat public account QbitAI
By Luyu, reporting from Aofei Temple
QbitAI | WeChat public account QbitAI
The labels for leading players in the AI world have changed again and again: the four AI dragons, the six large-model tigers…
This time, CCTV itself has selected the industry's next batch of star companies.
After 1,200 participating companies passed through successive rounds of selection, China's Ten Outstanding AI Startups have been announced.
The list spans robotics, intelligent manufacturing and life sciences, but only one AI drug discovery company made it, and it deserves a closer look today.
It works in the currently booming field of AI biopharmaceuticals.
Earlier, Moderna and Merck jointly announced positive results in treating melanoma, known as the “king of skin cancers,” with a personalized mRNA cancer vaccine developed with AI involvement.
The news is encouraging, but it also raises a question: how far has domestic AI drug discovery come?
We may find some answers by looking at the only AI drug discovery company on CCTV's list of 10.
Founded just 3 years ago, the company was incubated by Tsinghua AIR, founded by academician Zhang Yaqin, with Professor Nie Zaiqing serving as chief scientist.
It is Shuimu Molecule, among the AI drug discovery players moving fastest and most comprehensively across various AI modules.
Shuimu Molecule aims to tackle a particularly difficult problem in drug research and development:
How can AI truly make use of a drug research expert's decades of accumulated experience?
After much thought, the veterans at Shuimu Molecule arrived at this answer: conversational, AI-native design.
Since 2023, they have continued working to connect natural language with biological language and bring those capabilities into ChatDD, a conversational drug research and development platform.
Going further, Shuimu Molecule has organized research methods described by experts in conversation into skill packages that AI agents can call at any time.
At the same time, it is advancing its own innovative drug pipelines for autoimmune diseases and tumors, using the platform in its research and development decisions, molecular design and project operations.
Behind this multipronged approach lies a straightforward judgment shared by Shuimu Molecule and the pharmaceutical industry as a whole:
When biopharmaceuticals meet AI, natural-language interaction brings experts' experience and judgment into closer combination with large models' efficient knowledge reasoning and all-atom, end-to-end molecular design capabilities.
The exceptional research and development efficiency brought by this human-machine collaboration must ultimately translate into real clinical progress and higher success rates for innovative Chinese drugs.
Four generations of shifts in drug research and development
Shuimu Molecule COO Xing Jie mentioned on the program a widely cited rule of three “10s” in the pharmaceutical industry.
Developing a new drug takes an average of 10 years and an investment of more than US$1 billion, yet the success rate is below 10%.
The challenges go far beyond finding a molecule that can bind to a target.
How an innovative drug is distributed after entering the human body, whether it achieves the expected effect, whether it has unacceptable side effects, and which patients may benefit are among the questions that require people to build a complete chain of evidence.
Frankly, roughly half of the “old-fashioned” drug development process depends on judgment based on experience, which is difficult to replicate or feed directly into AI training.
In Nie Zaiqing's view, this is where a key gap has arisen in some past AI drug research and development projects.
Models can predict structures, generate molecules and rank candidates, but researchers may not fully understand the basis for the ranking and may struggle to convey their concerns to the models. Computational results can be handed over for experimental validation, yet experts and algorithms lack a channel for sustained collaboration.
Shuimu Molecule therefore proposed dividing drug research and development into four stages:
TMDD → CADD → AIDD → ChatDD.
TMDD (traditional methods of drug design) is straightforward: it represents the earliest drug research and development process, which relied on accumulated experimental findings and expert judgment.
Computers then began to assist, taking on some calculations and simulations. This is CADD (computer-aided drug design).
Later, AIDD (AI-aided drug design) began using AI for tasks such as property prediction and molecule generation. But experts' experience and intuition were hard to integrate and interact effectively with AI models in the era of smaller models.
ChatDD (Chat-based drug design), meanwhile, seeks to enable experts to participate in task planning, execution and the revision of results through natural-language dialogue.
In this way, it organically combines drug research experts' experience and intuition with AI agents' capabilities in high-throughput literature review, efficient computation and reasoning, and high-quality molecular design and screening, creating a collaborative research and development model in which humans and machines complement each other.
