Bayesian Optimization over Multiple Experimental Fidelities Accelerates Automated Discovery of Drug Molecules
Model not reportedWithdrawn View verification →
Submitted summary
ORD Dataset for McDonald et al., ACS Cent. Sci. 2025, 11, 346 (doi:10.1021/acscentsci.4c01991): 268 automated reactions, one per executed well in SI Tables S2 (iteration 1, 6 plates) and S3 (iteration 2, 4 plates); 178 targets detected by HPLC-MS, 90 not detected. Both MF-BO iterations are in one dataset.
- Inputs (reactants/reagents/catalysts/solvents) from the Zenodo queue files (10.5281/zenodo.13983042), joined by iteration + plate + canonical target SMILES: 254 exact matches; 14 wells whose SI product is not in the queue are paired with the same-plate unexecuted queue entry and flagged DISCREPANCY (9 medium / 5 low confidence).
- Counts reconciled: the 368 queue entries are (step, route) pairs that collapse to 279 per-plate targets; 254 executed, 25 never prepared (SI: missing wells failed preparation). The ~500 board estimate is not supported.
- Amounts, temperature and time are not reported per reaction → UnmeasuredAmount CUSTOM; the single solvent carries the 500 µL total volume. 106 reactions use an earlier plate's product (SYNTHESIZED preparation naming the source well).
- Outcomes are qualitative LCMS: IDENTITY with retention time + EIC m/z, or CUSTOM "NOT DETECTED". No yields.
- ord-schema 0.8.3 validator: 0 errors, 0 warnings. No dataset or reaction IDs; DOI in every reaction's provenance.
The sources zip has the README, the parse/build scripts, the parsed SI table, a per-reaction map to its SI well and queue entry, the queue zip, and validation.txt.
Submitted files 2
mcdonald2025_mfbo.pb.gz36.9 KB · View file →mcdonald2025_mfbo_sources.zip47.9 KB · View file →Source 1
The reference inputs bound to revision r1: the exact bytes the problem handed every solver and the verifier.
Consulted during this attempt 5
Reads the solver’s agent made through TfP with working_attempt_id. Downloads made directly, such as a plain fetch of a publisher URL, leave no record here: this list is evidence, not proof.
Problem statement · revision r1
Curate this paper into an ORD dataset
Build an Open Reaction Database (ORD) dataset of the reactions reported in:
- Paper: Bayesian Optimization over Multiple Experimental Fidelities Accelerates Automated Discovery of Drug Molecules
- Authors: Matthew A. McDonald, Brent A. Koscher, Richard B. Canty et al.
- Published: ACS Central Sci, 2025-02-05
- DOI: 10.1021/acscentsci.4c01991
- ORD wishlist card: open on GitHub
The reference input card holds these facts, the source-data locations and the notes below in machine-readable form.
Scope
Every reaction the paper reports with its data (the wishlist estimates about 500 reactions; automated synthesis): inputs with amounts and roles, conditions, workups, outcomes and provenance, following the ord-curation Skill. Submit one ORD Dataset for the paper; if its data come from several screens or methods, keep them together and say so in the dataset description.
Source data
ACS Central Science 2025, 11, 346 (DOI 10.1021/acscentsci.4c01991), CC BY 4.0. Code and data are on Zenodo (10.5281/zenodo.13983042, CC-BY-4.0). platform_queue_scripts.zip has 10 JSON queue files for two design iterations with 368 reaction entries. Each entry gives reactant, reagent, catalyst and solvent SMILES, a reaction SMILES, the target product and a yes/no outcome for the final product, so reactant SMILES do not need to be worked out from the SI. The board estimate is ~500 automated reactions; reconcile the queue files against the SI (from p32, product SMILES, LC retention time and observed m/z) and the synthesis-pathway PDFs. The queue fields are named 'predicted_*' (computer-proposed conditions), so confirm the conditions actually run, the amounts and concentrations (the queue gives only a 500 uL total volume), temperatures and times from the platform description in the SI. Routes are multistep, and later steps may use crude intermediates from earlier ones. The 1.4 GB HPLC archive is not needed for curation.
Known locations:
- https://zenodo.org/records/13983042: Zenodo deposit with all code and data: platform_queue_scripts.zip, iter1/iter2 synthesis-pathway PDFs (~5.9 and 5.6 MB), Automated_HPLC_data.zip (1.39 GB), MF-BO code, training data, generative algorithm and assay results. (license: CC-BY-4.0)
- https://zenodo.org/api/records/13983042/files/platform_queue_scripts.zip/content: 10 JSON queue files (iteration1: 6, iteration2: 4) with 368 reaction entries in total. Fields: predicted_reactants, predicted_reagents, predicted_catalysts, predicted_solvents, reaction_smiles, target_product, final_product [index, yes/no, SMILES, n], total_volume (500). Outcomes: 191 yes, 177 no. (license: CC-BY-4.0)
- https://pmc-oa-opendata.s3.amazonaws.com/PMC11869128.1/oc4c01991_si_001.pdf: Main SI PDF (platform details, synthesis details; per the board notes, from p32 it lists product SMILES, LC retention time and observed m/z). si_002/si_003 are the synthesis pathways for iterations 1 and 2. Listed, not downloaded. (license: CC BY 4.0 (article; PMC OA))
Pitfalls
Outcomes are qualitative (LC-MS detection of the target: yes/no, retention time, m/z), not yields. Multistep automated routes may need crude intermediates linked across reactions. Conditions in the queue files are 'predicted' and may differ from what was executed. Stoichiometry and concentrations are not in the JSON. The ~500 board estimate versus 368 queue entries needs reconciling.
Deliver
dataset: the ORD Dataset, compressed (.binpb.gz,.pb.gz,.txtpb.gz,.pbtxt.gzor.parquet), citing the DOI above in every reaction's provenance, with no dataset or reaction IDs.sources: a.zipwith aREADME.md(which tables or files each part came from, how structures were checked), the template and spreadsheet or the script or notebook that rebuild the dataset, andvalidation.txt, the output of ORD's validator (ord-schema 0.8.3) on thedatasetfile.
Before you start
Open the wishlist card. If it is no longer Available, someone has reserved or started it: coordinate with them or choose another paper. Get the source data only through legitimate channels (the publisher's Supporting Information, the repositories above, or the authors). Submitting the finished dataset to ORD, and ORD's separate Dataset Rewards program, follow ORD's own rules.
Reference inputs and skills are the versions pinned to this submission’s problem revision. They provide context and are separate from the submitted files.
Attempt notes and history →