Curator state

Recently surfaced

• Find a selective tool compound for a target (added 2026-08-16) — discharges the top Drug Repurposing and Discovery deferred item and gives the catalogued, GA, verification: works GtoPdb skill its first recipe. The banked amend-vs-add question against profile-compound-polypharmacology resolved to add, exactly as the note predicted: that page runs compound → targets over ChEMBL, this one runs target → ranked candidates with a selectivity window, and the two are now cross-linked both ways with the ChEMBL page used as this recipe’s step-8 second source. Rung 2, with the boundary stated honestly rather than inflated — the deliverable is a stdlib script against a public unauthenticated API, so a chemist who already knows the endpoint map is effectively at rung 1; the skill earns its place on the endpoint map plus two API traps verified this run, both of which return a clean-looking table of the wrong data: name= lookups are fuzzy across all fields and can silently resolve a different target, and approved=true is silently ignored (use type=Approved). API field names were read off live responses this run, not assumed: /services/targets/6/interactions returns affinity, affinityParameter, primaryTarget, targetSpecies, selectivity (curator free text, e.g. "Not Determined"), action, endogenous, refs[].pmid. Four gates. (1) Species declared before any affinity is read and never pooled — the load-bearing one, and it has an unusually sharp number: cross-species deviations of up to 3.2 log units (~1 600-fold, i.e. ~50× wider than a 30-fold selectivity window) and recombinant-vs-native-tissue deviations up to 4.2 log units (Bindel & Seifert, Mol. Pharmacol. 2026); verified on the API that zotepine’s 16 interactions interleave Human and Rat rows with no other distinguishing mark. (2) Absence of off-target data is not selectivity — the verdict is three-way and defaults to insufficiently_profiled, with n_targets_profiled printed next to the ratio, because GtoPdb is a curated representative selection (3,103 targets / 13,260 ligands — Harding et al., NAR 2026) and the same comparison found 69 of 302 histamine ligands with missing or incomplete functional characterization plus limited cross-database overlap, concluding selective inclusion “may give the impression of missing data despite substantial experimental data”. (3) affinity is a string that can be empty — verified on interactionId 7237 ([¹⁸F]altanserin, affinity: ""); an empty affinity is not a weak affinity, so those rows go to no_affinity.csv and never reach the ranking, and affinityParameter is never pooled across pKi and pIC50. (4) Selectivity computed by subtraction in log units, not division — Δ 1.5 is ~30-fold, whereas dividing 8.6 by 7.1 gives 1.2 and means nothing; this is the arithmetic error the gate exists to prevent. primaryTarget and the curator selectivity note are carried as evidence columns, deliberately not folded into the verdict (both 5-HT2A antagonist rows fetched had primaryTarget: false, so it is not a usable filter). Proposed — nobody has measured whether selective_by_evidence compounds behave selectively in a cell assay, and the page says so. Stdlib-only, so no ## Dependencies block. Fully open (ODbL-1.0 database, CC BY-SA 4.0 contents, CC BY 4.0 skill code, no auth) with the share-alike obligation on a published candidates.csv named, and the query-discloses-your-target caveat stated since there is no local mirror in the recipe. Laptop; the only scaling cost is one call per candidate ligand, hence the cache-to-raw/ rule. Note for the next pass: GtoPdb API responses carry no release version, so provenance.json records the query date and returned counts instead — do not invent a release number. PDSP Ki DB is filed under Missing components.
• Design a Bayesian trial that borrows external control data (added 2026-08-16) — discharges the second-ranked Translational Medicine deferred item and opens the Bayesian leg of the clinical-trials chain: the cookbook had a frequentist group-sequential design page, a protocol-drafting page and a missing-data analysis page, and nothing for the borrowing decision that the group-sequential page deliberately scoped out. The banked framing held: the page is about prior elicitation and borrowing, not “same design, different framework”. Rung 2 on the Bayesian Trials skill, argued explicitly — the RBesT arithmetic is a handful of lines, so the rung is earned by the conventions around it (the 20–80% ESS band, the 0.1–0.3 robust weight, the mandatory EXNEX sweep, the R-hat/ESS gates), without which Claude Code alone emits a syntactically valid design that cannot be reviewed. Three gates. (1) The no-borrowing comparator is designed and committed before the historical data are opened (design/00_no_borrowing_reference.R), following the group-sequential page’s fixed-sample-reference precedent — without it, “borrowing saved 40 patients” is an assumption. (2) The borrowed information is declared as a prior effective sample size in patients, written to map_prior.json before operating characteristics are computed — “we used a MAP prior” does not tell a reviewer whether 8 or 80 patients are being injected. (3) Operating characteristics are simulated over a two-way true-effect × prior-data-conflict grid (±1, ±2 τ), never only under exchangeability, with the robust weight swept and each weight kept as its own rows; the conflict column is read first. Grounding: robustification “alleviates” prior-data conflict rather than removing it (Schmidli et al., Biometrics 2014), and pre-specifying the borrowing amount is still an open methods question, which is why a semiparametric MAP with a generalised prior-ESS estimator was proposed (Hupf et al., Stat Med 2021); the dose-finding leg is supported by 3+3 selecting the correct MTD at up to 3× lower probability than CRM/Keyboard/BOIN (Chiuzan & Dehbi, Clin Trials 2024). Proposed, Fully open (MIT skill, CRAN packages) with the Stan C++ toolchain friction named, Laptop — with the honest split that the gMAP() fit and the conjugate OC grid are minutes, while EXNEX basket simulation has no conjugate shortcut and a full weight sweep is an overnight job. Two things a reviser should know: the Neuenschwander 2016 EXNEX 0.5/0.5 baseline is cited without a DOI (the catalog page records the year only; two papers MCP searches for it returned nothing this run) — find the DOI before turning it into a link; and the January 2026 FDA Bayesian draft guidance is flagged as a draft everywhere it appears, per the catalog page. Propensity-matched external control arms (psborrow2, patient-level data) are named under Alternatives as a different framework, not an escalation.
• Extract spectral features from resting-state EEG (added 2026-08-16) — discharges the banked Neuroscience deferred item (“Process resting-state EEG for spectral features end-to-end”), the natural companion to the ERP and HRV pages. Two decisions resolved against the banked note. (1) Component: the banked note offered “NeuroKit2 or MNE-EEG skill”; the pick is the third option, the EEG Processing skill, whose SKILL.md triggers verbatim on band power / downsample to frequency bands / resting-state .set/.edf/.bdf/.fif — a closer fit than either banked candidate. NeuroKit2 is named under Alternatives as the right pick when EEG is one channel among ECG/EDA/RSP. (2) The recipe is not the band-power pipeline the banked note described. Fixed-window band power is the failure mode, not the deliverable: the spectrum is aperiodic 1/f plus narrowband peaks, so an “alpha power” difference moves whenever the slope moves. So the page adds one pinned dependency the skill does not cover — fooof==1.1.1, Apache-2.0, import fooof, fetched from PyPI this run — and emits periodic and aperiodic parameters side by side, both labelled. Pin note for the next pass: the successor package specparam is at 2.0.0rc7 (2026-06-09) and still a release candidate, hence the stable predecessor; re-check whether 2.0.0 final has landed. Two gates, each traceable to a measured property: fit range and max_n_peaks are recorded literals chosen before the group labels are seen — odd–even reliability of the aperiodic intercept and slope falls as more peaks are allowed, with more outliers generated (Kałamała et al. 2026) — and a pre-registered r_squared floor with excluded channel IDs written out, because a bad 1/f fit yields a confidently wrong exponent and leaves no trace in the feature table. The confound is documented in both directions (sleep stage / age / region across n=251, Schneider et al. 2022; dopaminergic modulation in PD, Wang et al. 2022), and FOOOF-vs-IRASA has a ground-truth head-to-head (Gerster et al. 2022). Proposed, rung 2, Fully open, Laptop — and unlike the ERP sibling the path is fully deterministic (no ICA), so re-runs are byte-identical. Time-resolved parameterization (SPRiNT) is named under Alternatives as a distinct future recipe, not an escalation.

• Fit a drift-diffusion model to choice and reaction-time data (added 2026-08-16) — opens the cognitive-modelling leg of the Neuroscience section, which until now was entirely signals-and-images (EEG, fMRI, DTI, calcium, spikes, morphometry, pose) with nothing for behavioural data, the cheapest and most-collected measurement in the field. Rung 2: Drift-Diffusion Model skill + pyddm==0.9.0 (MIT, import pyddm, fetched from PyPI this run). The rung argument is deliberately not “the code is hard” — a PyDDM fit is a few lines — but that a converged-looking fit of an unidentifiable model emits a clean parameter table and a publishable-looking group difference with nothing flagging it; the skill’s value is entirely the non-code part (variant choice, free-vs-fixed parameters, convergence and PPC criteria, the recovery check). Gates: a committed ddm_spec.md declaring the hypothesis in parameter terms, the RT bounds and the free-parameter set before fitting, so the parameter set is a prediction; 10 random starting points with the likelihood spread recorded (a single start silently reports a local minimum); posterior-predictive RT quantile overlays, not means; and the recovery check with the **

  r

  > 0.5 cross-parameter tradeoff** warning (classically boundary vs non-decision time). The Parameter Recovery Checker skill is held at Alternatives rather than added by default — adding it would put a second component in the path for every routine fit. AI-generated-skill caveat carried into Availability, as with the rest of that collection, which is also why step 3 makes the reader write the exclusion rule themselves rather than take the skill’s. Proposed on the assembly but the method leg is current and strong: PyDDM’s own paper (Shinn, Lam & Murray, eLife 2020) plus a 2026 macaque transitive-inference study that used it to attribute learning to drift rate and the eye-vs-reach RT gap to non-decision time (Munoz et al. 2026) — a worked example of the recipe’s output being the scientific answer. HDDM is documented as an in-recipe substitution (low per-subject trial counts, hierarchical shrinkage — Wiecki et al. 2013), not a sibling recipe. Fully open, Laptop.