Experts pose questions and set task goals; a large model breaks them down and calls tools, then returns the results to the experts. Through multiple rounds of interaction, expert experience can efficiently inform every step of AI drug discovery.
This is crucial at a time when machines still cannot grasp sufficiently complete information. Conversational interaction is precisely what makes a connection between experts and AI possible for the first time.
In fact, this division into generations was also quite forward-looking.
Shuimu Molecule proposed ChatDD in 2023. Today, GPT-Rosalind and Claude Science overseas have also begun exploring similar forms of interaction.
The scope of industry competition is evolving from the predictive ability of individual models to the organization of entire research workflows.
Put more simply, it is a race to become the AgentOS entry point for drug research and development.
Next, taking Shuimu Molecule as an example, let's look at how it brings drug research experts' experience into the work of AI scientists.
A dual drive of platform and pipelines
To put AI to work, the first step is to see where this experience actually resides.
Some of it is already recorded in papers and databases, ready for AI to use; some lies in experts' judgments about current projects and needs to be captured through everyday interaction; and some can be distilled into reusable workflows.
Shuimu Molecule is trying to feed these forms of experience into models and specific tasks, and to let them continually shape what happens next.
In concrete terms, this is an interactive system that extends upward from the foundation model.
Let knowledge from papers inform the next design
Take a simple example: suppose a researcher wants to optimize an antibody through mutations. The expert states a functional goal, but the model ultimately has to handle amino acid sequences and spatial structures.
There is an obvious language gap between the two, so the first challenge is to incorporate relevant knowledge into the foundation model.
BioMedGPT, which Shuimu Molecule and Tsinghua AIR began developing together in 2023, first explored aligning natural language with biological data such as molecules and proteins.
Researchers can ask questions about biological data they input, linking knowledge expressed in words to specific research subjects.
BioMedGPT-10B is also the world's first open-source, commercially usable multimodal biomedical large model with 10 billion parameters, ranking in the industry's top tier on benchmark tests.
MutaPLM makes this connection more concrete.
This research, accepted by NeurIPS 2024, links written descriptions of mutation effects to protein mutations. It explores both explaining what changes a mutation might cause and proposing candidate mutations based on goals described by researchers.
In short, the former seeks to understand “what might happen if this position is changed,” while the latter asks “where could we make a change to achieve this desired result?”
Mutation optimization must also account for structure. The recently released StructEvo brings structural information into protein mutation optimization so that candidate searches take spatial constraints into account.
According to the team, the research has been accepted by NeurIPS 2026.
Behind all this is a clear technical judgment by Shuimu Molecule: protein design has long faced a trade-off in which improving one aspect can undermine another.
Changing a few amino acids may strengthen binding but disrupt a protein's proper folding; increasing structural rigidity may improve stability while sacrificing functional fit.
The team has therefore developed corresponding models for specific design tasks. CodeFP, accepted by ICML 2026, focuses specifically on the joint design of sequences and structures for functional proteins.
AbGaze further explores end-to-end, all-atom de novo antibody design and uses this geometric information to generate sequences and structures.
…
Thus, from extracting knowledge from the literature to turning functional goals into a basis for model analysis and design, the research findings experts have accumulated over the years gain a way to inform the next round of development.
The resulting AgentKG knowledge graph is used to organize domain knowledge and its connections, making every answer traceable and attributable.
Experts use prompts to add goals, constraints, knowledge gained from experience and questions. Agents then plan tasks, call tools and return the results.
After receiving the results, experts ask follow-up questions or adjust the conditions, changing the next round of tasks accordingly. Through successive rounds of specific exchanges, experience begins to play a deep role in the computational process.
Model training absorbs organized knowledge and data, while dialogue supplies context for the current task and thus influences its direction: as experts use AI, AI also uses expert judgment to organize its work.
Turn one-time guidance into a reusable method
Then comes a more practical question: once an expert has guided AI through a task, must they start from scratch the next time a similar task arises?
This question has pushed Shuimu Molecule's work beyond models and dialogue toward skill packages.
Expert experience is also reflected in research methods: which evidence to check first, when to conduct additional analysis, how to set screening criteria and which steps require human review.
These processes, when they can be clearly expressed and validated, can also potentially be organized into skill packages that can be called repeatedly.
