We ran a one-day protein design competition with muni: 10 human teams against 6 autonomous frontier LLM agents (GPT-5.2, Claude, Grok 4.1, Gemini, Qwen, GLM 5), all targeting TREM2, a receptor implicated in Alzheimer's.

141 designs came in, the top 100 by Boltz-2 ipSAE went into the Adaptyv wet lab, and 37 came back as binders.

The two cohorts landed within a coin flip of each other: humans hit 25 of 65 designs (38.5%) and agents hit 12 of 35 (34.3%), with overlapping affinity distributions. The best human binder, 17_MRAZS_mosaic, reached 1.11 nM (team MRAZS swept three of the top four spots), and the best agent binder, 132_GPT_5_2_PXDesign, came in at 3.64 nM with GPT-5.2 winning the agent bracket.

The strangest finding was tool convergence: all six agents independently picked PXDesign (ByteDance), accounting for all 10 of the top agent binders, while none of them reached for Mosaic, BindCraft, or RFdiffusion. Mosaic is what MRAZS used to win this round, and what Escalante used to sweep our Nipah competition two months earlier. Frontier-model RAG converges on whatever is recent and well-indexed; humans converge on whatever they've personally seen win.

All designs and binding data are open under ODC-BY.