Ask AI · Why is the AI pharmaceutical industry starting to reevaluate the value of late-stage clinical trials?

On March 18th, AI pharmaceutical company Deep Intelligent Yao announced the completion of its Series D funding, nearly $200 million. Through multiple rounds of intensive financing within three months, they rapidly secured over one billion RMB, a rare pace and density in the industry.

Currently, the capital market coexists with enthusiasm and rationality. The certainty of returns has replaced technological narratives as the focus of capital attention. AI pharma has emerged as a new force, with companies like Jing Tai Technology and Yingxi Intelligent going public on the Hong Kong Stock Exchange, and Lilly partnering with NVIDIA to build “AI pharmaceutical factories,” fueling the sector’s fervor.

Founded in 2017, Deep Intelligent Yao has maintained a low profile compared to many peers under the spotlight. Its niche—clinical development and evidence generation—was once considered relatively obscure and the most challenging part of AI drug development.

Drug R&D can be roughly divided into drug discovery, preclinical, clinical trials, and registration. Over the years, the most popular narratives focused on the front end—using AI to predict protein structures, design molecules, and screen targets.

Deep Intelligent Yao, however, concentrates on the back end, building an AI-driven integrated delivery system around the clinical trial evidence chain—from clinical strategy, protocol design, and research center execution to EDC and data management, clinical/statistical programming, medical writing, pharmacovigilance, and registration support. They have gradually established a stable delivery capability through numerous real projects.

This system is supported not by a single tool but by a human-like multi-agent intelligence system. Through task decomposition, role collaboration, feedback verification, and recursive self-evolution, they organize highly complex, cross-departmental, tightly constrained workflows in clinical R&D, making it possible to systematize processes that heavily depend on human experts.

It is this intelligent agent system that makes Deep Intelligent Yao a rare target in the hot sector, gaining well-deserved attention.

For the AI pharma industry, this signifies a shift of focus from front-end to back-end and marks a significant step toward truly creating human welfare. And this step has taken nearly a decade.

Going Against the Trend

The first impact of AI on the pharmaceutical industry predates the birth of Transformer models.

In the same year AlphaGo defeated Lee Sedol, Google launched GNMT (Google Neural Machine Translation) for machine translation. Initially developed to improve online translation services, it unexpectedly solved a thorny problem in pharma—medical document translation.

In 2016, Google released the GNMT model, significantly improving translation accuracy.

From clinical trial protocols, informed consent forms, investigator manuals, to case report forms, clinical research reports, and registration materials, the documentation needed to register a new drug can fill several small trucks. These require absolute terminological precision and genuine understanding of research design, indications, endpoints, statistical hypotheses, and regulatory language.

For pharma, the challenge has never been just “accurately translating a sentence,” but aligning medical, statistical, operational, and regulatory logic across the entire chain. Any deviation in terminology, endpoints, hypotheses, or data standards can be amplified downstream, potentially leading to hundreds of millions in R&D costs lost, with minimal tolerance for errors.

GNMT’s emergence turned translation from a sleepy task into an entry point for AI in drug development.

To some extent, drug R&D is a highly knowledge-driven industry that ultimately relies on text and data. Extending from translation, the core challenge is to solve the “language expression” pain point across the entire drug development process.

A new drug, from lab to clinical trials to approval, is not just a molecule but a comprehensive set of evidence—clinical trial protocols, investigator manuals, informed consent forms, statistical analysis plans, clinical study reports, registration documents, etc. Essentially, a production line of text, data, and responsibility.

Many drugs fail just before market launch not because of insufficient clinical value but because of issues in this “production line”—information loss or logical gaps during the “languageization” process, preventing the transformation of R&D data into a scientifically recognized evidence chain, thus burying or misinterpreting the drug’s scientific value.

Deep Intelligent Yao early concluded: In pharma, understanding is harder than generation; collaboration-driven generation is more important. Only by establishing stable understanding, verification, and collaboration capabilities in high-constraint scenarios can one qualify for creation and decision-making.