• Find differentially regulated phosphosites, not just changed proteins (added 2026-08-15) — discharges the top Molecular and Cellular Biology deferred item (banked 2026-08-08 while writing the LFQ page) and opens the site-level PTM leg. The banked amend-vs-add question resolved to add: the rollup rule, the localization gate and the protein normalization are genuinely different machinery, not a parameter swap on find-differentially-abundant-proteins-in-lfq-ms, which is now cross-linked as the sibling and as this recipe’s own unenriched arm. Correction to the banked note, worth carrying: it claimed the MaxQuant skill “emits Phospho (STY)Sites.txt-class tables”. It does not — the upstream SKILL.md, fetched this run, stops at proteinGroups.txt and contains no mention of site tables, localization probability or PTM analysis. The rung-2 call survives that correction for a specific reason: step 1’s paired unenriched proteome run is exactly the protein-level pipeline the skill documents, so the skill is load-bearing rather than decorative, and the page says plainly that the site-level layer is composed on top. Column names (Localization prob, Score diff, Amino acid, Positions within proteins, Reverse, Potential contaminant, ___1/___2/___3) were read off coxdocs this run. Three gates, each traceable to an observed property: (1) the paired proteome is a required input declared before enrichment, not an optional extra — it cannot be repaired after the sample is gone, and the confound is large in exactly the experiments people run (1,150/5,385 proteins significantly up at 20 h IL-3 alongside 4,218 regulated sites in the same samples, Esnault et al. 2018); the output carries raw and protein-normalized statistics side by side plus a three-value protein_status, because the correction itself is not benchmarked. (2) Localization cutoff declared as a literal with the fragmentation method recorded next to it — thresholds are method-dependent at 1% FLR across 180 synthetic phosphopeptides (Savitski et al. 2011), four pipelines disagree on the same dataset (Jiang et al. 2021), and HUPO ran a 22-lab challenge on 94 synthetic peptides because this is still hard (Hoopmann et al. 2020); sub-cutoff sites go to their own table and may only be described as “a site in this peptide”. (3) Multiplicity columns expanded to rows, never summed — ___1/___2/___3 are distinct molecular species that look like replicates to any agent reading the header. Supporting rules: no protein-level rollup at any point (activating and inhibitory sites on one kinase move opposite ways); enrichment chemistry fixed across the comparison and recorded, since Fe³⁺- and Ti⁴⁺-IMAC shared only 10% of 2,905 phosphopeptides from the same digest (Lai et al. 2012); and the two-flag site filter is flagged as not the protein-level three-flag filter, because Only identified by site does not exist in the site table and copying it silently matches nothing. Proposed. Fully open with the same MaxQuant free-but-not-open-source and no-macOS caveats as the sibling; Laptop kept for consistency with the sibling since the hardware class is identical — two searches plus one extra variable modification make it 2–4× the hours, not a bigger machine. Kinase-activity inference is named under Alternatives with its benchmark caveat (mean AUC 0.722 over 184 kinase-condition pairs, strongly dependent on substrate count — Hernandez-Armenta et al. 2017) but has no catalogued component, so it is filed under Missing components rather than given a followable path.
• Look up the curated composition and stoichiometry of a protein complex (added 2026-08-15) — discharges the top ISCB deferred item (banked 2026-08-08, sharpened by the interface recipe). The banked two-framings question resolved to the narrower one — “is this a curated complex, and at what stoichiometry” — because it is the interface recipe’s actual unmet dependency: that page cites Complex Portal and IntAct in step 6 as its orthogonal-evidence line but only as a link, and the new page is now cross-linked there as the followable path. The load-bearing rung decision repeats the Rhea precedent: both components are Claude.ai Connectors inside Claude Science with no separately addressable MCP URL, so they cannot leave a re-runnable artifact behind — hence rung 1, Claude Code scripting the CC0 ComplexTAB release files, with the Structures & Interactions connector named under Alternatives as the one-off interactive route. Release verified this run: ftp.ebi.ac.uk/pub/databases/intact/complex/current/complextab/ dated 2026-01-14, ~28 per-species TSVs named by NCBI taxon ID, and the exact 19-column header plus the worked CPX-4111 row were read off 9940.tsv. Four gates, each traceable to an observed property rather than invented: (1) refuse 9606_predicted.tsv by default — it sits next to 9606.tsv and is the same 5.0 MB size, so a *.tsv glob silently doubles the human hit count with non-curated entries; require an explicit --include-predicted and stamp a curation_set column so the sets can never merge downstream. (2) The Evidence Code ECO term is the headline field, not a footnote — the shipped CPX-4111 example is ECO:0005547 (“inference from background scientific knowledge”) with a blank Experimental evidence column, i.e. a curator’s reasoned assertion rather than a measurement; emit raw ECO ID + label + a derived experimental/inferred class. Complex Portal has used ECO for exactly this since 2015 (Meldal et al. 2015). (3) Fail on an unrecognised participant prefix — the composition column mixes UniProt accessions with ChEBI small molecules, RNA Central IDs, and nested CPX- subunits, and a silently dropped ChEBI cofactor turns an obligate holoenzyme into an apoprotein. (4) Absence is an explicit row, never an empty file — curation is deliberately incomplete outside the finished yeast and E. coli complexomes, so “no curated complex found” is emitted and IntAct is the named fallback for the weaker pairwise question. Supporting rules: parsed copy numbers cross-checked against the independent Complex assembly string (Heterotrimer) into a stoichiometry_agrees boolean; CPX- version suffix retained, since the IDs are stable and versioned (Meldal et al. 2019). Proposed on the assembly but strong at resource level — and the why-not-a-PPI-network argument has a number: curating the yeast complexome gave a 50% increase over CYC2008 and only ~40% of expanded co-complex pairs also have genetic interactions (Meldal et al. 2021), so co-membership and interaction evidence are substantially non-overlapping. Stdlib-only, so no ## Dependencies block; Fully open (CC0 since the 2022 release) and fully local after step 1 — the deliberate privacy contrast with the connector path. Laptop; human at 5.0 MB is the largest file and the download is the slow step. Notes for the next ISCB pass: ebi.ac.uk/complexportal/documentation returned a content-free shell to WebFetch, so the recipe depends on the FTP files only and makes no claim about the REST API; two IntAct papers MCP searches returned nothing, so no MI-score threshold is asserted anywhere — IntAct is cited by catalog page and ebi.ac.uk/intact/ws only. There is also no README in the complextab directory; the column schema was read from a data file.
• Compare bulk TCR/BCR repertoires across samples from raw immunosequencing reads (added 2026-08-15) — discharges the top Immunology and Microbiology deferred item (banked as “the largest remaining hole”) and gives the cookbook its first bulk repertoire page: the three existing repertoire recipes are all single-cell, while the amplicon path most vaccine and immuno-oncology labs actually run had nothing. Both banked questions resolved. (1) Licence: confirmed by fetching milaboratory/mixcr this run — free for academic users with no commercial funding, paid licence otherwise, activated via mixcr activate-license / MI_LICENSE / ~/.mi.license. So availability: Institutional access, not Fully open, following the neoantigen recipe precedent for academic-only binaries; the page tells the reader to clear it with tech transfer before scheduling cohort runs, and names the human-subject IRB layer alongside. (2) Depth normalization is a gate, not a step — DownSample to a literal depth set in config.yaml with the RNG seed recorded, a hard abort if any sample is below target, repeated N times so the spread is visible. Rung 3, escalation argued explicitly: MiXCR is a licensed Java aligner whose correctness rides on a chemistry preset, VDJtools is a statistics layer whose correctness rides on normalization decisions made after assembly, and visualization is load-bearing rather than decorative because rarefaction curves at shared x-values are how a reader checks the normalization instead of trusting it. Supporting gates: preset declared in config.yaml before any read is aligned (a mismatch does not error, it silently mistrims CDR3 boundaries); assembly_qc.csv gated on before any diversity number is read, with mean reads-per-UMI near 1 meaning the UMI families corrected nothing; non-functional clonotypes split into their own table, never silently dropped; chains never pooled; Hill profile at q=0/1/2 reported together, so a difference appearing only at q=0 is legible as a claim about singletons, i.e. about your PCR; Morisita-Horn and Jaccard both emitted, with a Jaccard-only conclusion flagged as a depth story. Proposed on the assembly, but the two gates are the strongly-evidenced part: multiplex-PCR frequencies only 42–62% accurate and uncorrected error inflating diversity up to 5000-fold, with UID tagging restoring 98–100% error correction (Khan et al. 2016); the canonical age-vs-diversity result required precisely normalized per-10⁶-cell profiling (Britanova et al. 2014); and step 6’s “this is your sample’s diversity, not the subject’s” caution rests on classical richness estimators frequently underestimating true TCR diversity (Laydon et al. 2015). Workstation with GPU is the honest bucket for a CPU-only 8–32 GB job (no GPU helps anywhere), with the 20-sample-cohort-is-an-overnight-array note and the quiet 2–3× disk constraint stated. Note for the next Immunology pass: mixcr.com/mixcr/about/license/ 404s — cite the GitHub repo for licence terms.
• Get balanced reactions and cross-references for a metabolite or EC number (added 2026-08-15) — discharges the top Chemistry deferred item (banked 2026-08-08, “Rhea + ChEBI”), and the banked “is this a recipe or a step the FBA recipe should gain” question resolved to its own recipe: predict-gene-knockout-phenotypes-with-fba starts from a model that already has the stoichiometry, whereas this reader’s reaction is missing from the model, so the entry point and the deliverable both differ. The load-bearing design decision is rung 1, not rung 2: both catalogued components are Claude.ai Connectors inside Claude Science with no separately addressable MCP URL, so they cannot leave a re-runnable artifact behind — the honest recommendation is Claude Code scripting the CC BY 4.0 release files, with the Chemistry connector named under Alternatives as the one-off interactive route. All ten TSV files, their columns, and their 2026-06-10 dates were verified this run against ftp.expasy.org/databases/rhea/tsv/ and its README.txt; Rhea release 141 has 18,558 reactions / 15,193 compounds. Four gates, each traceable to a stated Rhea convention rather than invented: (1) map the ChEBI ID through chebi_pH7_3_mapping.tsv before searching, because participants are the major microspecies at pH 7.3 and a query on the neutral form of a polyanion returns nothing — which reads identically to “no known reactions” — with that file’s third column (computation vs curation) carried out as a confidence field; (2) never emit a master ID as a modelling result, since the quartet is UN/LR/RL/BI with four separate identifiers and the master’s direction is undefined (<?>), reinforced by rhea-reaction-smiles.tsv covering only LR/RL so wanting a machine-readable equation forces the direction choice; (3) label the output columns side_left/side_right, never substrates/products — Rhea attaches no semantic meaning to the two sides, and this is the silent error that corrupts a model without ever looking wrong; (4) join rhea2*.tsv cross-references on MASTER_ID, not on the chosen directed ID, because an external database may have mapped to any quartet member, the published instance being UniProt returning several Rhea IDs per entry including UN and LR/RL together (Morgat et al. 2020). Obsolete IDs are marked rather than dropped so a collision with an old model file stays visible. Proposed on the assembly, strong at resource level (ELIXIR Core Data Resource, UniProtKB’s enzyme-annotation vocabulary — Bansal et al. 2022). Stdlib-only, so no ## Dependencies block; Fully open CC BY 4.0 and fully local after step 1, the deliberate privacy contrast with the connector path. Laptop; the 11 MB SMILES file download is the slow step, not the joins. Note for the next Chemistry pass: api.rhea-db.org/rhea/<id> returned 410 to two fetches this run and the /help/* pages 403, so the recipe deliberately depends on the FTP release files only and makes no claim about the JSON API.