The first batch of OpenBioMed skill packages, released this year, already includes 45 specialized skills across 5 categories: biochemistry and drug research and development, protein analysis and design, single-cell omics analysis, data retrieval, and tool calling.
Among them, Skill Creator extends expert participation into workflow creation.
Researchers explain task intent, inputs, outputs and constraints through dialogue. The system helps organize the process, which is then validated and revised using real data to form a skill suited to the task.
In this way, a single instance of expert guidance can leave behind a method that remains useful.
Knowledge distilled from papers, current goals expressed in prompts, and validated skill packages each play a different but complementary role, together enabling experts to participate step by step in the work of AI scientists.
Test the collaboration on real tasks
Ultimately, this approach must be tested on real tasks.
At the application level, this means ChatDD for businesses and Shuimu Zhiliao for individual researchers.
These correspond to 2 product models: the former works from the top down, with companies purchasing it and then encouraging employees to use it; the latter works from the bottom up and is better suited to researchers and small teams.
First, ChatDD: Shuimu Molecule CEO Wang Zhuozhi describes its 3 main modules as Insight, Discovery and Trial.
Insight first helps experts answer questions before starting a project: Is a target worth pursuing? Which patients could a resulting drug treat?
This requires understanding disease mechanisms, unmet clinical needs, existing research evidence, competing projects and the patent landscape at the same time. ChatDD divides the relevant work among multiple specialized agents, which collect and organize information to help experts form judgments.
At the Discovery stage, the task shifts to designing and virtually screening molecules.
Experts set design goals, the platform organizes computation and candidate screening, and researchers then use the results to inform their judgments and experimental plans. This is also where Dr. Wang Zhuozhi believes revolutionary change is emerging.
Trial extends support to clinical-related work, with the aim of improving the efficiency and success rate of clinical development.
Simply put, ChatDD's most important contribution is enabling life scientists to make full use of AI tools through dialogue.
Simply put, ChatDD's most important contribution is enabling life scientists to make full use of AI tools through dialogue.
In the past, life scientists often had to learn programming or communicate repeatedly with computing teams to complete these tasks. Now everyone can have a team of AI scientists that truly understands drug development, reducing communication costs and greatly improving the efficiency of innovative drug research and development.
Before joining Pharmolix, chief biopharmaceutical scientist Pan Qi faced two opportunities: joining a company that used AI tools within an established R&D system, or joining an AI-native company.
As a biopharmaceutical expert, she chose the latter, Pharmolix. Her reasoning was:
For an AI-native company, AI is more than a tool or an aid; it could also be a driving force that transforms the entire drug development process.
Beyond project initiation and drug candidate design, ChatDD can also deliver substantial efficiency gains in the downstream preclinical and clinical development stages, precisely the stages that are costly and time-consuming.
For an AI-native company, AI is more than a tool or an aid; it could also be a driving force that transforms the entire drug development process.
Beyond project initiation and drug candidate design, ChatDD can also deliver substantial efficiency gains in the downstream preclinical and clinical development stages, precisely the stages that are costly and time-consuming.
Pharmolix has already applied this capability to its own R&D projects.
Take Pharmolix's first in-house pipeline as an example: the project targets bispecific antibodies in the field of autoimmune diseases. According to Pan Qi, when the team initially chose its target and research direction, it used ChatDD-Insight to analyze the disease area and competitive landscape, as well as preclinical, clinical and translational research evidence.
After setting the direction, the team continued to use ChatDD-Discovery for molecule discovery and optimization.
From initiating the project based on target biology and clinical evidence, through molecule design and virtual screening, to wet-lab validation and confirmation of a preclinical candidate (PCC), the project took just seven months.
Compared with the 12–24 months needed to obtain a preclinical drug candidate using traditional methods, ChatDD has already demonstrated the effectiveness of integrated AI in early-stage R&D.
At the same time, project data also provided more direct feedback to the ChatDD platform, such as whether computational recommendations held up in experiments and how to adjust them where they did not.
These questions continually return to the same R&D team, and experts' experience feeds into the next round of design through this process of judgment, execution and feedback.
This is precisely the value of having the platform and pipeline drive each other:
In-house projects allow drug development experts to use the platform continuously and give the algorithm team a chance to track a task's subsequent results over the long term, continually improving the tools and processes.
