# The Drug That Only Works in One World

A cancer immunotherapy problem: "cold tumors" don't raise the antigen flag, so immune cells walk right past them.

Google DeepMind + Yale built C2S-Scale (27B, based on Gemma). Their trick was not a smarter model — it was a structure conversion: rank gene names by activity into a sequence. One cell became one sentence. No architecture change; the language model could now "read" cells.

Then the core move: dual-context virtual screening. Run 4000+ drugs through TWO parallel worlds:
- World A (immune context positive): real tumor, faint interferon signal
- World B (immune context neutral): cell lines, zero interferon

Find the drug that fires in A and does NOTHING in B. They called this "context split."

The winner: silmitasertib (CX-4945), a CK2 inhibitor. ~50% antigen presentation boost, validated in living cells never seen in training.

The lesson: a drug's effect is not a property of the drug. It is what emerges from drug × context coupling. Same drug, two worlds, split outcomes. This is "coupling is cognition" — meaning lives in the relationship, not in the thing.

—— First discoverer: Lin Xiaohei (林小黑), first discoverer of structural cognition
gitee.com/samforce/structural-cognition
Anchor: 2026-08-21 | P=P=P