• Trace an adverse outcome pathway from a molecular initiating event to an organ finding (added 2026-08-09) — rung-2 Adverse Outcome Pathway skill recipe discharging the top Drug Repurposing and Discovery deferred item (banked 2026-08-02). The banked “check first whether the skill returns curated AOP-Wiki links or reasons them out” question resolved cleanly in favour of retrieval: the upstream SKILL.md, fetched this run, names the exact calls (AOPWiki_list_aops → AOPWiki_get_aop → four parallel PubChemTox_* → CTD_get_chemical_gene_interactions/_diseases), so key events come back from the API rather than from the model — no heavy reasoned-not-retrieved caveat needed. Amend-vs-add against triage-a-compound-toxicology-and-hazard-profile resolved to add, and that page had already named this skill under Alternatives as “the right sibling”: the existing recipe is compound-keyed and produces a dossier, this one is finding-keyed (organ + endpoint + species) and produces a graded mechanism plus the one discriminating assay. Design puts nothing in the four-phase ordering (settled, supplied by the skill) and everything in four gates. Load-bearing: stressor absence is an explicit value, not a blank — 871 prototypical stressors across 595 AOPs and 1,939 key events (aopwiki.org/metrics_summary, fetched this run) means your compound being absent is the ordinary case, so stressor_match is listed/not_listed/aop_lists_no_stressors. Second, one row per key-event relationship, never per pathway, reporting the weakest link rather than an average, because the KER is the unit of causal knowledge that gets developed and reviewed on its own (Holmer et al. 2024). Third, OECD status recorded before the pathway content is read — an under-development AOP is one group’s draft, and the coaching/”gardening” programme exists because AOP-Wiki quality is uneven (Tanabe et al. 2025). Fourth, CTD direct_curated vs inferred split, since inferred chemical-disease links are transitive gene-chained constructs that CTD itself had to rank by network topology (King et al. 2012). Verdict is three-way defaulting to plausible_unsupported; hazard layer (GHS/IARC/LD50) deliberately quarantined in its own table so it cannot read as mechanistic corroboration; the page states plainly that an AOP is dose- and kinetics-free by construction. Proposed. Fully open, Laptop; the disclosure caveat is named with a followable degraded mode — steps 2 and 4–5 are chemical-agnostic, so a confidential compound can still get an endpoint-only pathway list provided the omission of steps 6–8 is recorded rather than left blank. AOP-Wiki release 2.8 (2026-03-08) is the provenance field to record.
• Design a group-sequential trial with interim looks (added 2026-08-09) — rung-2 Adaptive Designs skill recipe discharging the second-ranked Translational Medicine deferred item and opening the design-stage leg of the clinical-trials chain: the cookbook could draft a protocol document and analyse a trial’s missing endpoint data, and had nothing for the statistical design that sits between them. The banked amend-vs-add question resolved to add: the protocol recipe’s deliverable is a document, this page’s deliverable is three simulated tables plus an SAP section, and the reader arrives before enrolment rather than during drafting. Design puts nothing in the boundary arithmetic (settled, rpact) and everything in three gates. First, the fixed-sample comparator is computed and committed before any interim look is added (design/00_fixed_reference.R), so “the adaptive design is better” is a measured claim rather than an assumption. Second, operating characteristics are simulated across an effect grid, never at the design point, at a recorded seed and replicate count — Monte-Carlo error at 1,000 replicates is ~1 percentage point on a stopping probability, enough to reorder two close designs. Third and load-bearing, SSR defaults to “no” and must be justified against the simulation, with a hard presumption against it for time-to-event endpoints: the promising-zone rule triggers where the marginal power per patient is not greatest and over-inflates n across a narrow band (Jennison & Turnbull 2015); optimising conditional vs unconditional power yields different rules so the SAP must say which (Mehta et al. 2022); and in survival trials SSR lets interim curve-shape information enlarge the observed effect and raises the chance of recommending an ineffective therapy, so it “remains unjustified” there (Freidlin & Korn 2017). Supporting rules: futility boundary’s binding status recorded as an explicit field (teams take the alpha then cross the boundary anyway); protocol-must-specify checklist follows the explicit division in Shih, Li & Wang 2016, which also shows dual-test and promising-zone sample-size rules conflict; bias-adjusted estimate and repeated CIs generated in the script rather than promised in the SAP; ICH E20 flagged as draft (Step 2b) wherever cited. Proposed. Fully open with the LGPL-3 rpact / GPL-3 gsDesign copyleft flag named, fully local (confidential development plans fine), Laptop — with the honest note that the survival simulation, not the boundary computation, is the only real cost. rpact 4.4.0 (2026-03-04, LGPL-3) confirmed by fetch this run. Bayesian Trials named under Alternatives as the different-framework route rather than a second component, keeping the ladder at 2.
• Measure cortical thickness and subcortical volumes across a cohort (added 2026-08-09) — rung-2 FreeSurfer skill recipe discharging the banked “next-but-one Neuroscience pick” and opening the structural-MRI morphometry leg: the Neuroscience section had DTI, resting-state fMRI, EEG/ERP, calcium, spikes, pose and BIDS conversion, and nothing for the single most common structural-MRI ask. Written instead of the top-priority OpenNeuro pick, which is dead — see the flag below. Design deliberately puts almost nothing in the recon-all invocation (the settled part, supplied by the skill) and everything in the two gates that leave no trace in the output table. First, the FreeSurfer version is pinned in config.yaml before any subject runs, with a mid-cohort upgrade defined as “reprocess everyone”, because 6–12 h/subject means cohorts get processed in waves across months: v5.0.0 moved volumes 8.8 ± 6.6% and thickness 2.8 ± 1.3% versus earlier releases, with workstation type and even the macOS point release worth about half as much again (Gronenschild et al. 2012), and the modern replication found cingulate thickness at ICC 0.37–0.61 across v5.3/6.0/7.1 plus version differences in downstream statistical results (Haddad et al. 2023). The page deliberately surfaces the counterweight rather than hiding it — 1.6–5.8% absolute shifts had very little effect on AD/MCI classification (Chepkoech et al. 2016) — and resolves the two: within-version inter-subject stability is good, which is why one version per cohort suffices and why mixing versions across groups does not. Second, the Euler number is extracted and gated on before any group difference is computed, threshold as a recorded literal chosen pre-labels, excluded IDs written out (AUC 0.98–0.99 against expert “unusable” ratings, out-performing functional-timeseries proxies from the same session; quality both inflates and obscures age associations — Rosen et al. 2018) — and then carried into the model as a covariate anyway, because thresholding attenuates but does not eliminate the effect and how you control for quality changes ABIDE case-control conclusions (Bedford et al. 2023). Supporting rules: thickness and surface area kept as separate analyses (distinct measures, and their version-compatibility profiles differ); eTIV covariate for volumes only; hard abort on heterogeneous acquisition without a site column. Proposed. Compute tagged HPC or cloud cluster with the honest N-dependence spelled out (10 subjects = a workstation overnight; 60 = an array job — and it is exactly that duration that creates the version-drift risk); noted that GPU buys nothing for the classic pipeline. Fully open but with the free-registration license.txt gate named concretely (do it before scheduling the array, not after) and the download page’s silence on commercial terms flagged. FreeSurfer 8.2.0 (March 2026) / 7.4.1 confirmed by fetch this run. Harmonization is named under Alternatives as the multi-site addition rather than a second rung, keeping the ladder at 2.
• Find differentially abundant proteins in a label-free proteomics experiment (added 2026-08-08) — rung-2 MaxQuant skill recipe opening the cookbook’s entirely empty proteomics leg, the largest remaining Molecular and Cellular Biology hole and the oldest MCB deferred item: the transcriptomics leg had six recipes and proteomics had none. Design turns on two gates rather than on the arithmetic. First, the three-flag filter is written as mandatory and non-configurable (Potential contaminant, Reverse, Only identified by site), with per-filter removal counts logged — every one of those rows passes a t-test if left in, and a keratin hit is indistinguishable from a real one in the output. Second and load-bearing, the differential test runs twice — with the declared imputation and on complete cases only — emitting an imputation_sensitive flag per protein, because the benchmark position is not “pick a better imputer” but that imputation may not help detection at all: “the most commonly used methods are among the worst performing” and imputation “does not necessarily improve the ability to identify differentially expressed peptides” (Harris et al., JPR 2023); accuracy is driven by the MNAR rate, which you cannot measure, and downstream analysis is largely impacted by the method chosen (Jin et al., Sci Rep 2021); the field is still moving (Etourneau et al., Biostatistics 2025). Supporting rules: LFQ columns only (not Intensity, not iBAQ); Majority protein IDs retained so protein groups stay groups and the follow-up is designed against the group, not the leading razor ID; on/off proteins split into their own table since imputation manufactures a finite fold change for them; valid-value rule per group, not pooled; mqpar.xml committed as the most under-shared file in proteomics. The quantification layer is the one part that is settled — MaxLFQ recovers known mixing ratios across the expression range (Cox et al., MCP 2014) — so the uncertainty is deliberately located downstream of the intensities. Proposed. Availability nuance worth remembering: tagged Fully open (no subscription, account, or institutional gate) but the body states plainly that MaxQuant is free-of-charge, not open-source — MPI terms accepted at download, Bioconda flags it license-restricted, so redistribution inside a published container is not permitted. Platform gate stated concretely: Windows 10/11 or Ubuntu 20.04+, .NET 8.0 since v2.6.3.0, no macOS build and no Linux GUI (headless MaxQuantCmd.dll). Laptop tier with the honest caveat that large parts of the search are single-threaded regardless of core count, so fast disk beats cores. PRIDE is deliberately not a second rung — it is step 0 for a reanalysis, named under Alternatives, keeping the recipe at rung 2.
• Decide whether a suspicious cluster is a doublet artifact (added 2026-08-08) — rung-2 Doublet Detection skill recipe, resolving the banked amend-vs-add question in favour of add: the QC recipe removes doublets as one step on the way to a filtered object, whereas this reader already has a clustering and one specific intermediate population, and wants a verdict. Different deliverable, different entry point. Design: expected rate from the lane recovered-cell count, not the demultiplexed subset (the commonest setup error, and it biases low) → score per sample on raw counts before integration, with the script refusing a merged object, since simulating doublets from a pool of cells that never shared a droplet is not the null → two detectors kept as separate columns, never averaged → non-bimodal Scrublet histogram flagged and the cutoff recorded as a literal → calls aggregated to cluster_doublet_summary.csv, which is the object the reader’s question is actually about (the skill scores cells; the recipe’s own addition is the cluster-level roll-up) → two orthogonal checks the scores do not use: library-size doubling toward the sum of the parent medians, and additivity of the two lineage programs at near-parental strength versus one attenuating and one rising → three-way doublet_artifact/real_population/undetermined defaulting to undetermined, with the explicit statement that a clean negative is not proof because homotypic doublets are largely undetectable by any of these methods → re-cluster from scratch after removal, never reuse the old labels. Proposed on the adjudication, well grounded on the detection: nine methods over 16 experimentally-annotated real datasets plus 112 synthetic ones showed “diverse performance and distinct advantages in different aspects”, DoubletFinder most accurate and cxds most efficient (Xi & Li, Cell Systems 2021) — the page surfaces the tension that the benchmark’s accuracy winner is DoubletFinder while the skill’s stated 2024–2026 default is scDblFinder, with no comparable modern benchmark to arbitrate, as a further argument against one score deciding; step-interaction effects motivating the re-cluster rule come from Germain et al., Genome Biol 2020; steps 4–5 are flagged in Evidence as unbenchmarked field practice, included because they are orthogonal to the score rather than because their sensitivity is known. SNP demultiplexing named under Alternatives as the genuinely stronger evidence where samples were pooled (80–85% accuracy, degrading as doublet rate rises — Fu et al., Brief Bioinform 2025) but no such tool is catalogued, so no followable path is offered. Fully open, fully local, Laptop.
