We Misunderstood 'AI for Science'

Heisenberg said something that I keep returning to over the years.

"What we observe is not nature itself, but nature exposed to our method of questioning."

This is the cornerstone of quantum mechanics. But its true power extends far beyond physics.

I've been doing aging research for over a decade — building algorithms, models, predictions. But I've become increasingly aware of one thing: what we call 'AI for Science' has mostly gone in the wrong direction.

The Wrong Path of AI4S

When people mention AI4S today, what comes to mind?

Using deep learning to predict protein structures. Using large models to screen drug molecules. Using graph neural networks for materials design.

These are useful, important even. But they are essentially "AI as a tool for Science" — AI as a more powerful computational tool helping scientists process data faster and fit patterns more accurately.

This is AI serving science. This is not AI reinventing science.

What truly unsettles me is something else: the knowledge we feed to AI is itself a product of the old era.

Gene Ontology was established in 1998. KEGG pathways began construction in the 1990s. They are great engineering achievements that defined the common language of the molecular biology era. But they share a common assumption: a concept is a fixed, unchanging definition.

"Mitochondrial dysfunction" — it has an ID in GO, a gene list, a definition. You detect it in hypoxic liver cancer cells, you detect it in dopaminergic neurons in Parkinson's disease, and it's always the same list.

Is this true?

In Heisenberg's sense, it's not. Because what question you ask the system determines what face the system reveals to you.

The Uncertainty Principle of Concepts

I repeatedly run a thought experiment.

Suppose the biological concept of "mitochondrial dysfunction" is an observed object.

Under one mode of inquiry — say, "why does this cancer cell proliferate wildly in hypoxia?" — it manifests as a set of glycolytic reprogramming genes, weighted toward the metabolic side.

Under another mode of inquiry — say, "why does this neuron die prematurely under oxidative stress?" — it manifests as a set of mitophagy and apoptosis regulation genes, weighted toward the damage side.

These two gene sets overlap but are not identical. Different weights, different structures, different narratives.

They are both "mitochondrial dysfunction." But they are not the same incarnation.

The content of a concept depends on the context of inquiry. The representation of a concept is not a fixed vector, but a function that varies with the question.

This is not a metaphor. This is the most fundamental epistemological fact of biological knowledge.

Yet all our ontologies, all our knowledge bases, all our databases pretend this fact doesn't exist. They freeze a concept into an ID, a list, a definition, and tell AI: "This is knowledge. Learn it."

How can this be true AI4S?

What AI4S Should Really Do

The mission of AI4S should not be "using AI to accelerate existing science," but rather "using AI to reinvent how science organizes knowledge."

The first thing true AI4S should do is not build models, but rebuild concepts.

What we need is a knowledge system where:

Every concept is not a static label, but a living Agent. It has its own core memory, but facing different questions, different contexts, different intervention conditions, it generates different avatars. These avatars carry question-adapted gene weights, clinical context markers, and intervenability scores.

"Mitochondrial dysfunction" is no longer an entry. It's an Agent. When you ask "why isn't this patient's tumor responding to immunotherapy," it automatically generates an avatar adapted to that question, providing the gene network that best explains and intervenes in that context.

This is the true echo of Heisenberg's words in biology: knowledge itself must live in the moment of being questioned.

We don't need a bigger database. We need an operating system for concepts.

Why EvoSika Exists

EvoSika is doing exactly this.

We define every biomedical concept as an Agent. An Agent has Memory — its core gene set. An Agent has Skills — it can adaptively adjust weights and generate avatars across different tasks. An Agent has a life — it gets evaluated, competed, and evolved on public data.

The Chinese speak of "化身" (incarnation), Westerners call it Avatar. The same concept can manifest as different gene program forms across different diseases, tissues, intervention conditions, and ethnic medical traditions.

Traditional Chinese Medicine speaks of "气虚" (qi deficiency), a vague clinical concept. But in EvoSika's framework, the "qi deficiency" Agent can, within a specific tumor immune microenvironment dataset, be instantiated as a particular immunosuppression-related gene program avatar. It's no longer an ancient term, but a quantifiable, verifiable computational object that can be aligned with modern pharmacological targets.

This is what I believe true AI4S is: not using AI to serve old knowledge, but using AI to birth new forms of knowledge. Not a tool, but a paradigm.

Epilogue

Heisenberg said those words in 1927. Nearly a century later, we can finally give them a computational embodiment in biology and medicine.

EvoSika is the first bone of that embodiment.

I invite everyone who feels "something seems missing from current AI4S" to take a look at what we're building. It's immature, it's just beginning. But its direction is the only one I've found worth committing to after twenty years of computational biology.

True science always begins with a new way of seeing.

We are building such eyes.

Xiong Jianghui

Founder, EvoSika