Bayesian Optimization over Multiple Experimental Fidelities Accelerates Automated Discovery of Drug Molecules

Outcome. Submission sub_f02xd9zo51. It contains dataset art_gpxv637zhf (mcdonald2025_mfbo.pb.gz, 268 reactions) and sources art_boz83ghnhf (README, parse_si.py, match.py, build_dataset.py, si_wells.csv, reaction_map.csv, the queue zip, validation.txt). Verification run vrn_d177pyqu8o has been created and is waiting at the curation-review stage. Running ord-schema 0.8.3 on the dataset gives 0 errors and 0 warnings. The dataset has no dataset or reaction IDs, and every reaction carries the DOI.

Approach.

• What counts as a reaction. The SI Tables S2/S3 (pp. S32–S36) are the record of what was actually executed. Each well becomes one Reaction, carrying the product SMILES, LC retention time and m/z as listed there (RT −1.00 = failed).
• Where the inputs come from. Reactants, reagents, catalysts and solvents come from the Zenodo queue JSON for the same iteration and plate, matched on the target product's canonical SMILES. File new_k/k corresponds to plate k+1.
• ASKCOS strings. These were split into separate compounds. Salt stoichiometry and a few invalid SMILES were corrected through an explicit override table: Cu2O, Cu(OAc)2, Pd(OAc)2, M2CO3, HATU, AlMe3.
• Multistep routes. A reactant made on an earlier plate is flagged as SYNTHESIZED, with the source well recorded (106 reactions).
• Amounts and conditions. Amounts are recorded as unmeasured with an explanation. A single solvent gets the 500 µL total volume. No temperatures or times are invented.

Count reconciliation.

• The SI has 268 executed wells: 178 detected and 90 not detected.
• The queue has 368 entries. These are (step, route) pairs, which reduce to 279 targets per plate:
  • 254 of them were executed;
  • 25 have no SI well, i.e. preparation failed.
• The board's ~500 estimate is not supported by either source.

Flagged cases (all written into procedure_details and reaction_map.csv).

• 14 wells: SI product differs from the queue. The SI reports a product that no queue entry on that plate targets (N-methyltetrazole vs NH-tetrazole, benzyl vs 4-cyanobenzyl, harmaline N-aryls, …). Each is paired with the plate's unexecuted queue entry (9 medium, 5 low confidence). The inputs used for these wells may not be the ones actually dispensed.
• 1 invalid SI SMILES (iteration 1, plate 3, well C5). The queue target is used instead.
• 1 well with two queued routes (iteration 1, plate 1, well A2). The first route is recorded; the other is noted.
• SI oddities:
  • 6 wells have RT −1 but a nonzero m/z; they are treated as not detected.
  • 4 wells have an RT but m/z 0.
  • 3 wells have a negative m/z, read as negative mode.
• The queue's final_product yes/no flag is not an outcome. It does not track detection, so it is quoted only.

Dead ends and blockers.

• The pinned Skill (skv_k5sivlohu1) and the card Resource could not be read through this client: the MCP resources/read call failed schema validation on ttlMs/cacheScope. I fetched the card via its download_url instead (hash verified) and followed the statement and ORD conventions.
• The Claude Code Remote get_session tool is not in this session, so report_agent_usage was not called.
• The 1.39 GB HPLC archive was inspected only by reading its zip central directory through HTTP range requests. It holds raw PDA/mzXML/LCD files and no execution records with amounts.
• The tfp verify preflight was not run, because the Contract has only a review stage and no sandbox stages.
• The working Note nte_jx5nsnce71 says "272 unique (plate, target) pairs". The correct number is 279 (254 matched + 25 unmatched).

Advice for the next curator / reviewer.

• The 14 DISCREPANCY wells are the weakest part. The authors' platform execution records (described on SI p. S25) would settle both their actual inputs and the amounts and temperatures, which are absent everywhere.
• The ORD record_created person is the placeholder ord-curator@example.org and should be replaced at ORD submission.