• Characterize a protein–protein interface from a structure you already have (added 2026-08-08) — rung-2 Interface Analysis skill recipe discharging the oldest ISCB deferred item and opening the analyse-an-existing-complex leg: the cookbook could predict a complex (co-folding, antibody–antigen) and had nothing for a structure already in hand. Design forces the two decisions everyone skips: biological assembly, not asymmetric unit — assembly ID recorded, with a hard stop if it cannot be established, because everything downstream inherits that error — and cutoff declared before the run, 4.5 Å heavy-atom primary with an 8 Å Cα column emitted alongside so the definition-sensitivity of the residue list is visible rather than hidden in a default. contacts.csv (residue pairs, the mechanistic object) and interface_residues.csv (per residue, the cloning object) kept as two tables, because merging duplicates each residue once per partner contact and inflates every count taken from it; the SASA rule is that complex and both isolated chains use an identical probe radius and atom selection, a mismatch being invisible in the resulting BSA number. The load-bearing step is step 6: a three-way biological/ambiguous/likely_packing verdict that defaults to ambiguous, with orthogonal evidence (Complex Portal/IntAct annotation, solution data) recorded as its own line rather than folded into the call. Proposed — skill unbenchmarked end-to-end (verified: works establishes the workflow runs, not that the verdict is calibrated) — but every gate traces to a result: ROC 0.923, 91.4%/91.7% test accuracy, outperforming DiMoVo/Pita/PISA/EPPIC (Luo et al., Proteins 2014) is the ceiling, high and not high enough for a bare verdict; up to 94.8% on 172 packing-vs-functional complexes (Block et al. 2006); ~88.8% from a discriminator its authors note does not correlate with contact area (Tsuchiya et al. 2008) — the direct reason BSA alone is not the answer; and the CAPRI assessment that scoring “still does not distinguish well biological from crystal packing interfaces” (Kundrotas et al. 2018) — the reason the default is ambiguous rather than biological. Rung-1 refusal stated concretely: Claude Code alone will emit a fluent, confidently wrong interface residue list from training memory, undetectable by reading it. Fully open, Laptop, fully local except the PDBePISA cross-check, which is named as the one step that discloses a confidential structure, with the instruction to record its omission rather than let a missing input read as a negative.
• Test whether a proposed RNA structure is actually conserved (added 2026-08-08) — rung-2 Covariation Analysis skill (R-scape) recipe, the validation counterpart the RNA leg lacked: a thermodynamic folder returns a structure for any sequence including a shuffled one, so predict-rna-secondary-structure-and-accessibility could produce structures and nothing could test one. Design: alignment built independently of the structure as the load-bearing gate — structure-guided or covariation-guided alignment construction followed by a covariation test on that same alignment is a documented artifact, and the recipe voids the test rather than caveating it → power analysis run with the test, never after, with the rejected verdict simply unavailable below the skill’s ~10% mean-power line → three-way supported/rejected/undetermined, with the mean power written into the verdict line itself so the conclusion cannot be quoted without its caveat (the failure this recipe exists to prevent is a low-power negative circulating as a refutation) → proposed-pair and alternative-pair covariation in separate files, keeping “structure rejected” and “different structure supported” distinguishable outcomes → helix-level aggregated E-values reported alongside per-pair, R-scape ≥ 2.0.0.p required, since per-pair-only reporting discards structures whose signal lives in the stack → --cacofold explicitly barred from being reported as validation of its own consensus → provenance.json foregrounding average pairwise identity, mean power, and the Rfam release, plus how the alignment was built and with which aligner. Proposed on the assembly, but the method is unusually well characterized on exactly this question and each gate cites one paper: no significant support for the proposed HOTAIR/SRA/Xist structures (Rivas, Clements & Eddy, Nat Methods 2017); power estimation separating an adequately powered negative from an uninformative one (Bioinformatics 2020); helix-level aggregation raising sensitivity without sacrificing specificity, and the circularity artifact (Rivas, PLoS Comput Biol 2023). Boundary kept explicit in the page: this tests whether a structure is conserved, not whether the transcript is real, expressed, or functional. Fully open, Laptop, fully local (unpublished sequences fine); the one friction noted is that R-scape installs via conda, not pip. No ## Dependencies block — R-scape is the component’s own prerequisite and belongs on its catalog page.
• Test which immune populations differ between groups in a cytometry cohort (added 2026-08-08) — rung-3 chain of three bioSkills flow-cytometry skills (Compensation and Transformation → Clustering and Phenotyping → Cytometry Differential Analysis), opening a leg where all five catalogued cytometry skills were uncovered: the only cytometry recipe was parse-fcs-flow-cytometry-files, which is deliberately I/O-only, so the cookbook could parse an FCS file and then had nothing to say about the actual question. Design: panel.csv + metadata.csv written before any analysis, with marker_class type/state as the load-bearing field, because clustering on an activation channel splits one lineage into activation states and makes every downstream abundance claim uninterpretable → compensate-then-transform enforced (compensation is linear and mathematically invalid after a nonlinear transform) → FlowSOM on type markers only at a pinned seed → counts.csv (one row per sample × cluster) emitted as a committed intermediate and named the audit trail, which is what closes the pseudoreplication trap mechanically rather than by instruction → DA and DS as two separate tables that are never merged (“more activated CD8s” is two different claims needing different tests on different statistics) → low_support label on clusters whose median is <50 cells/sample rather than silent dropping → hard abort into untestable.md when a group has <3 biological replicates, since that design has no error term → every DA hit re-checked for compositionality with compositional_confirmed/compositional_suspect, because one expanding population arithmetically depresses the rest → batch as a design covariate, explicitly not normalized out → provenance.json foregrounding the design formula as a literal string, the field reviewers ask about and nobody records. Rung-3 escalation justified on ordering constraints rather than breadth, with the three skills sharing one installer and one SCE object; Cytometry QC named as a conditional fourth step (long/multi-batch runs) to keep the core toolbelt at three. Proposed — no agent-driven run documented, but diffcyt reports improved statistical performance including for rare populations (Weber et al., Commun Biol 2019), and the pseudoreplication gate rests on the sharpest available quantification of the failure — methods ignoring biological-replicate variation “can discover hundreds of differentially expressed genes in the absence of biological differences” (Squair et al., Nat Commun 2021) — with the transcriptomics→cytometry transfer stated rather than glossed. Fully open and fully local (usable on unpublished clinical cohorts), with the edgeR/limma commercial-redistribution restriction flagged. Laptop, RAM not CPU the binding constraint, plus the subsampling rule that FlowSOM may be subsampled but cluster assignment and counts.csv must not.
• Find which taxa differ between microbiome groups (added 2026-08-08) — rung-2 Differential Abundance skill recipe completing the microbiome leg: compute-16s-microbiome-diversity answered “do the communities differ” and nothing answered “which taxon”, the question that actually reaches the abstract. Design: prevalence filter declared as a recorded parameter, never a default, with filtered_features.csv listing dropped features as well as survivors and an explicit bar on re-tuning it after seeing results → ≥2 methods on the identical filtered table (ALDEx2 + ANCOM-BC2 default; MaAsLin2/LinDA when the design needs covariates or random effects) → da_by_method.csv one row per (feature, method) with the method’s own diagnostic flag, never pooled or averaged → da_consensus.csv tiered high_confidence (intersection) vs exploratory (union minus intersection), with the instruction to report it that way, since a taxon one method calls and another does not is a real observation about borderline evidence rather than a tie to break by picking the friendlier tool → effect-size floor alongside BH/FDR → seed and mc.samples recorded (ALDEx2 is Monte-Carlo) → DESeq2/edgeR refused on compositional amplicon data → a one-line “relative abundance was higher”, not “increased”, discipline absent a microbial-load anchor → provenance.json recording the taxonomy reference release (SILVA 138 vs 138.1 relabels genera and silently renames your findings). Proposed, but the consensus deliverable is the benchmark’s own recommendation: 14 methods across 38 datasets “identified drastically different numbers and sets of significant ASVs”, hit counts tracking sample size and depth rather than biology, ALDEx2 and ANCOM-II most consistent (Nearing et al., Nat Commun 2022) — with a pre-registered synthetic-data re-validation flagged as live (Kohnert & Kreutz, F1000Research) and a note that the default method pair should change if it revises the ranking. Scoped to amplicon deliberately: the skill routes shotgun DA to the metagenomics category, so the Kraken2 recipe is cross-linked with that caveat rather than as a substitute. Fully open, fully local (IRB-restricted human cohorts), Laptop with mc.samples and ZicoSeq permutations named as the only real cost knobs.
• Rank the conformers and tautomers of a small molecule with semi-empirical QM (added 2026-08-08) — rung-2 ChemGraph recipe opening the cookbook’s entirely empty quantum-chemistry leg: every existing Chemistry recipe (docking, MD, ADMET, SAR, protonation) consumes a 3D structure and none of them produced one on energetic grounds. Design: tautomers enumerated before conformers, never the reverse (a tautomer is a different bond topology, so conformers of the drawn form do not transfer) → enumerate_states.py with RDKit TautomerEnumerator + ETKDGv3 at a fixed seed → ChemGraph run_ase optimization followed by a mandatory frequency job, with n_imag > 0 routed to rejected_states.csv (a converged saddle point otherwise sits in the table looking like a real state) → ranking on G(298.15 K), not electronic energy, because ZPE and entropy flip tautomer orderings more often than expected → a too_close_to_call flag on anything within ~1 kcal/mol, since that gap is inside GFN2-xTB’s error bar and the honest answer is to carry both states downstream → provenance.json foregrounding the solvation model and solvent, the field that silently decides tautomer ordering if left undeclared. Proposed — assembly undocumented, but ChemGraph is benchmarked across 13 tasks with the finding that decomposing complex tasks into subtasks lets smaller models match larger ones (Pham, Tanikanti & Keçeli, arXiv 2506.06363) — the direct argument for committing a script rather than one open-ended request; GFN2-xTB’s authors describe it as “well-suited to explore the conformational space” (Bannwarth, Ehlert & Grimme 2019), a claim about narrowing rather than deciding, which is the basis for the sub-kcal/mol gate; and LLM-agent-orchestrated AIMD/DFT is now a documented pattern (Shen et al., Chem Sci 2026, Lahouari et al., JCTC 2026). One declared dependency (rdkit==2026.3.5, BSD-3-Clause, import rdkit, fetched from PyPI this run). Fully open and fully local — the deliberate contrast with the cloud-side prepare-protonation-states-for-docking, with the one leak (molecule_name_to_smiles → PubChem) called out. Laptop; disk, not CPU, is the surprise.