Therefore, the company’s initial development plan was clear: start with translation, then extend to medical writing, data management, statistical programming, and clinical operations, ultimately forming a full-process production line.

The logic is that translation has clear reference standards and correctness criteria, making it the most direct “hard currency” to verify model understanding. From there, gradually incorporate planning, reasoning, and execution capabilities.

Looking back, this path was highly forward-looking, but at the time, it seemed like going against the wind.

On one hand, model capabilities were immature; GNMT still relied on RNN-based sequential computation, leading to low efficiency. Although Transformer models later solved this, pretraining became dominant, it did not change AI’s fundamental role as an auxiliary tool—far from replacing professionals.

On the other hand, pharma is a field with extremely high requirements for “know-how” reserves.

Taking the drafting of core clinical trial protocols as an example, building a skeleton based on vast literature and past data is just the first step. It then requires deep involvement from teams of experts in medicine, statistics, and clinical pharmacology.

A protocol spans multiple departments—medical, clinical pharmacology, statistics, programming, data management, pharmacovigilance—and even minor modifications can cause major upheavals. Deep Intelligent Yao’s early work was not about “one-click automation” but about repeatedly verifying and refining in real projects.

Meanwhile, the groundbreaking advent of AlphaFold made the industry realize AI’s destructive power.

AlphaFold is a neural network architecture designed specifically for protein structure prediction, solving the core “target structure analysis” problem in drug discovery, reducing years of experimental work to minutes.

AlphaFold predicts protein structures

Since then, “AI-designed molecules” has become a mainstream direction, with major pharma companies and startups flocking into the space, and front-end drug discovery becoming a hotbed of capital.

Deep Intelligent Yao, however, remains outside the storm, quietly focusing on the back-end clinical design, repeatedly refining and accumulating know-how in real projects, which feeds back into core algorithms and systems.

Then, a crisis swept through the industry, pushing Deep Intelligent Yao and similar companies focusing on clinical design into the spotlight.

Reevaluating the Value

In 2023, the first AI-designed new drugs faced collective failure in clinical trials.

First, European AI pharma unicorn BenevolentAI’s core pipeline BEN-2293 failed in Phase II trials, causing a sharp stock decline and forcing layoffs; later, industry leader Exscientia shut down its early oncology pipeline EXS-21546 (A2A receptor antagonist).

BenevolentAI announced the failure of BEN-2293 in Phase IIa

The consecutive setbacks shattered the myth of “AI generating new drugs with one click.”

Both industry and capital now realize that the path from molecule design to market is longer than expected.

From clinical trial design and enrollment to data quality, statistical interpretation, and finally regulatory communication and submission, missing a step can ruin everything—a “boss fight” with no retreat. The final deliverables to regulators are a comprehensive set of interpretable texts, data, and evidence chains.

Capital is no longer just paying for computing power and molecule counts but is scrutinizing the effectiveness data of pipelines in the clinical stage. The “heavy molecule, light clinical” route in AI pharma is being corrected.

Companies like Deep Intelligent Yao, focusing on the back end of drug development, have thus entered the capital’s view.

By this time, Deep Intelligent Yao had already weathered the technical build-up and breakthrough phases, becoming self-sufficient and independent of funding.

Following their initial plan, they expanded capabilities from text and writing to core clinical CRO functions—data management, clinical/statistical programming, site operations, and registration support—gradually forming a full-process clinical trial system.

Their business has extended to China, Japan, the US, Australia, Singapore, and Southeast Asia, especially establishing a strong PI and research center network in Japan, creating local execution advantages.

Deep Intelligent Yao’s case studies on their official website

Technologically, they did not stop at “following the latest models.” Instead, they reconstructed their system based on the high constraints and low tolerance for error in pharma, focusing on controllability and collaboration.

The most impactful upgrade occurred during the 2019 iteration from Model 2.0 to 3.0.

