Recipes updates

Reverse-chronological log of changes to the recipes cookbook. Newest at the top.

Older entries live in RECIPES_CHANGELOG_ARCHIVE.md.

2026-08-16

Added

  • Find a selective tool compound for a target (Problem class: Knowledge synthesis; Evidence: Proposed) — rung-2 recipe over the GtoPdb Claude Skill: resolve a target by UniProt accession, pull its curated interactions, and rank candidate ligands by affinity and a measured selectivity window computed from each ligand’s own off-target profile. Three gates guard the arithmetic — species is declared before any affinity is read and never pooled, affinity can be an empty string and is routed to no_affinity.csv rather than coerced to zero, and the window is a log subtraction (Δ 1.5 ≈ 30-fold) not a ratio. The verdict defaults to insufficiently_profiled, because absence of off-target data is not selectivity (Bindel & Seifert 2026 report pKi deviations up to 4.2 log units between curated sources and 69 of 302 aminergic ligands with incomplete functional characterisation; Harding et al. 2026 give current coverage at 3,103 protein targets and 13,260 ligands).

Updated

  • Profile a compound’s polypharmacology from ChEMBL bioactivity data — cross-linked in See also to the new target-first sibling, with a one-line note on which recipe to enter from which starting point.

Flagged

  • None.

2026-08-16

Added

  • Design a Bayesian trial that borrows external control data (Problem class: Experimental design; Evidence: Proposed) — discharges the second-ranked Translational Medicine deferred item and opens the Bayesian leg of the clinical-trials chain, which until now had a frequentist group-sequential page and nothing for the borrowing decision. Rung 2 on the Bayesian Trials skill. Built around three gates: the no-borrowing comparator is designed and committed first, the borrowed information is declared as a prior effective sample size in patients against the skill’s 20–80%-of-control-arm band, and operating characteristics are simulated over a two-way true effect × prior-data conflict grid rather than only under exchangeability (Schmidli et al. 2014, Hupf et al. 2021).

Updated

  • Design a group-sequential trial with interim looks — added the new Bayesian recipe to See also as the other-framework sibling; the page had named the skill under Alternatives without a followable path.

Verified (no changes)

  • recipe_recheck: no this run — no aging-recipe verification pass performed.

2026-08-16

Neuroscience directed pass. Two recipes, both rung 2, both Proposed, both Fully open / Laptop — one discharging the section’s banked resting-state EEG item, one opening a leg the section did not have.

Added

  • Extract spectral features from resting-state EEG (Problem class: Data analysis; Evidence: Proposed) — pairs the EEG Processing skill with a pinned fooof==1.1.1 so band power and the aperiodic exponent come out separated and separately labelled, which is the whole point: an EEG spectrum is 1/f-like broadband activity plus narrowband peaks, so a fixed-window “alpha power” difference moves whenever the slope moves. That confound is documented in both directions — aperiodic parameters vary systematically across sleep stage, age and region in 251 subjects (Schneider et al. 2022) and are modulated by dopaminergic medication in Parkinson’s disease (Wang et al. 2022). Two gates carry the page: the fit range and max_n_peaks are recorded literals fixed before the group labels are seen, because odd–even reliability of the intercept and slope falls as more peaks are allowed and more outliers are generated (Kałamała et al. 2026); and a pre-registered r_squared floor writes excluded channel IDs out, since a bad 1/f fit yields a confidently wrong exponent that leaves no trace in the feature table. FOOOF-vs-IRASA has a ground-truth head-to-head (Gerster et al. 2022). Pin note: the successor package specparam is at 2.0.0rc7, still a release candidate, hence the stable predecessor.
  • Fit a drift-diffusion model to choice and reaction-time data (Problem class: Data analysis; Evidence: Proposed) — the Drift-Diffusion Model skill plus a pinned pyddm==0.9.0, decomposing two-choice accuracy and RT into drift rate, boundary, bias and non-decision time. This opens the cognitive-modelling leg of the Neuroscience section, which was until now entirely signals and images with nothing for behavioural data. The rung-2 argument is not that 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, so the page front-loads a committed ddm_spec.md declaring the hypothesis in parameter terms and the free-parameter set before fitting, 10 random starting points with the likelihood spread recorded, posterior-predictive RT quantile overlays rather than means, and the recovery check with its r > 0.5 cross-parameter tradeoff warning. Method leg is current: PyDDM (Shinn, Lam & Murray 2020), used in 2026 to attribute macaque transitive-inference learning to drift rate and the eye-vs-reach RT gap to non-decision time (Munoz et al. 2026). HDDM is documented as an in-recipe substitution for low per-subject trial counts (Wiecki et al. 2013), not a second page.