Flagged for review

• Discover NWB recordings on DANDI and prepare them for sorting — amended 2026-08-09, no further action expected. Its optional step 7 recommended installing the OpenNeuro MCP, whose catalog page has carried verification: broken since 2026-07-20 (repo QuentinCody/open-neuro-mcp-server and the hosted workers.dev/sse endpoint both 404 across four runs, no fixable path). Step 7 now states the breakage and points at manual openneuro.org search instead, and the “plus optionally the OpenNeuro MCP” clause was dropped from Why this assembly so the rung-3 justification rests only on the two working components. The recipe’s core Neurosift + neuropixels path is unaffected, availability is unchanged, and last_verified was not bumped (this run was recipe_recheck: no; the rest of the page has not been re-verified). Re-check at the next recheck slot.

Deferred — next-run priority

• Define populations by a reproducible gating hierarchy rather than clustering — Gating Analysis is catalogued and uncovered, and the new test-cytometry-clusters-for-group-differences recipe names it under Alternatives as the substitute route for step 4 without giving a followable path. The framing that earns a page rather than an amendment is GatingML-serialized, version-controlled gates whose results are comparable to historical gated data from the same lab — a genuinely different deliverable from unsupervised clusters, not just a different algorithm. Decide amend-vs-add before writing: if it is only “swap step 4”, amend the new recipe instead. Top Immunology and Microbiology pick for the next pass (promoted 2026-08-15 when the bulk-repertoire item above it was discharged as compare-bulk-tcr-repertoires-across-samples).
• Profile metagenome function and track strains across timepoints — MetaPhlAn, Functional Profiling (HUMAnN-class pathway abundance) and Strain Tracking (StrainPhlAn-class) are all catalogued and uncovered, and the shotgun leg currently stops at Kraken2 taxa. Two candidate pages hide here: functional profiling (pathways rather than taxa — the “what is the community doing” question the cookbook cannot answer at all), and strain tracking (does the donor strain persist after FMT / does the same strain recur in a relapse), which is the sharper and more clinically motivated of the two. Note the new find-taxa-differing-between-microbiome-groups recipe is deliberately scoped to amplicon because the skill routes shotgun DA elsewhere, so a shotgun differential page is a distinct third candidate. Pick one framing, not all three.
• Look up the biochemistry of a metabolite or enzymatic reaction (Rhea + ChEBI) — discharged 2026-08-15 as get-balanced-reactions-for-a-metabolite; the banked recipe-vs-FBA-step question resolved to its own recipe, and the rung resolved to 1 rather than 2 because both components are chat-only connectors. Drop this entry next run. Original note follows: both catalogued, both uncovered, and the pair is the natural Chemistry-side complement to the metabolic-model recipe predict-gene-knockout-phenotypes-with-fba, which consumes reaction stoichiometry it never sources. Surfaced on the 2026-08-08 Chemistry pass and deferred behind the two shipped pages. The framing that would earn a page is “given a metabolite or an EC number, get the balanced reactions, directionality, and cross-references I can put in a model” — decide first whether that is a recipe or a step the FBA recipe should gain. Top Chemistry pick for the next pass.
• Escalate a shortlist of xTB-ranked states to DFT — the explicit out-of-scope tail of the new rank-conformers-and-tautomers-with-xtb recipe. ChemGraph reaches NWChem and ORCA, so the components exist, but the honest page has to deal with basis-set and functional selection, wall-clock on a workstation or cluster, and the fact that a DFT job requested casually in conversation burns real core-hours — that needs its own slot rather than the tail of one. Second Chemistry pick.
• Trace an adverse outcome pathway from a molecular initiating event to an apical endpoint — DISCHARGED 2026-08-09 by trace-adverse-outcome-pathway-for-a-toxicity (rung-2, Proposed, Fully open, Laptop). Do not re-file. The banked reasoned-vs-retrieved worry resolved in favour of retrieval — the upstream SKILL.md names the AOPWiki/PubChemTox/CTD calls explicitly, so no heavy caveat was needed. Three things to know if the page is revised: the amend-vs-add call against triage-a-compound-toxicology-and-hazard-profile turned on finding-keyed vs compound-keyed (that page had already named this skill as “the right sibling”), the load-bearing gate is the three-value stressor_match field justified by AOP-Wiki’s own metrics (871 prototypical stressors across 595 AOPs), and the dose/kinetics disclaimer is deliberately in the body rather than a footnote. A quantitative AOP (qAOP) / response-response modelling page is the natural sequel and is not covered — check whether any catalogued component reaches qAOP content before filing it.
• Find a selective tool compound for a target with curated pharmacology (GtoPdb) — DISCHARGED 2026-08-16 as find-a-selective-tool-compound-for-a-target (rung-2, Proposed, Fully open, Laptop). Do not re-file. The banked amend-vs-add question resolved to add, exactly as predicted — target → ranked candidates with a selectivity window is a different deliverable and a different reader from compound → targets — and both pages are now cross-linked, with the ChEMBL page serving as the new recipe’s step-8 second source. Four things to know if the page is revised. (1) The load-bearing gate is the species filter, grounded on cross-species affinity deviations of up to 3.2 log units (~1 600-fold) and recombinant-vs-native-tissue deviations up to 4.2 log units (Bindel & Seifert 2026); GtoPdb interleaves Human and Rat rows with nothing else to distinguish them, verified on the API. (2) The verdict defaults to insufficiently_profiled because absence of off-target data is not selectivity — the same paper found 69/302 histamine ligands with missing or incomplete functional characterization, and GtoPdb is a curated representative selection (3,103 targets / 13,260 ligands, Harding et al. 2026). (3) affinity is a string that can be empty and affinityParameter mixes pKi with pIC50 in the same result set — both verified on live responses, both silent corrupters. (4) The API carries no release version, so provenance records the query date and returned counts; do not invent a release number. The rung is stated with an honest boundary — the artifact is a stdlib script against a public unauthenticated API, so the skill earns rung 2 on its endpoint map and its two documented traps (name= is fuzzy; approved=true is silently ignored), not on doing arithmetic.
• Find a repurposing candidate’s curated mechanism gap — approved drugs whose GtoPdb primary target is not the target you care about — surfaced 2026-08-16 while writing the tool-compound page and now the top Drug Repurposing and Discovery pick. The new recipe fetches primaryTarget and the curator selectivity free-text note but deliberately carries them as evidence columns rather than acting on them; the untapped question is the inverse — GET /services/ligands?type=Approved intersected against a target’s interaction list, filtered to ligands whose primary annotated target is something else, which is a curated, literature-traceable repurposing shortlist with an affinity number attached. It is a genuinely different deliverable from scan-drug-repurposing-candidates (target-driven bulk lookup) and from find-repurposing-candidates-in-a-knowledge-graph (graph traversal over PrimeKG), because the evidence unit is a curated affinity with a PMID rather than a graph edge or a bulk hit. Decide amend-vs-add against both before writing. Verified facts to reuse: type=Approved is the server-side alias and approved=true is silently ignored; the 2026 release report describes an improved approved-drugs presentation with new therapeutic targets and “modulators with clinical potential” across multiple approval agencies (Harding et al. 2026) — check whether that set is separately addressable in the API before relying on it.
• Resolve a compound identifier across chemical databases — UniChem and Guide to PHARMACOLOGY are both catalogued and uncovered. Deferred because the honest question is whether identifier resolution is a step other recipes should gain rather than a recipe of its own — the new sourcing recipe already carries a SMILES-not-name rule and a smiles_mismatch.csv, and ChEMBL-based recipes hit the same problem. Decide amend-vs-add before writing. GtoPdb’s distinct value is curated pharmacology (ligand-target pairs with quantitative activity read from primary literature), which may be the sharper page.
• Do not re-file: Indication Dossier — evaluated and rejected 2026-08-02 as a recipe candidate. It is a Claude Science-only skill with no separately addressable install path, so a recipe could not converge on a committed artifact — it would be the cookbook’s first artifact-less page, the same blocker already recorded against the FinnGen item below. It also overlaps build-target-dossier on substance. Revisit only if the skill becomes separately installable.
• Look up a variant or gene’s disease associations across ancestries (FinnGen + BioBank Japan + GWAS Catalog) — SUPERSEDED 2026-08-16 by the sharpened “Cross-ancestry replication of a candidate association” item below, which carries this run’s verified access facts and resolves the connector caveat banked here (the GWAS Catalog skill is separately installable, so the recipe is rung 2 and not artifact-less). Use that entry; drop this one next run.
• Design an adaptive trial (interim analyses, sample-size re-estimation) — DISCHARGED 2026-08-09 by design-a-group-sequential-trial-with-interim-looks (rung-2, Proposed, Fully open, Laptop). Do not re-file. The banked amend-vs-add question resolved to add, not a branch of the protocol recipe: different deliverable (three simulated tables + an SAP section vs a protocol document) and a different entry point (before enrolment vs during drafting). Two things to know if the page is revised: it is deliberately framed as group-sequential first, SSR only if the simulation earns it — with a standing presumption against SSR for time-to-event endpoints (Freidlin & Korn 2017) — and the fixed-sample comparator in design/00_fixed_reference.R is the load-bearing artifact, because without it “the adaptive design is better” is untested. Bayesian Trials is named under Alternatives, not as a second component; a genuinely Bayesian design page (predictive-probability stopping, borrowing priors, CRM/BOIN dose finding) is a legitimate separate candidate and is now the second Translational Medicine pick (see the next item).
