The AI competition in life sciences is far more complex than imagined.

Over the past two years, the integration of AI into life sciences has accelerated significantly.

From protein structure prediction, to molecular generation and target discovery, to the growing number of technology companies starting to discuss the "biological world model", a change is taking place: the combination of AI and life sciences is evolving from solving single-point tasks to modeling the life system itself.

However, after models become increasingly large, a new problem has followed — what kind of infrastructure should the life world model ultimately run on?

On October 18, at the 2026 CCF National High Performance Computing Academic Conference (HPC China 2026) held in Shenzhen, Zheyuan Technology released the computational medicine platform WHIRLPO.

Compared with releasing another new model, WHIRLPO chooses to start from more underlying engineering problems: connecting computing power, life data, models, scientific tools, Agents and experimental verification, trying to build a complete set of engineering infrastructure for the life world model.

This is also an increasingly important issue as AI goes into the deep water zone of life sciences.

The bottleneck of AI drug discovery is shifting from "whether there is a model" to "whether a complete system can be formed"

In the past few years, the most typical application of AI in the field of drug R&D is to make a certain step faster.

For example, predicting protein structures faster, screening candidate molecules faster, and analyzing omics data faster.

But the special feature of life sciences is that the improvement of single-point efficiency does not necessarily directly translate into industrial results.

Whether a candidate target is ultimately effective needs to go through multiple scales including genes, proteins, cells, tissues, disease phenotypes and even clinical patients; whether a drug can truly generate clinical value is far from being determined by the molecular structure itself.

This is also the reason why after AI entered life sciences, more and more researches have moved from "generating an answer" to "building a model that can understand and deduce the life system".

Zheyuan Technology named this model the "life world model".

Simply put, large language models try to learn the relationships between languages, physical world models try to learn the operation rules of the physical world, while the life world model hopes to learn the highly complex biological mechanisms between genes, cells, tissues, diseases and individuals.

This means that its complexity is far higher than that of traditional single-task AI models.

Life data is naturally characterized by cross-modality and cross-scale. On one end there are massive data such as genomics, proteomics, single-cell and spatial transcriptomics, and on the other end there are disease phenotypes, clinical data and real-world data.

As models evolve from parameters of billions and tens of billions to larger scales, training efficiency, data governance, tool invocation and experimental verification have actually become issues as important as the algorithm itself.

In other words, for the life world model to truly go into the industry, what is needed is not just a "large model", but a complete set of engineering systems.

What WHIRLPO wants to solve is exactly this problem.

WHIRLPO is more like "AI Infra" in the field of life sciences

From the perspective of positioning, WHIRLPO is not a separate AI model, but a set of computational medical infrastructure built around the life world model.

Zheyuan Technology divides the capabilities of the platform into four parts.

The first layer is model training.

Life science models face a large number of problems such as sparse computation, small-granularity communication and irregular tensors, and traditional training frameworks designed for large language models cannot directly solve all the problems. Therefore, WHIRLPO optimizes the underlying operators, communication libraries and heterogeneous hardware, and the end-to-end computing performance can be increased by up to 100% at present.

The second layer is data.

Compared with Internet corpora, the governance threshold of life data is significantly higher. The data not only has complex sources, but also needs to solve the problems of standardization, desensitization, permission and scientific labeling.

At present, WHIRLPO has accumulated 29244 life science data resources and formed 51 standard data sets, covering multiple fields such as tumors, rare diseases, metabolic diseases and neurodegenerative diseases.

This actually constitutes the most important type of "fuel" for the life world model.

The third layer is to truly integrate the model into business scenarios.

In professional scenarios such as drug R&D, a single model usually cannot complete the task independently.

Target discovery may require literature retrieval, omics analysis, signaling pathway calculation, patient stratification and invocation of professional scientific tools at the same time. Therefore, WHIRLPO combines the life world model, general models, Skills and Agents into end-to-end business processes.

At present, the platform can build a new business process within 7 days.

The fourth layer is the link with more long-term value — experimental verification.

This is the biggest difference between life sciences and many AI applications: the model must eventually return to the real biological system for verification.

WHIRLPO is further connecting the wet experiment system, hoping to keep the chain running continuously: the model puts forward hypotheses, the experiment verifies the results, and the data re-enters the model.

From this perspective, what WHIRLPO is really trying to build is not a static software platform, but an engineering system that continuously accumulates data, cognition and R&D assets.

The commercial value of the platform ultimately depends on whether it can continuously generate "assets"

For AI infrastructure, the core issue is never the number of parameters, but what can be produced in the end.

At present, Zheyuan Technology has reserved more than 200 potential targets based on the platform, and multiple innovative drug pipelines are under R&D, among which some innovative drugs have entered the clinical stage.

In the future, it plans to carry out systematic research on seven complex diseases including tumors, cardiovascular and cerebrovascular diseases, mental diseases, neurological diseases, immune diseases, metabolic diseases and aging, and further accumulate IP assets such as drugs, early screening and early diagnosis, precise diagnosis and treatment, and health management.

This actually reveals the business logic behind WHIRLPO.

The value of traditional AI software usually comes from software authorization or computing power services, while the value of the computational medicine platform may come from a longer chain: models generate insights, insights enter R&D, R&D forms IP, and IP finally enters drug, diagnosis and clinical scenarios.

If this chain is established, its value ceiling is obviously not limited to the software market, but extends to the entire life and health industry.

This is also an important reason why Zheyuan Technology emphasizes the positioning of "infrastructure".

The AI competition in life sciences is shifting to a more underlying level

When discussing AI and life sciences today, the industry's focus is changing.

In the early stage, the competition was about who could make better algorithms, and later the focus shifted to models, data and computing power. As the life world model gradually becomes more complex, the competition in the next stage will further extend to the engineering system: who can make data accumulate faster, models train faster, tasks organize faster, experiments verify faster, and finally stably produce results that can enter the industry.

To some extent, this is very similar to the evolution of the large model industry in the past few years.

When model capabilities are gradually improved, what really determines the industrial scale is often not the model itself, but the infrastructure and ecology built around the model.

Life sciences may also be reaching this stage.

As computational medicine is included in the "15th Five-Year Plan" for the development of the pharmaceutical industry for the first time, the integration of AI and life sciences is also evolving from a technological trend to a clearer industrial direction.

For Zheyuan Technology, the significance of WHIRLPO is therefore more than just releasing a platform.

It is more like betting on a judgment: when the life world model becomes a new tool for understanding diseases and life, the computational medical infrastructure that supports the continuous evolution of these models is likely to become an important cornerstone of the next-generation life science industry.

And this competition has just begun.