🏆 SPICE first_mut — First survivors from the clean-code pipeline

Milestone date: 2026-08-14 The first time the SPICE end-to-end loop (Pre-train → RL → pseudo-label reflow) produced real, MD-validated surviving mutants under the clean, fixed codebase. (All earlier attempts were blocked by the q-metric bug, NaN contamination, or the Head B' blob — this run had the q-fix + NaN guards + the 0.5-config new pre-train all in place.)

Protein

7QF3 = miniSOG (R57Q mutant), Arabidopsis thaliana, a flavoprotein photosensitizer (binds FMN/riboflavin, produces singlet oxygen)
116 residues (the modeled chain used by the pipeline), 1.10 Å X-ray crystal
Reference: Lafaye et al. 2022, Photochem. Photobiol. Sci. 21, 1545 (paper ref 21)

Rescue environment (Env_fail)

pH 7.5–8.0 / 330 K / ionic 0 — mild thermal + weakly alkaline stress (not extreme pH)
7QF3 collapses here; the mutants below survive the full 20-step window here

Eight surviving mutants (two episodes, two collapse environments)

Group 1 — episode 1 (pH 7.5–8.0 collapse) → Figure 3 (main paper)

The first clean-code episode produced 5 survivors in Env_fail (pH 7.5–8.0 / 330 K). Mutations converge on a small set of hotspot rescue positions (50 / 100 / 102 / 103).

File	Mutations	Q	steps	Strategy
`pseudo_7qf3_0_20.npz`	50:Q>Y; 102:V>Y	0.91	20	aggressive
`pseudo_7qf3_1_20.npz`	52:T>N; 104:L>N	0.90	20	aggressive
`pseudo_7qf3_2_20.npz`	50:Q>Y; 100:I>N	0.93	20	aggressive
`pseudo_7qf3_3_20.npz`	50:Q>W; 100:I>M; 102:V>E	0.91	20	aggressive
`pseudo_7qf3_4_20.npz`	50:Q>W; 102:V>Y; 103:Q>S	0.92	20	aggressive

Q-gate threshold 0.5; all scored 0.90+ → genuine fold retention, not "fell apart but didn't crash"

Group 2 — later episodes (pH 10.0 alkaline collapse)

After episode 1, the run probed an alkaline collapse environment (pH 10.0 / 330 K) and produced 3 more survivors, rescued by a different mechanism — N-terminal substitutions at residues 1–3 (rather than the central hotspot cluster).

File	Mutations	Q	steps	Strategy
`pseudo_7qf3_5_20.npz`	1:M>A; 2:E>L; 3:K>T	0.91*	20	aggressive
`pseudo_7qf3_6_20.npz`	1:M>P; 2:E>K	0.91*	20	aggressive
`pseudo_7qf3_7_20.npz`	1:M>W; 2:E>Q	0.91*	20	aggressive

* Q re-derived, not from logs: the per-mutant Q for these later episodes was not captured in the run logs (HPC outputs cleaned before download). It was re-computed from the archived time-averaged coordinates using the run's exact native-contact definition (native_contact_q, reference = wild-type 7QF3 Cα, cutoff 8 Å); the method was validated by reproducing the five episode-1 survivors' logged Q (0.90–0.93) to within 0.02.

Reflow: 8 survivors total → confidence-weighted ×8 → pseudo.tfrecord

Structure quality (MD time-averaged Cα, `analysis_metrics.csv`)

Metric	Value	Reading
Rg	14.47–14.56 Å	correct compact size for 116 aa (random coil ≈ 22 Å) ✅
Adjacent Cα bond	3.94–3.96 Å	proper chain geometry ✅
Helix content	23.3–27.6%	plausible α-helical level for a flavoprotein ✅
Nearest non-adjacent pair	3.89–4.19 Å	no clashes ✅
Consistency across the 8	highly consistent	one shared folded state ✅

Metrics recomputed uniformly for all 8 survivors (helix: d(Cαᵢ,Cαᵢ₊₃) < 6.5 Å; contact: non-adjacent Cα pairs < 7 Å). Q for survivors 5–7 was not archived (see above).

Mutation hotspots (ES convergence signal)

Residue 50 (Q→Y/W): 4/5 — aromatic packing stabilization
Residue 102 (V→Y/E): 3/5 — V102→E adds surface negative charge (deprotonated at pH 8), charge-balance mechanism (m5)
Residue 100 (I→N/M): 2/5
Easter egg: 103:Q>S hits the Q103 position that the miniSOG literature independently identifies as a functional hotspot (Q103L markedly raises singlet-oxygen yield, via an oxygen-access channel) — independent convergence on a biochemically meaningful site

Group 2 (alkaline) — N-terminal charge relief

Common target: residue 2 (Glu), 3/3 — the sole acidic residue in the N-terminal region; all three remove or reverse its negative charge at pH 10:
- M6: E2→L (charge −1→0) plus K3→T (+1→0) — neutralizes the E2–K3 charge pair
- M7: E2→K (−1→+1) — charge reversal (net +2)
- M8: E2→Q (−1→0) — removes the charge (net +1)
Direction tied to pH: the alkaline mirror of the episode-1 charge signal — episode 1 adds negative charge at pH 7.5–8 (V102→E); the alkaline trio eliminates the N-terminal negative charge at pH 10. Precise framing: "eliminate the acidic terminus charge", not a uniform net-charge balance (M7 actually adds +1; M6 nets 0). Hypothesis pending single-mutant controls (does E2→Q alone rescue? E2→K alone?)

Honest caveats

The Q-gate validates structural retention, not function (cofactor binding / photosensitizer activity unverified)
First-episode results from an ongoing run; cross-episode survival, ΔΔG, and baselines still pending
For the pH-10 survivors (5–7), per-mutant Q was re-derived from the archived coordinates (run logs not captured; method validated against survivors 0–4, see Group 2 table)
Construction currently falls back to the wild-type backbone + sidechain rebuilding: Head B' (= Head A fold of the mutant) yields a real fold but still carries local Cα clashes that are not yet engine-buildable, so a geometric sanity guard (Rg + local-clash check) rejects it and falls back (2026-08-14)

Paper placement

articles/SPICE_paper.md §3.3 "End-to-end operation (first results)" — written up
Reference added: ref 21 (Lafaye 2022)

Reproduce

# Pseudo-label structure analysis
cd /Users/redelectricity/Documents/Projects/SPICE
python - <<'PY'  # see the analysis script (Rg/bond/helix/contact/clash)
PY

Files

first_mut/
├── README.md                    # this milestone record
├── analysis_metrics.csv         # structure-quality metrics
├── run_log.txt                  # raw HPC run-log excerpt
├── pseudo_7qf3_{0..7}_20.npz    # 8 pseudo-labels (seq + env + coords)
└── 7QF3.pdb                     # native reference (if downloaded)