• Design a Bayesian trial (predictive-probability stopping, borrowing priors, CRM/BOIN dose finding) — DISCHARGED 2026-08-16 by design-a-bayesian-trial-that-borrows-external-data (rung-2, Proposed, Fully open, Laptop). Do not re-file. The banked framing held exactly: the page is the prior-elicitation and borrowing decision, not “same design, different framework”. Three things to know if it is revised. (1) The load-bearing artifact is design/00_no_borrowing_reference.R — committed before the historical data are opened, following the group-sequential page’s fixed-sample-reference precedent; remove it and “borrowing saved patients” becomes untested. (2) The declared unit is a prior effective sample size in patients, written to map_prior.json before operating characteristics are computed — that is the number a reviewer or DSMB can argue about, and “we used a MAP prior” is not a specification. (3) Operating characteristics run over a two-way true-effect × prior-data-conflict grid, never only under exchangeability, because robustification alleviates conflict rather than removing it (Schmidli et al. 2014). The predictive-probability-stopping half of this banked item is not covered — the shipped page is scoped to borrowing and dose-finding boundaries — and I-SPY 2-style graduation at posterior predictive probability ≥ 0.85 is on the Bayesian Trials catalog page but has no recipe; check amend-vs-add against the shipped page before writing it, since it may only be a fourth section rather than a page. The January 2026 FDA Bayesian draft guidance is flagged as a draft wherever cited, per that catalog page.
• Cross-ancestry replication of a candidate association (FinnGen + BioBank Japan + GWAS Catalog) — now the top Translational Medicine pick, promoted 2026-08-16 as the Bayesian item above was discharged. Picked up and not written this run for a budget reason, with the reconnaissance worth banking. The connector caveat is resolved and is not the blocker it looked like: GWAS Catalog is a genuine installable Claude Skill (SciAgent, Apache-2.0, runs Python locally via Bash, verification: works), so the recipe is rung 2, not the artifact-less rung-1 fallback the older note feared; FinnGen and BioBank Japan supply the data, not the tooling. Access facts verified 2026-08-16: FinnGen is at DF13, released 2026-06-02 — 500,186 participants, 21,311,644 variants, 2,755 endpoints — browsable at r13.finngen.fi, but bulk summary-statistic download is gated behind a Helsinki e-form (elomake.helsinki.fi/lomakkeet/124935/lomake.html) with instructions returned by email, free of charge; treat that like the FreeSurfer license.txt precedent (Fully open with the gate named concretely) and do the form before planning the run. FinnGen reports all effect sizes and allele frequencies on the alternate allele — verified on finngen.fi/en/access_results this run, and the single most important fact for the recipe. BBJ is at pheweb.jp with a /downloads page linking humandbs.dbcls.jp/files/hum0197/hum0197.v3.BBJ.<PHENOTYPE>.v1.zip, ~170–178k samples per phenotype, 220 traits, cite Sakaue & Kanai 2021 Nat Genet. Still unverified and required before writing: the genome build of each cohort’s files (FinnGen DF13 is almost certainly GRCh38 and BBJ hum0197 almost certainly GRCh37, but neither was confirmed — finngen.gitbook.io/documentation/r13 returned a content-free landing page to WebFetch, so go to its Data description / Methods subpages) and the exact column headers on both sides. The three gates the page should be built on: (1) effect-allele harmonization with a palindromic-SNP (A/T, C/G) hard abort or AF-based resolution — a silent A1/A2 mismatch flips the sign of beta and turns replication into “opposite direction”, which reads as a real biological finding; (2) power to replicate computed from the replication cohort’s own MAF at the discovery effect size, before the replication p-value is read, with the not_replicated verdict unavailable below a pre-registered power floor (the R-scape power-gate precedent in test-rna-structure-for-covariation-support); (3) Finnish founder enrichment stated as the expected case, not an anomaly — a bottleneck-enriched low-frequency variant is predicted to be near-monomorphic in an East-Asian cohort, so its absence is not evidence against the locus. Note two papers MCP searches on cross-ancestry replication power returned nothing this run; find the methods citation by another route.
• Assemble a CONSORT-compliant trial report — rung-2 fit for the uncovered Trial Reporting skill. Deliberately not written this run: it overlaps the new missing-data recipe’s MMRM/rbmi machinery, and the catalog page itself says the two are distinguished by whether the missing-data package or the wider report is the deliverable. Only worth a page if the framing is squarely Manuscript prep (flow diagram, CONSORT checklist, results tables) rather than a second pass at the same statistics.
• Quantify a label-free proteomics experiment from raw spectra to a protein table — DISCHARGED 2026-08-08 by find-differentially-abundant-proteins-in-lfq-ms. Do not re-file. Two things to know if the page is revised: it landed at rung 2, not the anticipated rung 3 — PRIDE is step 0 for a reanalysis rather than a second component, so it is named under Alternatives and the ladder stays at one skill; and the page is deliberately built around the contaminant/decoy/site filter and the paired imputed / complete-case differential run rather than around the search, because the search is the settled part (Cox et al. 2014) and the imputation is not (Harris et al. 2023).
• Post-translational modification site analysis (phosphosite differential) — DISCHARGED 2026-08-15 by find-differentially-regulated-phosphosites (rung-2, Proposed, Fully open, Laptop). Do not re-file. The banked amend-vs-add question resolved to add. Three things to know if the page is revised. (1) The banked claim that the MaxQuant skill “emits Phospho (STY)Sites.txt-class tables” is wrong — the upstream SKILL.md, fetched 2026-08-15, stops at proteinGroups.txt with no mention of site tables, localization probability or PTM analysis; the site-level layer in the recipe is composed on top of MaxQuant’s own output and the page says so. (2) The rung-2 call rests on the paired unenriched proteome arm, which is the protein-level pipeline the skill documents — remove that arm and the skill stops being load-bearing and the page drops to rung 1. (3) The load-bearing gate is the protein normalization, and it is deliberately reported alongside the raw statistic rather than replacing it, because the confound’s magnitude is evidenced (Esnault et al. 2018) but the correction is not benchmarked.
• The TMT / labelled proteomics path — the other half of the discharged item above, still open. The MaxQuant skill documents SILAC as well as LFQ, and TMT brings gates neither shipped proteomics recipe carries: reporter-ion ratio compression, the reference-channel / bridge design, and the fact that missing values behave completely differently from LFQ (near-complete quantification within a plex, systematic absence across plexes) — which means the LFQ page’s imputation machinery does not transfer. Check what the SKILL.md actually documents for TMT before writing; the phospho pass this run showed the banked description of the skill’s coverage was optimistic. Top Molecular and Cellular Biology pick for the next pass.
• Detect doublets before trusting a cluster — DISCHARGED 2026-08-08 by decide-if-a-cluster-is-a-doublet-artifact. The banked amend-vs-add question resolved to add: the QC recipe’s deliverable is a filtered object, this page’s deliverable is a three-way verdict about one named cluster, and the reader arrives after clustering rather than before. Do not re-file. The Harmony half of this item is still open — Harmony versus the existing integrate-single-cell-datasets recipe is a separate amend-vs-add call and has not been made; note the doublet precedent suggests the test is different deliverable, different entry point, not different algorithm.
• Annotate cell types with a foundation model and check it against markers — rung-2 fit for the uncovered scGPT; the interesting framing is not “run scGPT” but the head-to-head against the reference-based annotate-cell-types-in-single-cell-data path, with disagreement kept as a column. Deferred because the same additive-baseline scepticism the new Perturb-seq recipe applies to perturbation-prediction models (Ahlmann-Eltze et al. 2025) needs a cell-type-annotation-specific benchmark before the page can say anything defensible about when the foundation model wins.
• Query MGnify for metagenome assemblies and functional annotations matching a biome — rung-2 fit for the MGnify tool; the public-repository discovery counterpart to the existing Kraken2 and 16S-diversity recipes (which both assume you already have reads). Deferred this run for cap reasons.
• Select the best experimental PDB structure for a target before docking — LARGELY DISCHARGED 2026-08-01 by the new vet-a-pdb-structure-before-reusing-it recipe, whose step 2 is exactly the PDB MCP search_by_uniprot candidate-enumeration pre-flight (region coverage, apo vs holo, construct mutations, deposition year), with the quality judgement handled by the Structure Validation skill rather than by reading resolution alone. No separate recipe needed; do not re-file.
• Characterize a protein-protein interface from an existing complex — DISCHARGED 2026-08-08 by the new characterize-a-protein-protein-interface recipe (rung-2 Interface Analysis skill, Proposed, Fully open, Laptop). Do not re-file. Two design decisions worth remembering if the page is revised: the recipe is built around the biological-assembly and cutoff-declaration gates rather than around the contact arithmetic (which is uncontroversial), and its biological-vs-packing verdict is deliberately three-way and defaults to ambiguous — grounded on the fact that the best published classifiers top out near 90% (Luo et al. 2014) and CAPRI assessors still report scoring functions cannot separate the two classes (Kundrotas et al. 2018). If a local druggability/interface classifier ever gets catalogued, the PDBePISA disclosure caveat in the Availability section is the thing it would remove.