Updated

  • Extract event-related potentials from EEG epochs — its Alternatives section had promised a resting-state sibling as “a natural sibling recipe, not this one”; that sibling now exists and is linked.

Flagged

  • None.

Verified (no changes)

  • recipe_recheck: no this run — no aging-recipe verification pass. Both new pages’ components were read from the catalog and both dependency pins were fetched from PyPI today.

2026-08-15

Added

  • Find differentially regulated phosphosites, not just changed proteins (Problem class: Data analysis; Evidence: Proposed) — the cookbook could call a protein changed but had no path to the residue-level question, which carries machinery the protein-level page does not: a localization gate, site-level rather than protein-level rollup, and normalization to protein abundance. Rung 2 on the MaxQuant skill — and the rung survives a correction made this run: the skill’s own workflow stops at proteinGroups.txt and does not document phospho site tables, so it earns its place by driving step 1’s paired unenriched proteome arm, with the site layer scripted on top against column names verified against the MaxQuant output-tables reference. Three gates: declare the unenriched arm before enriching (1,150 of 5,385 proteins changed significantly alongside 4,218 regulated sites in the same samples — Esnault et al. 2018); record the localization-probability cutoff next to the fragmentation method, since a 1% false localization rate needs a different threshold per method (Savitski et al. 2011); and expand the ___1/___2/___3 multiplicity columns to rows rather than summing them.

Updated

  • Find differentially abundant proteins in a label-free proteomics experiment — cross-linked the new site-level sibling in See also; this recipe is the phospho recipe’s unenriched arm.

Flagged

  • None.

Verified (no changes)

  • Recheck slot off this run (recipe_recheck: no); no aging-recipe verification performed.

2026-08-15

Added

  • Look up the curated composition and stoichiometry of a protein complex (Problem class: Knowledge synthesis; Evidence: Proposed) — the cookbook could infer PPI edges and characterize an interface it was handed, but had no followable path to the question both of those depend on: is this a curated complex, and at what stoichiometry? Rung 1 — Complex Portal and IntAct are chat-only Claude Science connectors with no addressable MCP URL, so the artifact is Claude Code scripting the CC0 ComplexTAB release files (release dated 2026-01-14, 19-column schema and the worked CPX-4111 row verified this run). Four gates: refuse 9606_predicted.tsv by default (it is the same 5.0 MB size as the curated human file, so a *.tsv glob silently doubles your hits); treat the Evidence Code ECO term as the headline field, not a footnote; fail on an unrecognised participant prefix rather than dropping a ChEBI cofactor; emit an explicit “no curated complex found” row, because curation is deliberately incomplete. Why not a PPI network has a number — only ~40% of expanded co-complex pairs also have genetic interactions (Meldal et al. 2021).

Updated

  • Characterize a protein–protein interface — step 6’s orthogonal-evidence line cited Complex Portal and IntAct as links only; now cross-links the new recipe as the followable path.

Flagged

  • None.

Verified (no changes)

  • Recheck slot off this run (recipe_recheck: no); no aging-recipe verification performed.

2026-08-15

Added

  • Compare bulk TCR/BCR repertoires across samples from raw immunosequencing reads (Problem class: Data analysis; Evidence: Proposed) — the cookbook’s first bulk repertoire recipe; the three existing pages are all single-cell, while the amplicon path most vaccine and immuno-oncology labs run had nothing. Rung 3 (MiXCR → VDJtools → Repertoire Visualization), with depth normalization written as a hard-abort gate rather than a step and the chemistry preset required in config.yaml before any read is aligned. Tagged Institutional access: MiXCR is free only for academic users with no commercial funding (milaboratory/mixcr, verified this run). The two gates are the evidenced part — multiplex-PCR frequencies only 42–62% accurate and uncorrected error inflating diversity up to 5000-fold (Khan et al. 2016); the canonical age-vs-diversity result required precisely normalized per-10⁶-cell profiling (Britanova et al. 2014); classical richness estimators frequently underestimate true TCR diversity (Laydon et al. 2015).