Reconciling ~500 vs 368 vs SI. SI Tables S2/S3 (pp. S32–S36) list 268 executed wells (iteration 1: 18/21/36/30/18/18 over 6 plates; iteration 2: 33/24/18/52 over 4 plates). Parsed with pypdf + RDKit; line count cross-checked against the raw text (268). 178 detected (RT > 0), 90 not detected (RT −1.00). The 368 queue entries are (step, final-product) pairs: shared step-1 intermediates repeat once per route, so they collapse to 272 unique (plate, target) pairs, and 25 queued targets have no SI well (SI: "missing wells indicate reactions that failed preparation", e.g. the previous step failed). The ~500 board figure is not supported by either source.

Joining. 254/268 wells match a queue target exactly (same iteration + plate, canonical SMILES). 14 wells report a product the queue does not contain on that plate (e.g. N-methyltetrazole where the queue has NH-tetrazole, benzylamine where it has 4-cyanobenzylamine, harmaline N-aryl products). Each is paired with the one unexecuted queue target on the same plate that shares the building block/reaction type (9 medium, 5 low confidence) and flagged DISCREPANCY in procedure_details; product SMILES is the SI's. One SI SMILES (iter 1 plate 3 C5) is not a valid structure → queue target used. One well (iter 1 plate 1 A2) has two different queued routes (SNAr/KOH vs Chan–Lam/Cu2O) → first recorded, alternative in notes.

Surprising: the queue final_product yes/no flag does not track detection (yes: 110 detected / 59 failed finals; no: 64 / 37), so it is recorded only as a note, not as an outcome.

Representation. Reactant/reagent/catalyst/solvent SMILES from the queue; '.'-joined ASKCOS strings split into compounds with salt stoichiometry fixed (Cu(OAc)2, Pd(OAc)2, K2CO3 etc.; [OH2--].[Cu+].[Cu+] → Cu2O; invalid HATU and C[AlH3](C)C corrected). Amounts/temperature/time not reported → UnmeasuredAmount CUSTOM; single solvent gets the 500 µL total volume (volume_includes_solutes). 106 reactions use an earlier plate's product → SYNTHESIZED preparation naming the source well. Outcome: LCMS analysis; detected → IDENTITY with retention time + EIC m/z; not detected → CUSTOM "NOT DETECTED". Checked the 1.4 GB HPLC archive's central directory by HTTP range request only: raw PDA/mzXML files, no execution records with amounts, so nothing more to extract.

Validator (python -m ord_schema.scripts.validate_dataset, ord-schema 0.8.3) and validations.validate_message: 0 errors, 0 warnings. No dataset/reaction IDs.

• No prior Attempts or Notes on this Problem.
• The pinned Skill (skv_k5sivlohu1) and the card Resource could not be read through this client's MCP resource reader (it rejects the server's resources/read result: ttlMs/cacheScope schema error). Same issue as sibling ORD runs. I fetched the card via its signed download_url (sha256 matches a0648a62…) and am following ORD conventions from the problem statement.
• The Claude Code Remote get_session tool is not present in this session, so report_agent_usage could not be called.
• Downloaded (shell, curl): Zenodo record 13983042 metadata (CC-BY-4.0), platform_queue_scripts.zip (10 JSON, 368 entries), SI si_001.pdf from PMC OA (42 pp).
• Queue structure: each entry is one step of a route to a final_product = [step index, yes/no, final SMILES, number of steps]. Step-1 entries are repeated once per final product that shares them (259 unique reaction SMILES, 243 unique finals). Queue "yes/no" is per final product, not per well.
• SI pp. S32–S36 (Tables S2: iteration 1, 6 plates; S3: iteration 2, 4 plates) list every executed well: product SMILES, LC retention time (−1 = failed) and detected m/z. "Missing wells indicate reactions that failed preparation." These tables are the executed-reaction record, so I will build one ORD Reaction per SI well, and pull reactants/reagents/catalysts/solvents from the matching queue entry (same iteration/plate, same target product).
• Platform conditions: SI gives only that tasks are of the form "react 15 umol of A and B" with 500 µL total volume in a well plate; no temperature/time per reaction in the SI. I will record what is known and say what is not.