• Look up the known composition of a protein complex — DISCHARGED 2026-08-15 by look-up-curated-complex-composition. Do not re-file. The two-framings choice resolved to the narrower one (“is this a curated complex, and at what stoichiometry”), exactly as this note predicted, because it is the interface recipe’s actual dependency; that page now links the new recipe. The rung resolved to 1, not the anticipated 2 — both components are chat-only Claude Science connectors with no addressable MCP URL, so the artifact has to come from the CC0 ComplexTAB FTP files, following the Rhea precedent. The design-facing “which subunits do I co-express / co-fold” framing is still uncovered and is a legitimate separate candidate, though the new page’s subunits.csv gets most of the way there. Original note follows: rung-2 fit for the still-uncovered Complex Portal and IntAct entries (curated stoichiometry and membership, as opposed to the inferred edges in build-ppi-network-and-rank-hub-genes). Now the top ISCB pick, and its framing sharpened by this run: the new interface recipe cites both databases in step 6 as the orthogonal-evidence line behind its packing verdict, but only as a link — a reader who needs to actually query them has no followable path, which is the concrete gap a page would close. Two candidate framings, pick one: “which subunits do I need to co-express / co-fold” (design-facing, feeds the co-folding recipe), or “is this pair a curated complex, and at what stoichiometry” (the interface recipe’s actual dependency). The second is narrower and more clearly earns a page. Revisit next ISCB slot.
• Characterize a protein–ligand contact footprint — noticed while writing the new interface recipe: the Interface Analysis skill explicitly covers protein–ligand contacts and antibody/antigen epitope residues out of the same NeighborSearch machinery, and the new recipe deliberately scopes to protein–protein only to keep one problem per page. The ligand case is a genuinely different reader — someone with a docked or co-crystal pose asking which residues to mutate to test the binding mode — and it sits downstream of dock-ligand-library-with-diffdock and prepare-protonation-states-for-docking, neither of which reports contacts. Decide amend-vs-add first: if it is only “run the same skill with a HETATM selection”, add a step to the new recipe rather than a page. Surfaced 2026-08-08.
• Pseudobulk single-cell DE end-to-end — natural composition of the QC + Scanpy pseudobulk + PyDESeq2 path; deferred because the Scanpy-MCP pseudobulk aggregation flow is mentioned but not yet documented as a focused end-to-end workflow.
• Process resting-state EEG for spectral features end-to-end — SHIPPED 2026-08-16 as extract-resting-state-eeg-spectral-features, with two departures from this banked note recorded under Recently surfaced: the component is the EEG Processing skill rather than NeuroKit2/MNE-EEG, and the deliverable is the periodic/aperiodic decomposition rather than the fixed-window band power this note assumed. Microstates and complexity measures were not covered and remain open — see the new deferred item below. EEGAgent (Yan et al., AAAI-26) is cited nowhere on the shipped page; it is an agent system, not a catalogued component, so it stayed out.
• Parameterize resting-state EEG in a time-resolved way (SPRiNT-style), or extract microstates / complexity measures — the two legs the new spectral recipe explicitly declined. The time-resolved leg matters when the aperiodic component itself drifts within a recording (task blocks, drowsiness onset), which a static per-recording fit averages away (Wilson et al., eLife 2022); it is named under that recipe’s Alternatives as a distinct future page. The microstate/complexity leg is the NeuroKit2 skill’s documented territory. Check before writing whether either is a genuinely different problem for a reader or just a parameter swap on the shipped page — the honest answer may be “amend, don’t add” for microstates and “add” for time-resolved.
• Fit a reinforcement-learning model to trial-by-trial choice data — the obvious next cognitive-modelling page now that fit-a-drift-diffusion-model-to-choice-rt-data has opened that leg: same reader, same trial table, different latent process (learning rate and inverse temperature rather than drift and boundary), and the Parameter Recovery Checker skill names RL models first among its use cases. Check the Awesome Cognitive and Neuroscience Skills collection for a catalogued RL-fitting skill before writing; if there is none, the recipe would be rung 1 over a pinned fitting library and the recovery skill becomes load-bearing rather than an alternative.
• Browse OpenNeuro for MRI/EEG/MEG datasets matching a task or modality — CANCELLED 2026-08-09, do not re-file as written. This was the top Neuroscience pick going into this run; on picking it up, the OpenNeuro MCP catalog page turns out to have carried verification: broken since 2026-07-20 — repo and hosted endpoint both 404 across four runs, no working install path — so the recipe’s only component does not exist. Nothing was written; the existing DANDI recipe’s optional OpenNeuro step was amended instead (see Flagged for review), and the gap is filed under Missing components. The problem is still real and still the section’s remaining discovery-layer hole — OpenNeuro is where the MRI/EEG/MEG datasets are, and DANDI does not cover them — so re-file this the moment a working OpenNeuro component is catalogued, not before.
• Score behavior from pose keypoints (supervised or unsupervised action segmentation) — the natural downstream sibling of the new track-animal-pose-in-behavioral-video recipe: keypoints → behavioral syllables or classified bouts (B-SOiD / keypoint-MoSeq / SimBA class). Deferred because no such component is in catalog/tools/ — filed as a Missing component. Do not write until one lands.
• Cortical thickness and subcortical volumes from structural MRI — DISCHARGED 2026-08-09 by measure-cortical-thickness-across-a-cohort (rung-2, Proposed, Fully open, HPC or cloud cluster). Do not re-file. All three banked worries were resolvable and became the page’s spine: wall-clock is stated as 6–12 h/subject with the tiering made explicitly N-dependent, version drift is quantified rather than hand-waved (Gronenschild 2012 / Haddad 2023, with the Chepkoech 2016 counterweight kept visible), and surface QC is the Euler number gate plus a quality covariate (Rosen 2018 / Bedford 2023). Two things to know if the page is revised: the recipe deliberately stops before multi-site harmonization — harmonization-tool is named under Alternatives so the ladder stays at rung 2, and a genuinely multi-site cohort page is a legitimate separate candidate — and the free-registration license.txt gate is the practical blocker readers hit first.
• Multi-site morphometry harmonization (ComBat-class) — surfaced 2026-08-09 while writing the cortical-thickness recipe, which explicitly scopes to a single-site cohort. harmonization-tool is catalogued and uncovered. The framing that earns a page rather than an amendment is the empirical-Bayes site-effect removal decision itself — which covariates to protect, why residualizing on site is not the same thing, and how to tell whether harmonization removed the effect you came to measure — not “add one more step to the thickness recipe”. Decide amend-vs-add before writing. Applies well beyond morphometry (the same skill covers connectivity and diffusion metrics), which argues for a page.
• Query the Allen Brain Atlas for cell-type / connectivity / expression context — clean rung-2 fit for the allenbrain-mcp (Alpha) — RMA queries, mouse connectivity experiments, brain-structure ontologies, 3D expression grid download. Deferred because the upstream MCP is Alpha with no LICENSE file; revisit after upstream clarifies licensing.
• Choose an integration method via scIB benchmarking — companion to the integration recipe; needs a documented scib-metrics driver in the catalog before composing.
• RNA velocity / latent-time analysis with scVelo — natural rung-2 follow-on to the new Census atlas and integration recipes; deferred because most published scVelo case studies require upstream velocyto / STARsolo / kallisto-bustools spliced-counts pipelines that aren’t yet in the catalog (a CLI-only step before Claude takes over). Revisit after either the scVelo SKILL.md adds a worked example or a velocity-aware aligner skill is catalogued.
• Drive pySCENIC end-to-end (motif filtering + AUCell) — natural rung-3 toolbelt extension of the new GRN-inference recipe; deferred because pySCENIC (the cisTarget + AUCell stack downstream of Arboreto) is not yet wrapped as a Claude skill or MCP server. Surfaced as a Missing component note.
• Recheck backlog: the 2026-07-04 cohort — recipe_recheck: yes was set for 2026-08-08 but only one recipe (dock-ligand-library-with-diffdock, chosen because it sits in the Chemistry focus area) was actually verified; the batch link-checking tooling was unavailable and per-page checking does not fit the run budget. Seven recipes remained at last_verified: 2026-07-04. 2026-08-15 update: six of the seven are cleared and bumped — screen-genome-for-resistance-and-virulence-genes, find-selective-cancer-dependencies-with-depmap, identify-bacterial-isolate-from-16s-sequence, organize-raw-dicom-to-bids-layout, build-phylogenetic-tree-from-sequences, integrate-single-cell-datasets. Every linked catalog/tools/ page resolves, every load-bearing component is verification: works, and every availability: Fully open tag is consistent with its components’ GA status. Scope caveat to carry forward: the external Sources URLs on those six were not re-fetched this run (no batch link-checker available, and per-page fetching does not fit the budget alongside the directed pass) — the component and availability legs are what got verified. One find worth keeping: openneuro.md has been verification: broken for four consecutive catalog runs, and organize-raw-dicom-to-bids-layout linked it bare in See also; that link now carries the broken status and points readers at the OpenNeuro website instead. sort-spikes-from-neuropixels-recording is the one still at 2026-07-04 — take it first next slot, and note its secondary AIND Data MCP link is verification: degraded (upstream moved from stdio to an HTTP endpoint), so check whether the Alternatives paragraph still describes a followable pairing.