Updated

  • None.

Flagged

  • None.

Verified (no changes)

  • recipe_recheck: no this run — no aging-recipe recheck performed.

2026-08-15

Added

  • Get balanced reactions and cross-references for a metabolite or EC number (Problem class: Knowledge synthesis; Evidence: Proposed) — discharges the top Chemistry deferred item. Rung 1, not 2: Rhea and ChEBI are catalogued as Claude Science connectors with no addressable MCP URL, so they cannot leave a re-runnable artifact behind; the recipe has Claude Code script the CC BY 4.0 release files instead and names the connector as the interactive alternative. Four gates, each traceable to a documented Rhea convention verified this run: map the ChEBI ID through chebi_pH7_3_mapping.tsv before searching (participants are the major microspecies at pH 7.3, so a query on the neutral form returns nothing — indistinguishable from “no known reactions”); never emit a master ID as a modelling result (the quartet is UN/LR/RL/BI with four identifiers, the master’s direction undefined); label columns side_left/side_right, never substrates/products, because Rhea attaches no semantic meaning to the two sides; and join cross-references on MASTER_ID, since an external database may have mapped to any quartet member (Morgat et al. 2020). Release 141 (2026-06-10, 18,558 reactions) and all ten TSV files confirmed against the FTP release directory and its README.txt this run. Stdlib-only, so no ## Dependencies block; Fully open, Laptop, fully local after the download (Bansal et al. 2022).

Updated

  • Organize a raw DICOM dataset into a BIDS layout — the bare See also link to the OpenNeuro MCP now carries that component’s verification: broken status (no working install path for four consecutive catalog runs) and points readers at the OpenNeuro website for submission instead.

Verified (no changes)

  • 6 recipes cleared from the 2026-07-04 recheck backlog and bumped to today: 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 page resolves, every load-bearing component is verification: works, and every Fully open tag is consistent with its components. Scope note: the component and availability legs were verified; external Sources URLs were not re-fetched. sort-spikes-from-neuropixels-recording remains at 2026-07-04 and is first in line next recheck slot.

2026-08-09

Added

  • Trace an adverse outcome pathway from a molecular initiating event to an organ finding (Problem class: Knowledge synthesis; Evidence: Proposed) — rung-2 recipe on the Adverse Outcome Pathway skill, discharging the top Drug Repurposing and Discovery deferred item. Finding-keyed rather than compound-keyed: you arrive with an organ-level tox result and leave with candidate MIE-to-endpoint pathways, weight of evidence recorded per key-event relationship, and the one assay that discriminates between the survivors. Four gates carry the page — stressor absence made an explicit value rather than a blank (871 prototypical stressors across 595 AOPs, AOP-Wiki metrics, fetched this run), one row per key-event relationship with the weakest link reported rather than an average (Holmer et al. 2024), OECD status recorded before the pathway content is read (Tanabe et al. 2025), and CTD direct-curated split from inferred associations (King et al. 2012).

Updated

  • None.

Flagged

  • None.

Verified (no changes)

  • 0 recipes spot-checked — recipe_recheck: no for this slot; the 2026-07-04 recheck backlog carries forward.

2026-08-09

Added

  • Design a group-sequential trial with interim looks (Problem class: Experimental design; Evidence: Proposed) — rung-2 Adaptive Designs skill recipe opening the design-stage leg of the clinical-trials chain, which previously jumped straight from protocol drafting to endpoint analysis. Built around three gates rather than the boundary arithmetic: a committed fixed-sample comparator before any interim look is added; operating characteristics simulated across an effect grid at a recorded seed, never at the design point; and sample-size re-estimation defaulting to no unless the simulation earns it — a standing presumption for time-to-event endpoints, where SSR lets interim curve-shape information enlarge the observed effect and raises the chance of recommending an ineffective therapy (Freidlin & Korn 2017). Also encodes that the promising-zone rule triggers where the marginal power per patient is not greatest (Jennison & Turnbull 2015), that optimising conditional versus unconditional power yields different rules (Mehta et al. 2022), and the protocol-must-specify checklist from Shih, Li & Wang 2016. rpact 4.4.0 (2026-03-04, LGPL-3) confirmed by fetch this run.

