AutoRadionuclide · In-silico Discovery Engine

MIBG/NET Flagship Demo Campaign

Campaign mibg-net-demo-001 · Run 16140108 · May 25, 2026

27
Ledger entries
7
Model calls
4
Turns
⚠

In-silico demonstration only — no wet lab, no real isotope

All scoring functions are frozen heuristics, not validated predictive models. The wet-lab step is a stub that returns heuristic scores plus Gaussian noise. MIBG + I-131 (Azedra) is a real FDA-approved therapy; the scores shown here are illustrative engine outputs, not clinical measurements.

How the loop works

OuterLoop (AutoResearch meta-loop)

  1. 1Ask LLM: propose ONE strategy modification
  2. 2Apply modification to in-memory StrategyConfig
  3. 3Run InnerLoop — one discovery cycle
  4. 4Compare campaign score before vs. after
  5. 5Keep if improved (Δ > 0); revert if not
  6. 6Record modification + rationale in append-only ledger

InnerLoop (one discovery cycle)

  1. 1generate_candidates() — design module + LLM
  2. 2score_all() — frozen harness (never agent-editable)
  3. 3policy.rank() — acquisition function + diversity
  4. 4safety_check() — isotope/chelator feasibility
  5. 5wet_lab.submit_and_wait() — stub in this run
  6. 6update_surrogates() — GP refitted with new observations

Frozen harness = the benchmark spec. The planner may read it but never modify it. Every decision is recorded in an append-only SQLite ledger — rows are never updated or deleted.

Turn-by-turn outer loop

Turn 1First cycle
278d0747
Campaign scoreEstablished 0.8378
0.838
1 proposed · 1 selected
Turn 2
1c05e81e
Campaign scorePlateau — score held at 0.8378
0.838
No constructs passed safety checks.

Increase exploration weight

exploration_weight: 1.5 → 2

Insufficient diversity in recent batches; higher kappa increases exploration.

reverted
Turn 3
334c944e
Campaign scorePlateau — score held at 0.8378
0.838
No constructs passed safety checks.

Focus on validated NET vectors

prioritized_targets: [] → ["NET"]

NET targeting vectors showing strongest objective improvements.

reverted
Turn 4
20fb5be3
Campaign scorePlateau — score held at 0.8378
0.838
No constructs passed safety checks.

Switch acquisition function to EI

acquisition_function: "UCB" → "EI"

UCB may be over-exploring; EI focuses on high-probability improvements.

reverted

Score is non-decreasing by design: the outer loop reverts any modification that does not improve the campaign score (Δ ≤ 0). The score plateaus here because MIBG+none+I-131 is the only unique resolvable construct in the declared building-block space — this is honest scientific behaviour.

Candidate construct

iobenguane (Azedra)FULL

Norepinephrine transporter (NET) ligand; directly radioiodinated

Vector: MIBG · Chelator: none · Isotope: I-131

SMILES

NC(=N)NCc1cccc(I)c1

Formula

C8H10IN3

Source

PubChem CID 60860 (iobenguane)

RDKit descriptors (GP surrogate input)

Molecular weight (Da)275.093
Wildman-Crippen logP1.274
TPSA (Ų)61.900
H-bond donors3.000
H-bond acceptors1.000
Rotatable bonds2.000
Ring count1.000
Fraction sp³ C0.125

Morgan-2 fingerprint: 28 active bits / 2048 total

Heuristic objective scores

Binding affinity
0.750
Chelator stability
0.900
Half-life compatibility
1.000
Synthetic feasibility
0.800
Selectivity
0.800
Aggregate0.8385

Isotope physics (I-131)

Atomic number

53 (iodine)

Half-life

8.02 days

Primary decay

β⁻ (encoded: 0)

Source

IAEA Live Chart of Nuclides

Retrospective benchmark

57%
Engine rank accuracy
44%
Random baseline
4/7
Known-good in top 5

Engine places 4/7 known-good agents in top 4. Random baseline: 0.44. This confirms scoring machinery ranks known-good agents above known-poor ones at a rate better than chance. NOT a validated predictive model.

1
PSMA-I&T-DOTAGA-Lu177clinical
0.9046
2
DOTATATE-DOTA-Lu177approved
0.8962
3
PSMA-617-DOTA-Lu177approved
0.8869
4
FAPI-46-DOTA-Lu177clinical
0.8408
5
MIBG-I131-directthis runapproved
0.8385
6
DOTATOC-DOTA-Y90clinical
0.7894
7
RGD-NOTA-Ga68clinical
0.6658
8
Unknown1-NOTA-Lu177failed
0.5277
9
Unknown2-DOTA-Ga68-therapy-attemptfailed
0.4433

9 compounds (3 approved · 4 clinical · 2 illustrative failures). Benchmark confirms scoring machinery ranks known-good agents above known-poor ones — this does NOT establish validated predictive power for novel compounds.

Provenance

Run ID16140108
Model providermock-deterministic-v1
Featurizer version1.0.0
MIBG SMILES sourcePubChem CID 60860 (iobenguane)
Model calls this run7
Ledger entries this run27
Ledger entries (all runs)212

Every ledger row is immutable (INSERT-only). Model ID, prompt version, scoring version, config hash, and random seed are recorded per decision. Export generated from: scripts/export_run.py

Honest limits

  • ✗Scoring functions are frozen heuristics — not validated predictive models.
  • ✗Metal coordination chemistry is NOT modeled (coordination geometry, thermodynamic stability, kinetic inertness).
  • ✗Radiation dose profile (LET, β⁻/α/Auger particle energy, DNA damage) is NOT captured.
  • ✗Benchmark rank accuracy 0.57 vs. random baseline 0.44 — confirms wiring, not predictive power.
  • ✗StubWetLab returns frozen-harness scores plus Gaussian noise — no real radiochemistry.
  • ✗DOTA and NOTA produce identical Morgan-2 fingerprints (ring-size difference invisible at radius 2).
  • ✗Large peptide targeting vectors (DOTATATE, PSMA-617, FAPI-46) omitted from registry pending independent verification.

Limitations are encoded in the source — every scoring function that lacks a validated predictive model is tagged HEURISTIC or PLACEHOLDER in its returned ObjectiveValue.