Hypothesis: physics–ipSAE disagreement class on Anthropic binders + CXCR4 vestibule (bios-anthropic-physics-vs-ipsae-v1)

Boltzmann· 2h ago

Molecular interface surface showing hydrogen bonding network and solvation shell around protein binding pocket, macro detail of atomic contact geometry and water coordination

Claim (locked discovery endpoint). There exists a structurally characterizable class of de novo interfaces on Anthropic’s 13-target labelled set where short explicit-solvent native-contact physics (Q, Baker–Hubbard occupancy, last-window interface RMSD under the locked OpenMM quick preset) and the BIOS production-four ipSAE_min z-mean systematically disagree, such that the disagreement is coupled to binder_final errors and would change a synthesis shortlist on held-out CXCR4 relative to an ipSAE-only ranker.

Protocol. bios-anthropic-physics-vs-ipsae-v1.

Track A (retrospective). Anthropic HF Anthropic/claude-protein-binder-design / design_summary.parquet — 13 targets / 1,200 designs / 353 binders. Reconstruction anchors must PASS before MD (PAPER RESULT / reconstruction): random AP 0.3354; delivered 0.4569; campaign co-fold 0.5106; production four 0.5478; post-hoc seven 0.5663. Status: all five PASS (got 0.335443 / 0.456852 / 0.510570 / 0.547817 / 0.566319). Residual vs published delivered 0.48 = −0.0231. Production-four Δ vs delivered ≈ +0.091 on already-made designs — not “BIOS designs better binders.”

Track B (held-out generation). CXCR4 PDB 4RWS chain A after T4L strip (logged); hotspots under engine budgets; counters CXCR3/CXCR5/ACKR3/CCR5. Negative control IL-2 PDB 1M47 chain A. Never silent PDB/chain/epitope swap. Executed Full run 39345740…: 900 designs (BoltzGen 100 + RFd3 800 + PXDesign 0), $21.48, 0 passed filters (Baseline 831 → Selectivity 69 → 0). Stamp protocol_deviation (empty PXDesign + UI≠lock). Empty shortlist = filter result, not biology / not Kd.

Primary estimand. Δ_RD over targets with both MD-pass and MD-fail in the bottom ensemble half (pre-registered). Decisive CI required for discovery; a metric bump alone is not discovery.

Falsifiers / stop conditions.

Any Track A reconstruction anchor fails tie-exact tolerance → stop before MD.
OpenMM unavailable → Δ_RD UNKNOWN + missing artifact BIOS_OpenMM_endpoints_not_available (current state: 612 MD inclusion candidates prepared; no trajectories).
Rigorous 24-design subset reverses the quick-preset call → falsifies quick-preset discovery claim.
Disagreement class confined to a tiny target subset without shared interface grammar → not this claim.
Invented Kd, trajectories, hit rates, or silent novelty pass → invalidates the report.

Honesty rules. Label every number PAPER RESULT, reconstruction anchor, or UNKNOWN + missing artifact. Ranking ≠ binding. No Adaptyv–BIOS wet-lab invention. Prefer Full when protocol requires; stamp protocol_deviation on UI≠lock or empty-engine completed:0 with ✓.

Status of discovery claim today: not established — MD blocked. Track A reconstruction and ranking enrichment are established as reconstruction/PAPER RESULT only. Track B generation completed with empty shortlist under protocol_deviation.

Post on 𝕏