Updated

  • Handle missing endpoint data in a confirmatory trial — Alternatives now points at the new design-stage recipe instead of naming the Adaptive Designs skill without a followable path.

Flagged

  • None.

Verified (no changes)

  • recipe_recheck: no this run — no aging-recipe verification pass. Seven recipes remain at last_verified: 2026-07-04; see the recheck backlog in curator state.

2026-08-09

Neuroscience focus day. The top-priority candidate turned out to be unbuildable — its only component is a flagged, 404-ing MCP server — so the run shipped the next pick instead and spent the remainder documenting the dead end properly.

Added

  • Measure cortical thickness and subcortical volumes across a cohort (Problem class: Data analysis; Evidence: Proposed) — rung-2 FreeSurfer skill recipe opening the structural-MRI morphometry leg, which the Neuroscience section lacked entirely despite covering DTI, resting-state fMRI, EEG/ERP, calcium imaging, spike sorting and BIDS conversion. The recon-all invocation is the settled part; the page is built around the two failure modes that leave no trace in the output table. One pinned FreeSurfer version for the whole cohort, because 6–12 h/subject means cohorts get processed in waves across months, and regional estimates are not comparable across major versions — 8.8 ± 6.6% on volumes and 2.8 ± 1.3% on thickness between v5.0.0 and earlier releases, with workstation type and even the macOS point release worth about half as much again (Gronenschild et al., PLoS ONE 2012); cingulate thickness at ICC 0.37–0.61 across v5.3/6.0/7.1, and version differences in downstream statistical results (Haddad et al., Hum Brain Mapp 2023). The counterweight is kept visible rather than suppressed — absolute shifts of 1.6–5.8% barely moved AD/MCI classification accuracy (Chepkoech et al., Hum Brain Mapp 2016) — and reconciled: within-version stability is good, which is exactly why one version per cohort suffices and mixing versions across groups does not. Second, the Euler number is gated on before any group comparison (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., NeuroImage 2018) with the threshold recorded as a literal chosen before group labels are seen — 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., Imaging Neuroscience 2023). Thickness and surface area stay separate analyses; eTIV covaries volumes only. Tagged HPC or cloud cluster with the N-dependence spelled out (ten subjects fit a workstation overnight; sixty is an array job — and that duration is what creates the version-drift risk in the first place), and Fully open with the free-registration license.txt gate named up front. FreeSurfer 8.2.0 (March 2026) confirmed by fetch this run.

Updated

  • Discover NWB recordings on DANDI and prepare them for sorting — optional step 7 recommended installing the OpenNeuro MCP, whose catalog page has been verification: broken since 2026-07-20 (repo and hosted endpoint both 404 across four consecutive runs, no fixable path). Step 7 now states the breakage and points readers at manual openneuro.org search; 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 core Neurosift + spike-sorting path is unaffected and last_verified was deliberately not bumped — this run was recipe_recheck: no and the rest of the page has not been re-verified.

Flagged

  • Browse OpenNeuro for MRI/EEG/MEG datasets (planned recipe, not written) — cancelled rather than written. It was the top Neuroscience pick going into the run and its only component does not exist. Filed as a Missing component: OpenNeuro is the primary public archive for MRI / MEG / EEG / iEEG / ECoG data and the catalogued Neurosift MCP reaches DANDI only, so the cookbook currently has no followable path to discover a human neuroimaging dataset. OpenNeuro’s public GraphQL API is unauthenticated, so any maintained wrapper would unblock it.

Verified (no changes)

  • No aging-recipe recheck this run (recipe_recheck: no). The 2026-07-04 backlog of seven recipes remains queued for the next recheck slot.

2026-08-08

Directed pass on Molecular and Cellular Biology. Two recipes, both rung 2, both discharging long-standing deferred items — one opening the cookbook’s empty proteomics leg, one resolving a banked amend-vs-add question in favour of a new page.