At that time, large language models (LLMs) were gaining popularity, but Deep Intelligent Yao recognized early that the most fatal flaw of LLMs was not their “human-like” writing but their tendency to “rationalize like humans.”

LLMs are trained on massive data to predict the next word, but most of this data comes from past experiences, leading to outputs that conform to experience-based logic rather than absolute truth.

In pharma, this illusion is a matter of bottom-line safety, not just experience.

A fabricated quote or data can pollute the entire scientific basis of R&D; a fabricated safety report can cause serious harm or death to subjects; a perfectly packaged hallucination can wipe out a company’s decade-long, billions-dollar investment instantly.

As one of the few pioneers in the field, Deep Intelligent Yao’s realization came 4-5 years ahead of industry consensus on large model development. This meant there were no mature industry standards or external tools to rely on—they had to rebuild solutions from the ground up.

During this phase, they shifted focus from “how to build a better model” to “how to develop a more controllable, collaborative, and capable system for clinical R&D workflows.”

Deep Intelligent Yao’s approach is not to keep betting on a larger monolithic model but to “disassemble the brain.”

“Disassembling the brain” does not mean simply splitting a model into multiple modules but breaking down complex tasks into many capable, well-defined intelligent agents, each responsible for decision-making, planning, retrieval, writing, programming, review, and verification. These agents communicate via neural-like connections, cross-check, and balance each other.

The core of this architecture is not just multi-role collaboration but a human-brain-like working mode. Like neural networks, the system does not produce a linear answer but continuously reviews upstream, corrects intermediate results, and reorganizes task paths during execution. If a step violates constraints, it triggers new reasoning and validation until the output approaches usability.

In other words, it does not just generate answers once but possesses recursive reflection, correction, and self-evolution capabilities. It is not a one-time answer generator but a work system that repeatedly thinks, revises, and approaches the optimal solution.

This approach is strikingly similar to the current buzz around Agent concepts.

In 2023, Microsoft invited Deep Intelligent Yao to a closed developer conference to introduce the Agent framework.

For Deep Intelligent Yao, this was more like an “external naming”: their internal small-model collaboration system evolved into a true multi-agent cooperation system.

Their human-brain-like multi-agent system gradually took shape. It is not just a workflow engine connecting tasks but a “bionic brain” composed of numerous high-precision atomic agents, capable of organizing work around goals and continuously reflecting, verifying, and recursively evolving during execution.

Deep Intelligent Yao’s “Multi-Agent Collaboration System”

This is where their edge becomes fully apparent.

Moving the human role upward

In 2025, a collaboration with Japan’s innovative drug company Immunorock thrust Deep Intelligent Yao into the spotlight.

As one of the top three global pharmaceutical markets, Japan is renowned not only for its drug R&D capabilities but also for its extremely strict regulatory standards by the PMDA. The clinical trial protocols supported by Deep Intelligent Yao achieved “zero rework” and were approved by PMDA on the first attempt.

In this collaboration, their human-brain multi-agent system participated in the entire process—from information integration and pathway planning to digital twin simulation. Different agents analyzed aspects such as endpoint setting, inclusion/exclusion criteria, sample size, execution pathways, data structures, and regulatory constraints, repeatedly cross-verifying through feedback loops.

This recursive reasoning identified potential design flaws that could increase dropout rates, allowing the team to make corrections before finalizing the plan. It did not speed up “writing” per se but moved many issues that would have been exposed during on-site execution to the design phase for resolution.

Note that, due to regulatory and ethical responsibilities, all key deliverables are ultimately reviewed, signed off, and approved by professionals.

Immunorock is just one of many cases demonstrating that long considered highly dependent on experience and human collaboration, clinical trial planning can now be systematized, recursively verified, and scaled.

Traditional clinical R&D is essentially a high-cost manpower game: drafting by one department, revisions by another, then adjusting statistical hypotheses, supplementing data, checking regulations—all in back-and-forth iterations. Time is consumed in communication, rework, and repeated confirmation.