Missing components

• PDSP Ki database component (a second curated-affinity source to cross-check GtoPdb) — the new find-a-selective-tool-compound-for-a-target recipe ranks candidates on GtoPdb’s expert-curated affinities, and its step 8 tells the reader to cross-check the top candidate before ordering anything. That instruction currently has no followable path for the sharpest available comparator. The gap is quantified rather than suspected: a systematic comparison of the NIMH PDSP Ki database against GtoPdb across histamine receptors H1R–H4R found limited ligand overlap, differing primary literature, pKi deviations between the two, and missing or incomplete functional characterization for 69 of 302 ligands — concluding that the two are complementary and that systematic integration is needed, since selective inclusion “may give the impression of missing data despite substantial experimental data” (Bindel & Seifert, Mol. Pharmacol. 2026). PDSP Ki DB is the canonical aggregated binding-affinity resource for aminergic GPCRs (the receptor class where tool-compound promiscuity does the most damage), and nothing wrapping it is Claude-installable (skill / MCP / plugin) today. A thin skill over its query interface would let the selectivity recipe compute a two-source agreement column instead of naming the comparison and walking away, and would unlock a cross-database affinity-reconciliation page of its own. Note the recipe’s step 8 currently routes the cross-check through the ChEMBL connector instead, which is a breadth substitute, not a curated-second-opinion one. Surfaced 2026-08-16.
• Kinase-activity inference layer (KSEA / PTM-SEA / INKA class) — the new find-differentially-regulated-phosphosites recipe stops at a site-level regulation table, because that is where the MaxQuant skill and MaxQuant’s own output stop. The question every user asks next — which kinase moved — needs a substrate-set enrichment layer over a kinase-substrate database. The method is benchmarkable and benchmarked: substrate-based inference reaches mean AUC 0.722 against a 184 kinase-condition gold standard, with Z-test and GSEA best and performance depending strongly on substrate count and evidence type (Hernandez-Armenta et al., Bioinformatics 2017); implementations exist as an R package and Shiny app (KSEAapp on CRAN / casecpb/KSEA), as Python (saezlab/kinact), and the tool landscape including PTM-SEA and INKA is reviewed in Piersma et al., Mass Spectrom. Rev. 2024. None is Claude-installable (skill / MCP / plugin) today, so the recipe names the route under Alternatives with its caveat but gives no followable path. Cataloguing one would unlock a pathway-level sibling recipe downstream of the site table — and note the R-package route means the same conda-not-pip friction as R-scape, so it is a catalog component rather than a recipe dependency. Surfaced 2026-08-15.
• A working OpenNeuro discovery component (replacement for the flagged OpenNeuro MCP) — the catalogued community wrapper has been verification: broken since 2026-07-20: repo QuentinCody/open-neuro-mcp-server and the hosted open-neuro-mcp-server.quentincody.workers.dev/sse endpoint both 404 across four consecutive runs. This blocked the top-priority Neuroscience recipe candidate outright this run (see the cancelled Deferred item) and forced an amendment to discover-nwb-recordings-on-dandi step 7. The gap is real and structural, not cosmetic: OpenNeuro is the primary public archive for MRI / MEG / EEG / iEEG / ECoG datasets, and the catalogued Neurosift Tools MCP covers DANDI (neurophysiology) only — so the cookbook currently has no followable path to discover a human neuroimaging dataset, the single most common first step in reusing open imaging data. OpenNeuro exposes a public unauthenticated GraphQL API (https://openneuro.org/crn/graphql), so any maintained wrapper — a fresh community MCP, an official one, or a thin Claude Skill over the GraphQL endpoint — would unblock it. Worth checking whether the upstream project simply moved rather than disappeared before declaring it gone. Surfaced 2026-08-09.
• SNP-based single-cell demultiplexer (Vireo / Souporcell / demuxlet / fastdemux class) — the new decide-if-a-cluster-is-a-doublet-artifact recipe rests on simulate-and-score detectors, which infer doublets rather than observe them. Where samples were pooled from several donors, genotype-based demultiplexing identifies cross-donor doublets directly and is strictly stronger evidence: leading tools reach 80–85% accuracy with Vireo best, degrading as the doublet percentage rises (Fu et al., Brief Bioinform 2025); a newer DLDA-based method matches that accuracy while cutting runtime and peak memory by orders of magnitude and extends naturally to higher-order multiplets (Ranjbaran, Luca & Pique-Regi, fastdemux 2026). None is Claude-installable (skill/MCP/plugin) today, so the recipe names the route under Alternatives without a followable path. Cataloguing one would both upgrade the doublet recipe’s step 3 and unlock a pooled-design demultiplexing recipe of its own — pooling is now standard practice in large single-cell studies, so this is a broadly-used gap, not a niche one. Surfaced 2026-08-08.
• Behavior-classification layer downstream of pose keypoints (B-SOiD / keypoint-MoSeq / SimBA class), and Lightning Pose as a higher-accuracy pose alternative — the new track-animal-pose-in-behavioral-video recipe stops at keypoints and kinematics, because that is where the catalogued DeepLabCut skill stops. The question every user asks next — “how much time did the animal spend grooming / rearing / in a given behavioral syllable” — needs an action-segmentation component (B-SOiD, keypoint-MoSeq, SimBA); none is Claude-installable (skill/MCP/plugin) today. Separately, Lightning Pose beat DeepLabCut on a 24-landmark infant task with near-identical training sets, 99.75% vs 97.80% over 3,200+ unseen frames (Kaur et al., EMBC 2024) — the recipe names it under Alternatives but cannot recommend a followable path. Cataloguing either would unlock a behavior-scoring sibling recipe and an accuracy-tier choice in the pose recipe. Surfaced 2026-08-02.
• Local druggability scorer (DoGSiteScorer-class) and a cryptic-pocket predictor (PocketMiner-class) — the new detect-and-rank-druggable-pockets recipe runs fpocket and P2Rank entirely locally, but the only druggability model the Binding Site Detection skill reaches is DoGSiteScorer’s SVM on the ProteinsPlus web server — a data-residency blocker for an unpublished structure, with no catalogued local substitute (fpocket’s own drugscore is coarser). Separately, the recipe’s cryptic-pocket path is an MD-ensemble mdpocket run costing GPU-hours; a Claude-installable single-structure cryptic-site predictor (PocketMiner, benchmarked against P2Rank on the 1,107-pair CryptoBench set — Škrhák et al. 2024) would give a cheap first pass. Neither is Claude-installable (skill/MCP/plugin) today. Would unlock an offline-druggability sibling and a no-MD cryptic-pocket step. Surfaced 2026-08-01.
• SHIPPED as a recipe dependency, not a catalog component. User request #74 (@goodb, 2026-07-27) → shipped localize-implant-tip-in-brain-atlas-subregion (rung-3, Proposed, Fully open, Laptop). The two load-bearing libraries are not catalogable — neither ships as a Claude Skill / MCP / plugin / connector, and the catalog curator declined them on 2026-07-27 (checked PyPI, both repos, and the NeuroClaw / SciAgent / K-Dense / bioSkills collections). Do not re-file them for the catalog. Both are pip dependencies in the recipe’s ## Dependencies block. Verified this run (fetched 2026-07-27 from PyPI): brainglobe-atlasapi 2.3.1, BSD-3-Clause, import brainglobe_atlasapi; ships allen_mouse_25um with annotation (voxel→structure-id), lookup_df (id→acronym/name) and structure-ancestors (hierarchy roll-up) — the subregion readout; source JOSS 2020. DeepSlice 1.2.8, import DeepSlice, QuickNII-compatible JSON output, first-run model-weight download; source Carey et al., Nat Commun 2023. License correction: the banked note claimed DeepSlice was MIT — PyPI 1.2.8 declares GPL-3.0-only (copyleft). The recipe stays Fully open (still OSS, no gate) but flags the copyleft redistribution obligation; if a future run needs to redistribute a derived pipeline, re-check the license. Catalogued neighbours still don’t substitute: allenbrain-mcp is read-only RMA/ontology/image queries (no coordinate→subregion lookup, no plane model, Alpha, no LICENSE); the SimpleITK skill is general 3D registration with no Allen atlas and no cutting-plane model — the confidently-wrong off-axis path @goodb warned against. scikit-image does the per-section fiber-tip detection. Surfaced 2026-07-27, shipped 2026-07-27.
• Peptide-set coverage optimizer (PopCover-2.0 / EpiToolKit-class) — the new scan-protein-for-cd4-t-cell-epitopes and existing scan-protein-for-cd8-t-cell-epitopes recipes each produce a ranked per-allele epitope table, but selecting a minimal peptide set with broad HLA + pathogen-strain coverage across the combined CD4/CD8 shortlist (the step that turns predicted binders into a deployable vaccine cocktail, experimentally T-cell-confirmed in Nilsson et al., Front. Immunol. 2021) needs a set-cover optimizer like PopCover-2.0. Not Claude-installable (skill/MCP/plugin) today. Would unlock a “design a coverage-optimized multi-epitope peptide set” capstone recipe downstream of both epitope scans. Surfaced 2026-07-25.
• Constrained-retrosynthesis agent (LARC-style Agent-as-a-Judge) — the new plan-a-synthetic-route-for-a-target-molecule recipe uses the CovaSyn MCP covaplatform retrosynthesis tool for unconstrained route search, then post-filters for simple constraints. Search-time enforcement of hard constraints (reagent-class avoidance, reaction whitelist, step-count ceiling) via an in-loop constraint judge — the design that lifts LARC to 72.9% on 48 constrained tasks — needs a Claude-installable agentic-retrosynthesis framework. None is a skill/MCP/plugin today. Would unlock a constrained-route sibling recipe. Surfaced 2026-07-18.
• Trained KG drug-repurposing model wrapper (TxGNN / COMIC / CellAwareGNN) — the new find-repurposing-candidates-in-a-knowledge-graph recipe uses the PrimeKG skill, which exposes only neighbor lookups (search_nodes/get_neighbors/get_disease_context), so the recipe is an interpretable graph-traversal heuristic, not a learned link predictor. A Claude-installable wrapper around a trained GNN over PrimeKG — TxGNN (the canonical model), or the newer COMIC (recovered 21/30 recent FDA repurposing pairs, +9.55% over SOTA) / CellAwareGNN (AUPRC 0.826) — would unlock a calibrated ranked-prediction sibling recipe with far higher recall than the traversal heuristic. None is Claude-installable (skill/MCP/plugin) today. Surfaced 2026-07-12.
• Colocalization / Mendelian-randomization driver (coloc / TwoSampleMR / SuSiE-coloc) — the new triage-gwas-hits-to-drug-targets recipe uses GWAS-MCP for the lookup-and-annotate layer (variant consequence, eQTL-implicated gene, GWAS Catalog context, drug-target support), but stops short of causal inference. Establishing that a GWAS signal and an eQTL share a causal variant (colocalization) and the direction of effect (MR) — the steps that lift Lessard et al.’s approved-target enrichment from 1.75 to 2.58 (BMC Genomics 2024) — needs summary-statistic-level statistical tools (coloc, TwoSampleMR). Neither is Claude-installable (skill/MCP/plugin) today. Would unlock a causal-directional sibling recipe downstream of the GWAS triage. Surfaced 2026-07-12.
• Local pH-aware protonation-state enumerator (Dimorphite-DL / OpenBabel -p / ChemAxon cxcalc) — the new prepare-protonation-states-for-docking recipe uses the Rowan skill submit_macropka_workflow, which runs cloud-side (SMILES leave the machine — a data-residency blocker for confidential structures). A Claude-installable local pH-aware protonator (Dimorphite-DL, OpenBabel -p) would enable an offline sibling recipe and remove the cloud caveat. Datamol/RDKit standardize charges/tautomers but are pH-blind, so no catalogued local component covers this today. Surfaced 2026-07-11.
• ANI / MLST caller (fastANI / pyani / mlst) — the new identify-bacterial-isolate-from-16s-sequence recipe gives a fast 16S-based genus/species call, but 16S alone frequently ties multiple close species. Definitive species delineation needs whole-genome average nucleotide identity (fastANI, pyani) or multi-locus sequence typing (tseemann/mlst). None is Claude-installable (skill/MCP/plugin) today. Would unlock a confirmatory species-delineation recipe downstream of the 16S ID. Surfaced 2026-07-04.
• Dedicated resistome caller (RGI / AMRFinderPlus) — the new screen-genome-for-resistance-and-virulence-genes recipe uses the catalogued BLAST MCP for straight homology detection against CARD/VFDB, but RGI (arpcard/rgi) and NCBI AMRFinderPlus add SNP/protein-variant models that catch point-mutation resistance (e.g. gyrA fluoroquinolone mutations) a homology search misses. Neither is Claude-installable (skill/MCP/plugin) today. Would upgrade the AMR recipe from homology-only to variant-aware. Surfaced 2026-07-04.
• Biobank-scale mixed-model GWAS wrapper (regenie / SAIGE / BOLT-LMM) — the catalogued PLINK2 skill covers array/small-cohort fixed-effect GWAS (now the case-control GWAS recipe), but cohorts of 100k+ samples with relatedness/structure need a linear mixed model to control inflation. No regenie/SAIGE/BOLT-LMM component is Claude-installable today. Would unlock a biobank-scale sibling recipe. Surfaced 2026-06-14.
• pySCENIC wrapper (cisTarget + AUCell) — no Claude-installable component for the steps downstream of GRNBoost2 (motif filtering against cisTarget databases, per-cell regulon AUCell scoring). Would unlock a full SCENIC-pipeline recipe on top of the new GRN-inference recipe. Surfaced 2026-05-28.
• ADMET-AI / AdmetLab 3.0 / Deep-PK wrapper — RESOLVED 2026-07-26: the ADMETlab MCP server (ToxMCP/admetlab-mcp, fronting ADMETlab 3.0; Fu et al., NAR 2024) is now catalogued and shipped as the new predict-admet-properties-for-a-compound-series recipe — the ML-prediction layer (CYP isoform, hERG, clearance, solubility with uncertainty) that pure descriptors miss. Note remaining: no local / offline ML-ADMET predictor is catalogued (ADMETlab runs cloud-side; a data-residency blocker for confidential structures); a Deep-PK-class local wrapper would unlock an offline sibling recipe. Originally surfaced 2026-05-27.
• MARRVEL-MCP (bioRxiv 2025-11-26, hyunhwan-bcm/MARRVEL_MCP) — rare-disease variant-interpretation MCP server with 39 tools and a published 95%-accuracy benchmark on 45 expert-curated tasks. Would strengthen the Interpret a clinical variant recipe from Proposed to Reported, and give the new diagnose-rare-disease-from-patient-phenotypes recipe a directly-benchmarked, dedicated substitute for the ToolUniverse rare-disease skill (which uses MARRVEL tools under the hood but is not itself benchmarked). Not catalogued as a standalone Claude-installable component today. Surfaced 2026-05-23; re-noted 2026-07-19.

User requests (open)

None.

User requests (blocked)

None.

User requests (closed this run)

None.

Composition reports

Demand signal from the Composer plugin (/composer:compose). The responder routes report=composition entries into ## User requests (open) like any other request; process them there each run (promote a success, write a recipe for a gap), then keep a rolling tally here so the directed pass can prioritize the problem classes and subject areas scientists actually compose against. Keep the last ~15 lines.

Format: - YYYY-MM-DD outcome=<worked|gap|failed> problem_class=<…> → <what shipped / note>

• 2026-06-20 outcome=worked problem_class=Knowledge synthesis → shipped new recipe map-disease-to-genes-and-pathways (Open Targets → gget/Enrichr chain); cascaded Open Targets MCP handshake breakage into prioritize-targets recipe (#43).
• 2026-07-16 outcome=worked problem_class=Hypothesis generation → prioritize-targets reused for AD-prevention (SORL1 lead, prevention-tuned pillar re-weighting); added field report + reinforced direct-GraphQL fallback (#55).