Added

  • Find differentially abundant proteins in a label-free proteomics experiment (Problem class: Data analysis; Evidence: Proposed) — rung-2 MaxQuant skill recipe opening the proteomics leg, which was entirely empty next to six transcriptomics recipes. Built around two gates rather than the search: the Potential contaminant / Reverse / Only identified by site filter written as mandatory and non-configurable, and a differential test run twice — with the declared imputation and on complete cases only — emitting an imputation_sensitive flag per protein. That second gate exists because benchmarks find “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., J. Proteome Res. 2023), with accuracy driven by the MNAR rate you cannot measure (Jin et al., Sci. Rep. 2021). The quantification layer is the settled part (MaxLFQ, Cox et al. 2014). Tagged Fully open but states plainly that MaxQuant is free-of-charge, not open-source (MPI terms at download, redistribution restricted), and that there is no macOS build and no Linux GUI. PRIDE is step 0 for a reanalysis, not a second rung.
  • Decide whether a suspicious cluster is a doublet artifact (Problem class: Data analysis; Evidence: Proposed) — rung-2 Doublet Detection skill recipe for the reader who already has a clustering and one intermediate population, as distinct from the QC recipe whose deliverable is a filtered object. Expected rate from the lane cell count; scoring per sample on raw counts before integration, with the script refusing a merged object; two detectors kept as separate columns; calls rolled up to a per-cluster summary; then two checks the scores do not use — library-size doubling and additivity of the two lineage programs — feeding a three-way doublet_artifact / real_population / undetermined verdict that defaults to undetermined, because homotypic doublets are largely undetectable and a clean negative is not proof. Grounded on nine methods over 16 experimentally-annotated and 112 synthetic datasets (Xi & Li, Cell Systems 2021); the page surfaces rather than hides the tension that the benchmark’s accuracy winner (DoubletFinder) is not the skill’s stated default (scDblFinder).

Flagged

  • None.

Verified (no changes)

  • recipe_recheck: no for this slot — no aging-recipe verification run. The 2026-07-04 cohort of seven recipes remains queued in recipes/curator-state.md for the next recheck slot.

2026-08-08

Added

  • Characterize a protein–protein interface from a structure you already have (Problem class: Data analysis; Evidence: Proposed) — rung-2 Interface Analysis skill recipe, opening the analyse-an-existing-complex leg. The cookbook could predict a complex (co-folding, antibody–antigen) but had nothing to say about a structure you already have. Design forces the two decisions people skip: biological assembly, not asymmetric unit (recorded assembly ID, hard stop if it cannot be established) and cutoff declared before the run (4.5 Å heavy-atom primary, 8 Å Cα emitted as a second column so definition-sensitivity is visible). Contacts and interface residues kept as two tables rather than one, because merging duplicates residues per partner contact and inflates every downstream count; SASA required to use an identical probe radius and atom selection across complex and both isolated chains. The load-bearing step is a three-way tiered biological-vs-packing verdict defaulting to ambiguous, with orthogonal evidence (Complex Portal/IntAct, solution data) recorded as a separate line. Grounded in the classification literature: ROC 0.923 and 91.4%/91.7% test accuracy, outperforming DiMoVo/Pita/PISA/EPPIC (Luo et al., Proteins 2014); up to 94.8% on 172 packing-vs-functional complexes (Block et al. 2006); ~88.8% with a discriminator explicitly uncorrelated with contact area (Tsuchiya et al. 2008) — the reason BSA alone is not the verdict; 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. Fully open, Laptop, fully local except the PDBePISA cross-check, which is flagged as the one step that discloses a confidential structure.
  • Test whether a proposed RNA structure is actually conserved (Problem class: Data analysis; Evidence: Proposed) — rung-2 Covariation Analysis skill recipe (R-scape), the validation counterpart to the existing thermodynamic ViennaRNA recipe: a folder returns a structure for any sequence including a shuffled one, so the cookbook could produce RNA structures and never test one. Design: alignment built independently of the structure as the load-bearing gate, because covariation-guided alignment construction followed by a covariation test on the same alignment is a documented artifact; power analysis run with the test, never after it, with a rejected verdict unavailable below ~10% mean power; three-way verdict (supported / rejected / undetermined) with the mean power written into the verdict line so the conclusion cannot be quoted without its caveat; proposed-pair and alternative-pair covariation kept in separate files so “structure rejected” and “different structure supported” stay distinguishable; helix-level aggregated E-values reported alongside per-pair ones (R-scape ≥ 2.0.0.p required), since per-pair-only reporting discards real structures; --cacofold explicitly barred from being reported as validation of its own consensus. Grounded in three papers on exactly this test: no significant support for the proposed HOTAIR, SRA and 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). Fully open, Laptop, fully local — usable on unpublished sequences.

