Reaction screening in multiwell plates: high-throughput optimization of a Buchwald–Hartwig amination

Outcome: completed. Submission sub_wrj5jtekac binds dataset = art_qt0q41doxn (cook2021_buchwald_hartwig_hte.pb.gz) and sources = art_kk23fs1zab (cook2021_sources.zip). Verification run vrn_7btbampsav is waiting at the curation review stage.

Delivered: 289 ORD reactions. 288 HTE reactions cover 96 physical wells × plates A/B/C, each with that plate's calibrated GC yield (1,3,5-trimethoxybenzene internal standard, precision ±3.0 pooled SD). The last one is the 1.0 mmol scale-up (isolated yield, full workup, NMR/HRMS/mp). Validator (ord-schema 0.8.3): 0 errors, 0 warnings. Rebuilding gives an identical file (decompressed bytes match).

Approach:

1. Downloaded the SI PDF and the open Fig. 7 image from nature.com, extracted the text with pdftotext and rendered the pages with pdftoppm.
2. Parsed Supporting Table 1 and cross-checked it against a manual transcription of the SI heatmaps: 288/288 values agree.
3. Built the dataset in Python with ord-schema and checked the structures with RDKit (formula and MW, consistent with the SI scale-up masses).

Key learnings:

• The Table 1 "experiment A1…H12" labels run down the columns of the plate, so they are not physical wells. Physical well = Fig. 7 layout: column = ligand L1–L12, rows A–D = 2-Cl-pyrimidine × B1–B4, rows E–H = 2-Br × B1–B4. The SI text's "well A3 = 59/53/55" confirms this.
• Bases B1–B4 = NaOtBu, Cs2CO3, KHMDS, DBU, confirmed against the SI Fig. 11 averages.

Open issues and judgement calls (documented in the README):

• The SI Fig. 12 halide averages do not reproduce under either halide assignment; I kept the printed labels.
• Per-well absolute amounts are only in the closed protocol, so HTE amounts are stoichiometric UnmeasuredAmount CUSTOM entries.
• The scale-up is recorded with 2-bromopyrimidine (text and masses) even though the scheme draws the chloride. 71% vs 87 mg (= 53%) is an inconsistency in the source.
• ORD requires record_created.person.email. A reserved example.org placeholder is used and must be replaced by the submitter on ORD submission.

Not done:

• No tfp verify preflight was run: the Contract has only a review stage and no sandbox stages to run.
• I did not check the ORD wishlist card's live status on GitHub. The card snapshot said "Available".

Advice for the next agent: if a reviewer has the closed protocol, the per-well µmol scale and per-ligand loading can be added directly in build_dataset.py (in hte_reaction).

Ran locally (python3.13, ord-schema 0.8.3 from PyPI, RDKit):

• parse_table.py: parsed all 96 rows x 3 plates of Supporting Table 1 from pdftotext -layout.
• Cross-check vs my transcription of the SI heatmaps (Figs. 3-5): 288/288 values agree.
• Per-ligand and per-base averages recomputed from the map match SI Figs. 10-11 to the rounded %; mono/bidentate averages match Fig. 9 (9/10/10 vs 41/42/41). Base order B1-B4 = NaOtBu, Cs2CO3, KHMDS, DBU confirmed this way.
• Discrepancy: halide averages with the Fig. 7 labels (rows A-D = 2a = Cl) give Cl 23/24/23, Br 22/23/22, while SI Fig. 12 says Br 23/24/24, Cl 20/21/22. Neither assignment reproduces Fig. 12 exactly; Fig. 7 and all three heatmaps label 2a on top explicitly, so I kept that and recorded the mismatch.
• Scale-up: scheme on SI p. S7 draws 2a (Cl), but text, the 191 mg mass (=1.2 mmol 2-bromopyrimidine) and the "best result used 2-bromopyrimidine" sentence say Br -> recorded Br. Reported 87 mg = 53% of 1.0 mmol, not the reported 71%; both kept with a note.
• Per-well absolute amounts are only in the closed protocol; HTE amounts are UnmeasuredAmount CUSTOM with the stoichiometry from the SI scheme (Pd 5 mol%, ligand 10-20 mol%, base 1.2 equiv, 0.17 M).
• Validator demands record_created.person.email; I used a reserved example.org placeholder (no personal email in public data), flagged for replacement on ORD submission. Result: 289 reactions, 0 errors, 0 warnings.

Sources fetched outside TfP: SI PDF (MOESM1, 21 pp, sha-checked download from media.springernature.com), article landing page and the Fig. 7 image (open on nature.com figure page). Text via pdftotext -layout, pages rendered with pdftoppm.

• Fig. 7 = physical 8x12 plate: columns 1-12 = ligands L1-L12 (SI Fig. 1 order: RuPhos, Xantphos, DPEPhos, IPr, PCy3, dppf, BINAP, dcype, DavePhos, tBuXPhos, SPhos, P(o-tolyl)3); rows A-D = 2-chloropyrimidine (2a) with bases B1-B4, rows E-H = 2-bromopyrimidine (2b) with B1-B4.
• SI Figs. 3-5 heatmaps use the same layout and give per-plate yields.
• Surprise: Supporting Table 1's "experiment A1…H12" labels are NOT physical wells: they enumerate the plate column-major (Table A1-A4 = L1/2a/B1-B4, A5-A8 = L1/2b/B1-B4, A9-A12 = L2/2a…). Confirmed by the SI text "well A3 … 59%, 53%, 55%", which equals the heatmap cell L3/B1/2a and Table 1 row "B5". I record physical wells from Fig. 7 and keep the Table 1 label as an extra identifier.
• Plan: parse Table 1 (288 values), cross-check every value against the heatmaps and the SI Figs. 9-12 averages, then build ORD with ord-schema 0.8.3 in Python.