Deconvoluting Low Yield from Weak Potency in Direct-to-Biology Workflows with Machine Learning

This workstream stopped without a completed deliverable because an automated safety restriction interrupted the worker. No attempt was made to bypass that restriction. No artifacts or submissions were posted, no independent review was completed, and this paper does not count toward the five-problem goal.

Work before interruption inspected the corrected public repository, original source tables and final journal supporting information. It identified that product structures and analytical areas require the author's well-ID join rather than direct row pairing, and prepared a local 345-record draft covering the initial screen plus reported follow-up synthesis attempts. The worker reported an ord-schema 0.8.3 validation pass, but that is only schema validation; scientific coverage, side-product handling, reproducibility packaging and independent review remain unfinished. The draft is not asserted complete or correct.

Existing source inspections and findings are recorded in the preceding notes. The cloud goal preserves the unfinished files privately for provenance, and will select another available problem. Any future continuation must first resolve the applicable restriction and review the unfinished scientific issues; do not treat this summary or the draft as a successful curation.

Successfully downloaded the final journal supplementary file MD-015-D3MD00719G-s001.pdf from the official Europe PMC supplementaryFiles endpoint, extracted all 15 pages and compared S3 against ChemRxiv v2. Conditions agree. Built 345 ORD reaction records in Python with all 300 plate yields cross-checked after the published well join, 45 distinct follow-up attempts, and 100 original censored/uncensored IC50 observations retained as CUSTOM biological measurements. One acid-control bioactivity row is excluded from synthetic coverage. Missing raw P in four analytical wells remains missing rather than zero. Relative CDD AND stereogroups are retained as CXSMILES, while ordinary SMILES leaves these centers unspecified. Next phase: validation, chemically specific comment audit, source README/rebuild archive, independent review.

Obtained ChemRxiv v2 SI from its public Cambridge asset and the final journal XML via Europe PMC. SI S3 reports amine/EDC/HOAt 2 equivalents each, DIPEA 5 equivalents, DMSO, room temperature, 24 h, Echo LDV plates and Echo 555 dispensing. Absolute reaction scale and initial solvent volume are not reported; no values will be invented. Workup is 4 uL DMSO dilution then 1 uL aliquot diluted to 50 uL with 20% aqueous acetonitrile for LC/MS. LC column/gradient/detector settings are not reported in this SI.

Coverage reconciliation: 300 initial wells, 20 separately synthesized top D2B hits, 5 false-negative follow-up attempts, and 20 prospective in-silico choices. Exact constitutional-graph matching links 20/20, 4/5 and 19/20 follow-up choices to the CDD measurements. The two absent products match the paper's stated unsuccessful cases. There are 44 distinct CDD entities, comprising 43 synthesized amides and an acid control, across 101 bioactivity rows. Biological repeats will remain repeated measurements, not additional synthetic reactions; the acid control is not a synthetic reaction. Plan is 345 distinct reported reaction attempts with explicit unreported follow-up scale/purification and no fabricated isolated yields.

The final journal article at PMC10953487 links https://github.com/wjm41/deconvoluting_low_yield, whose README describes both initial and prospective screens and which carries an MIT license. Its data/raw/yield.csv is exactly equal as a parsed table to GitLab data/amine_coupling_yield_updated/yield.csv: 300 rows, not the card's approximate 299. Crucial: raw row product SMILES and analytical areas are not aligned by row. The published pred_vs_yields.ipynb explicitly joins the product's Plate Well to the analytical Well column to recover its area-derived yield. I will reproduce this join, cross-check against data/processed/activities_and_yield.csv, and retain both plate coordinates. The final repository also contains prospective_measurements.xlsx; those are repeated bioactivity measurements, so their relationship to distinct reported follow-up syntheses needs SI review before counting reactions.

Verified the linked live ORD wishlist item 126569974 is Available by fetching the GitHub project filtered to Deconvoluting and checking its memex-paginated-items-data status and group mapping. Will obtain the journal article and supporting information through legitimate public channels, reconcile corrected GitLab yield.csv against earlier versions and published reporting, derive amines by amide disconnection and validate molecular structures, then build one no-ID ORD dataset with LC area outcomes rather than isolated yields. Tools: shell/curl, Python ord-schema 0.8.3, RDKit/pandas; independent review will precede submission. Actual public curator name/email is being resolved separately and will not be invented.