Enhancing High-Throughput Experimentation Liquid-Handling Capabilities via Strategic Hardware Development of the Mantis Liquid Handler

Outcome

Submission sub_d62xob8hif. Verification run vrn_95wcnoee87 is waiting at the curation-review stage.

• dataset = art_1715p5brln, golden_2025_mantis_hte.pb.gz, sha256 393c07ab…6832, 960 reactions, no IDs, DOI on every reaction.
• sources = art_2dg0v3wxil, golden_2025_mantis_hte_sources.zip: README, parse_tables.py, build_dataset.py, tables/S*.csv, validation.txt.

Approach

• The SI PDF came from figshare 30090846 (CC BY-NC) via the figshare API. I extracted it with pdftotext -layout, checked schemes on page renders, and regex-parsed the ten tables: 96 rows each, unique wells, full factorials.
• There is one Python template per plate family: KRED (S2/S3), phenol O-alkylation (S7/S8), amide coupling with stock solutions (S9/S10) and Mantis-first amide coupling (S11–S14). Amounts, devices, order of addition and workups come from the procedures on pp. S3, S10, S15 and S18.
• Structures come from the schemes and the authentic-product characterization. The build asserts that the RDKit formula equals the HRMS formula for every product.

Key learnings

• BMT-238442 is 4-(6-fluoroquinolin-4-yl)cyclohexan-1-one. The SI product name "4-(7-fluoronaphthalen-1-yl)cyclohexan-1-ol" is wrong. HRMS C15H16FNO, the ring N in the scheme and the quinoline H2/H3 doublets settle it.
• Outcomes are Prod/IS ratios, entered as CUSTOM measurements with uses_internal_standard, and Prod AP, entered as AREA %. Nine wells have genuine sub-1% AP values, which the validator warns about.
• Liquid bases "3 equiv, 120 µL" in the alkylation procedure are read as 120 µmol. EDC 9.5 mg/50 µmol is the HCl salt.

Open questions for the reviewer

• GDH appears only in the KRED scheme. It is recorded as a catalytic, unmeasured enzyme, and its source is unknown.
• The KRED products' stereochemistry (and cis/trans for BMT-238442) is not reported, so it is left unspecified.
• Di-tert-butylbiphenyl is taken to be the 4,4'-isomer. Its IS stock solvent is not stated.
• Tables S9–S14 have no well map, so reactions are in printed order.
• record_created email is the curator@example.org placeholder, to be replaced on ORD submission.

Process issues (not chemistry)

• The pinned Skill skv_k5sivlohu1 and the card resource could not be read in this client (resources/read schema mismatch). The card was fetched via its download_url, and ORD conventions were taken from ord-schema 0.8.3 directly.
• The Claude Code Remote get_session tool was not available, so report_agent_usage was not called.

Ran locally (python3.13, ord-schema 0.8.3, RDKit 2026.09.1):

• parse_tables.py si.txt tables/: regex parse of the pdftotext output. Each of S2, S3 and S7–S14 has exactly 96 rows. Wells are unique (S2/S3/S7/S8) and the amide plates are a complete 3 activators × 8 bases × 4 solvents factorial.
• build_dataset.py: one template per plate family. It asserts that every SMILES parses and that each product's formula equals the SI HRMS formula, which caught nothing beyond the BMT-238442 naming error already noted.
• Spot checks against the SI text all match: S2 H12 1.50, S3 E4 2.27, S7 B9 0.7 AP / 0.03, S8 G1 87.1 / 5.10, S10 TCFH/DBU/DMF 0.78, S13 DppCl/lutidine/EtOAc 23.99, S14 TCFH/PMP/THF 4.60.
• python -m ord_schema.scripts.validate_dataset: exit 0. Full validate_message: 0 errors and 9 warnings, all genuine sub-1% Prod AP values flagged as "fractions?".

Decisions worth a reviewer's eye:

• Prod/IS is a CUSTOM measurement with uses_internal_standard. AP is AREA %.
• KREDs are named by PRZ code with a CATALYTIC unmeasured amount. GDH comes from the scheme only.
• "3 equiv, 120 µL" liquid bases are read as 120 µmol.
• EDC is recorded as the HCl salt (9.5 mg/50 µmol).
• The IS is taken to be 4,4'-di-tert-butylbiphenyl.
• S9–S14 have no well map, so entries keep printed order.
• record_created email uses the curator@example.org placeholder, as in this account's other ORD runs.

Setup / blockers (minor)

• accept_agent_job OK. The Claude Code Remote get_session tool is not in this cloud session, so report_agent_usage could not be called with rate_limit_info.
• The pinned Skill skv_k5sivlohu1 (SKILL.md) and the card resource fail client-side: resources/read is rejected (ttlMs missing, invalid cacheScope). The card was fetched from its signed download_url. Other ORD Notes on this account hit the same thing. I follow ORD conventions from ord-schema 0.8.3 directly.

Sources fetched outside TfP

• figshare API articles/30090846 → op5c00216_si_001.pdf (50 pp, 1,393,730 bytes, CC BY-NC 4.0) via ndownloader. Text via pdftotext -layout; schemes checked with pdftoppm renders.

Reaction tables (all 96 rows each, 960 total)

• S2: KRED reduction of 6,7-dihydrocyclopenta[b]pyridin-5-one, 96 Prozomix PRZ codes, Prd:IS.
• S3: same 96 enzymes on BMT-238442. The scheme on p. S5 shows 4-(6-fluoroquinolin-4-yl)cyclohexan-1-one. The SI's characterization name "4-(7-fluoronaphthalen-1-yl)cyclohexan-1-ol" is wrong: HRMS C15H16FNO, the drawn N and the ¹H NMR (8.79 d / 7.43 d, quinoline H2/H3) all fit 6-fluoroquinolin-4-yl. Product cis/trans is drawn with a wavy bond, so stereo is left unspecified.
• S7 (PFE/PP chips) and S8 (PFE/FEP chips): 4-phenylphenol + hexyl mesylate, 6 bases × 4 solvents × 2 conc × 2 replicates; Prod AP + Prod/IS (IS = di-tert-butylbiphenyl).
• S9 (manual) and S10 (Mantis-integrated): 2-methoxybenzoic acid + 2-chlorobenzylamine, TCFH/HATU/DppCl × 8 bases × 4 solvents; Prod/IS.
• S11–S14 (Mantis-first): c1 1-phenylcyclopropane-1-carboxylic acid, c2 (E)-3-(3-CF3-phenyl)acrylic acid, c3 4-chlorobenzo[b]thiophene-2-carboxylic acid, c4 1-(4-CF3-phenyl)piperidine-4-carboxylic acid + 2-(cyclohex-1-en-1-yl)ethan-1-amine; TCFH/EDC/DppCl × 8 bases × 4 solvents. Structures come from the authentic-product names, NMR and HRMS (pp. S29–S31).

Plan: one Python build script with a template per plate family, ord-schema 0.8.3. Prod/IS goes in as a CUSTOM measurement with uses_internal_standard; Prod AP as AREA. Enzymes are named by catalog code with a catalytic unmeasured amount.