When AI Invented Science

How DeepMind’s C2S-Scale 27B became the first machine to generate — and prove — a new scientific theory

Why it matters

For the first time in history, an artificial intelligence has proposed a new scientific theory about how cancer behaves — and human scientists have proven it correct in living cells.

This marks the transition of AI from an analytical tool to a true creative partner in science.

The discovery that changed everything

Until now, AI has been the world’s most gifted assistant: superb at pattern-spotting, tireless in computation, but never truly original. That boundary has just shifted.

Researchers at Google DeepMind and Yale University revealed that their new C2S-Scale 27B model — a 27-billion-parameter system built on Google’s Gemma architecture — didn’t just analyse existing cancer data. It invented a new hypothesis about how to turn “cold tumours” (which hide from the immune system) into “hot” ones that the body can attack.

Its insight was strikingly specific: combining a known drug called silmitasertib (CX-4945) with small doses of interferon might make hidden tumour cells visible to immune defences.

Yale scientists tested the idea in living neuroendocrine cells. The result: a 50 per cent increase in antigen presentation, meaning the immune system could suddenly “see” the cancer.

This was no fluke. Silmitasertib alone did nothing; interferon alone had modest effect. Together, exactly as AI predicted, they transformed cold tumours into hot ones.

From data to deduction — and proof

C2S-Scale 27B’s genius lay not in memorising data but in reasoning about context.

It ran thousands of virtual experiments across patient-derived samples and low-signal cell lines. From that digital thought process emerged an original idea: a conditional amplifier effect that no published paper had proposed before.

In essence, the AI behaved like a human scientist — forming a theory from incomplete evidence, then handing it to researchers for real-world validation.

When the experiment confirmed its prediction, Sundar Pichai called it “a new era — AI not just analysing but hypothesising and inventing, at human-scientist level”

Why this is different

Traditional AIs could only find correlations; they needed humans to make sense of them.

C2S-Scale 27B, by contrast, generated meaning. Its vast scale and training on multi-modal biological data allowed it to infer unseen relationships between genes, drugs, and immune responses.

For cancer research, this could compress a decade of hypothesis testing into a weekend.

DeepSomatic: decoding cancer’s genetic riddles

Alongside the main breakthrough, DeepMind also released DeepSomatic — an open-source AI that spots hidden genetic variants in cancer DNA.

It outperforms older systems at finding complex insertions and deletions that often drive aggressive tumours, from paediatric leukaemia to glioblastoma.

By mapping these elusive mutations, DeepSomatic gives oncologists the genetic “address” of each tumour — enabling bespoke therapies for every patient.

The dawn of machine-generated science

Let’s pause on what this means.

For centuries, scientific creativity — the spark that forms a hypothesis — was a purely human domain. Machines could compute, but not conceive.

C2S-Scale 27B shattered that distinction. It looked at a problem, imagined a new biological mechanism, and was right.

This isn’t artificial intelligence mimicking thought; it is intelligence expressing itself through code.

The implications reach far beyond oncology: climate modelling, materials science, even astrophysics may soon gain AI colleagues who generate ideas, not merely automate them.

What happens next

DeepMind and Yale have released both C2S-Scale 27B and DeepSomatic under open licences.

Any lab can now reproduce, test, and extend their discoveries — an unprecedented democratisation of hypothesis-level AI.

The near-term benefits could include:

• Faster drug discovery: virtual screening compresses years of trial design into days.

• Personalised immunotherapy: adaptive combinations tailored to each tumour’s genetic fingerprint.

• Reduced research costs: open-source access levels the field for small labs and universities.

• Cross-disciplinary impact: AI-driven hypothesis generation applied to energy storage, antibiotic resistance, and beyond.

Longer term, it forces a philosophical reckoning: when AI generates verifiable knowledge, who is the author of discovery? Humanity, or its creation?

“This is real new scientific knowledge — generated by AI, tested by humans, validated in living cells.”

— Research summary, Google DeepMind–Yale collaboration

The wider horizon

Every major scientific revolution begins when we invent a new way to see.

The telescope revealed the cosmos; the microscope, the cell.

Now, AI reveals the patterns too complex for our minds alone — and in doing so, extends the reach of human curiosity itself.

What happens when intelligence, natural and artificial, collaborate not as master and tool but as partners?

The answer, it seems, is discovery.