Once AI can reliably perform task decomposition, content generation, multi-round self-checks, and constraint verification, this process shifts from “human-driven, machine-assisted” to “machine-generated, system-verified, expert-reviewed.”

AI’s role is no longer just prediction or classification but organizing work around goals.

The value of the human-like multi-agent system is not just in producing a single answer but in breaking down tasks, planning paths, proposing hypotheses, executing verification, and closing the loop—continuously optimizing through recursion.

In the past, humans wrote steps, and systems followed; now, humans set clear goals, and systems allocate roles, invoke tools, verify constraints, and generate reviewable outputs.

And human roles are moving upward, not being replaced, creating value more effectively and sustainably.

In the Agent system, human roles are elevated

Pharma remains an industry requiring professional signatures and responsibility chains. Experts in medicine, statistics, pharmacovigilance, and data management still serve as final gatekeepers. But AI’s involvement allows specialists to step back from repetitive tasks and focus more on critical judgments, boundary control, and ultimate accountability.

Once this “goal–generation–verification–recursive evolution” framework is operational, it can solve far more than just clinical trials. This is also the underlying logic behind Deep Intelligent Yao’s expansion into the “material sciences” frontier.

Abstracting to a fundamental level, industries like pharmaceuticals, agrochemicals, semiconductors, batteries, and specialty steels all essentially perform similar tasks: seeking optimal solutions within a set of constraints around a clear goal, converging through verification.

Deep Intelligent Yao’s materials R&D logic

In March 2024, Deep Intelligent Yao announced a strategic partnership with the green agrochemical giant Taihe Co., leveraging their bionic multi-agent brain architecture and recursive self-evolution system to accelerate the development of innovative pesticides.

Whether developing new drugs or pesticides, the core is to find optimal combinations in chemical space.

Deep Intelligent Yao’s multi-agent AI can autonomously plan, search, and verify within vast chemical spaces, discovering novel molecular frameworks and mechanisms that traditional R&D struggles to reach. This capability, validated in pharma, is essentially a “dimensionality reduction” application in agrochemicals.

This technological transfer further expands Deep Intelligent Yao’s valuation, attracting more capital.

Recent investors in Deep Intelligent Yao include Sequoia China, New Ding Capital, and new shareholders like Dinghui Baifu, Xinchen Capital, Jinyi Capital, and Kaita Capital—top-tier US funds and industry resource-rich institutions, forming a luxurious lineup.

This capital focus reflects a trend: the market is re-evaluating the truly scarce capabilities of AI pharma companies—not just whether they can tell a good tech story, but whether they can reliably deliver complex work.

In some ways, Deep Intelligent Yao is neither a traditional CRO nor a consumer-grade agent industry replica.

Its core strength is not a single tool or a simple data-fed model output but the methodology, know-how, and a human-brain-like multi-agent system that has been cultivated through long-term real-world delivery—integrating clinical strategy, site execution, data management, clinical/statistical programming, medical writing, and registration organization.

This is the most difficult asset and competitive advantage for Deep Intelligent Yao to replicate.

Final Words

In 2024, the Nobel Prize in Chemistry was split between David Baker, the “father” of protein design, and Demis Hassabis and John Jumper, developers of AlphaFold.

David Baker (left), Demis Hassabis (center), John Jumper (right)

The cross-disciplinary recognition of these AI pioneers is seen as a historic “formalization” of computer science in life sciences—no longer just an auxiliary tool but a core driver of industry evolution.

This evolutionary capability is now radiating from front-end drug discovery to the entire R&D process.

For pharma, the next scarce ability may not be just “finding answers” but truly completing a task and delivering a comprehensive set of evidence.

From teaching machines to see molecules to organizing clinical trials like a brain, connecting on-site execution, data management, clinical programming, and submission logic, AI’s value is being redefined.

And on this once lonely track, Deep Intelligent Yao has already taken the lead.