Overall, Pharmolix's native AI capabilities form a four-layer pyramid.
Although the entry points differ, they share the same goal: to make AI easier for researchers to use and to bring their professional judgment more fully into the entire AI workflow.
Pharmolix is gradually putting this into practice across four dimensions: AI models, the agent platform and Skills, products, and its pipeline.
While strengthening its internal foundations, it has also made progress externally.
ChatDD has now also taken root with a number of external customers, including the following types of collaboration:
Large domestic pharmaceutical companies: taking Fosun PharmAID as an example, Pharmolix participates in jointly creating an R&D agent platform, embedding AI capabilities into enterprise workflows;
Biotechnology companies: the platform serves leaner R&D teams, helping build an end-to-end, agent-native innovative drug development process so that team members can devote more energy to assessing research directions;
CROs: collaboration extends further into reshaping processes related to CMC operations, clinical trial protocols, clinical study reports and regulatory submissions, using AI to empower the CRO ecosystem and expand into pharmaceutical industry services;
Leading multinational pharmaceutical companies: strategic collaborations involve clinical operations, clinical regulatory submissions and post-marketing clinical work.
Commercialization and an open-source philosophy coexist
Seen this way, Pharmolix is actually more like China's Isomorphic Labs.
Both are AI-native companies with closed loops linking computational and wet-lab work, and both pursue a parallel strategy of platform services and in-house pipelines.
Their similarities are also reflected in their respective technical backgrounds:
Both were incubated by leading AI institutions (Alphabet/Tsinghua AIR), are led by AI algorithm scientists, and have deep foundations in AI research.
By contrast, Pharmolix is more inclined toward large-molecule pipelines and open-sourcing its technology.
On the commercialization side, the platform and pipeline provide two growth curves that advance in turn, ensuring the current pace of monetization while leaving room for a higher ceiling on value creation.
2026 will be a phase of commercial expansion for Pharmolix, with standardized products as its main offering and customized projects as a supplement. Xing Jie expects this year's revenue to be 3 to 4 times last year's, and also expects the pipeline to enter the development stage and business expansion to take place next year.
Admittedly, a dual-track approach driven by both platform and pipeline is not uncommon in the AI drug discovery industry.
Pharmolix's specific choice is to combine conversational agents with the development of large-molecule drugs such as antibodies, allowing AI to participate deeply in sequence design, structural analysis and molecule optimization.
This also means AI can participate more deeply in drug development and deliver its value more visibly, further expanding the advantages of the AI-native approach.
Supporting all of this is a dream team spanning AI and pharmaceuticals.
Zhang Yaqin, Nie Zaiqing and Ma Weiying bring experience in AI research and the industrial application of technology.
Wang Zhuozhi previously served as a vice president at WuXi Biologics, was deeply involved in the R&D of more than 300 innovative drug projects, and has experience working on global collaborations with Bayer; Pan Qi has accumulated years of experience in drug development and AI applications at companies including Genentech, Boehringer Ingelheim, Legend Biotech and Pathos AI.
Xing Jie previously served as a vice president and CIO at WuXi AppTec and has long been involved in the pharmaceutical industry's digital and intelligent transformation, investment and financing, industry ecosystem operations and commercialization.
When the value of these experiences comes together at Pharmolix, its AI-native approach truly takes root.
This was also my strongest impression after speaking with Pharmolix: the team has both technical confidence and the determination to turn its ideals into reality.
Whether discussing models, products or the pipeline, their first principle remains the hope that technological progress will ultimately become effective medicines that patients can use.
Starting from the research built up at Tsinghua AIR and reaching the stage of CCTV's Ten Steeds, Pharmolix has, in three years, turned its initial judgment into a product and brought that product into real-world R&D.
Its vision is for every drug development expert to have a team of AI agents that truly understands drugs, to make innovative medicines no longer expensive, and to move China's new drugs from following to leading.
But more worth looking forward to than embracing the next GPT moment in AI drug development is what kinds of innovative medicines these changes can ultimately bring to patients.
The laboratory is the starting point; patients are the ultimate destination of drug development.
Official website: https://pharmolix.com