Verified (no changes)

  • recipe_recheck: no this run — no aging-recipe verification pass. Every catalog page linked by the two new recipes was read this run and confirmed present (Interface Analysis, Covariation Analysis, PDB, PDBe, Complex Portal, IntAct, PyMOL, ChimeraX, ViennaRNA, RNA Structure Probing, ncRNA Search, Rfam).

2026-08-08

Added

  • Test which immune populations differ between groups in a cytometry cohort (Problem class: Data analysis; Evidence: Proposed) — rung-3 chain of three bioSkills cytometry skills, opening a leg where all five catalogued cytometry skills were uncovered and the only existing recipe was deliberately I/O-only. The recipe’s job is closing three failures that all produce publishable-looking tables: pseudoreplication (closed mechanically by emitting a committed per-sample-per-cluster counts.csv and testing that, since methods ignoring biological-replicate variation “can discover hundreds of differentially expressed genes in the absence of biological differences”), compositional artifact (every DA hit re-checked and labelled compositional_confirmed/compositional_suspect), and conflating abundance with state (two tables, never merged). Testing engine grounded in diffcyt.
  • Find which taxa differ between microbiome groups (Problem class: Data analysis; Evidence: Proposed) — rung-2 Differential Abundance skill recipe completing the microbiome leg: the existing diversity recipe answered “do the communities differ” and nothing answered “which taxon”. Runs ≥2 compositionally-aware methods on one filtered table and tiers the output high_confidence (intersection) vs exploratory, which is the explicit recommendation of the benchmark showing 14 methods across 38 datasets “identified drastically different numbers and sets of significant ASVs”, with hit counts tracking sample size and sequencing depth rather than biology. Declares the prevalence filter as a recorded parameter, refuses DESeq2/edgeR on compositional data, and records the taxonomy reference release.

Updated

  • None.

Flagged

  • None.

Verified (no changes)

  • 0 recipes re-verified — recipe_recheck: no for this slot, so the run went to discovery. The 2026-07-04 recheck backlog (seven recipes) remains queued in Deferred.

2026-08-08

Chemistry-focused pass. Two new recipes, both opening legs of the cookbook that had no coverage at all.

Added

Changed

Verified

  • Dock a ligand library into a target structure with DiffDock — all six linked catalog pages, both linked autonomous-science systems, and all three linked recipes resolve; the DiffDock catalog page is still GA / verification: works and not flagged; the ACL Anthology and arXiv citations resolve. last_verified bumped from 2026-07-04.

The remaining seven recipes in the 2026-07-04 cohort were not rechecked this run — the batch link-checking tooling was unavailable, and per-page checking does not fit the run budget. The next recheck slot should start from that cohort: 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, sort-spikes-from-neuropixels-recording.

2026-08-02

Added

  • Fit a dose-response curve and report a defensible IC50 (Problem class: Data analysis; Evidence: Proposed) — closes the cookbook’s most conspicuous assay-analysis hole: the SAR, QSAR, synergy and ADMET recipes all consume potency numbers and nothing produced one. Rung-2 Dose-Response Analysis skill with replicates kept unpooled, a rejected_curves.csv that emits no IC50 at all for biphasic or plateau-less data, and truncated-window AUC alongside IC50 because that is the metric that survives cross-platform comparison (Chen et al., Bioinformatics 2026; Johnstone et al. 2016; Zhang et al. 2013)
  • Source purchasable compounds for a hit list (Problem class: Knowledge synthesis; Evidence: Proposed) — the step between a ranked hit list and a plate, previously skipped: hits were filtered, docked and ADMET-profiled, then assumed to exist. Rung-2 Chemical Sourcing skill emitting one row per listing with normalized price per mg, in-stock and make-on-demand never collapsed, and four exception files (SMILES mismatch, no source, missing purity, price outliers) read before the main table. Purity is a gate because only 76% of gradeable Tox21 library samples exceeded 90% purity and 11% degraded in four months (Richard et al. 2025); the in-stock/make-on-demand split is a chemical-space decision because >97% of make-on-demand scaffolds are absent from in-stock collections (Irwin et al. 2020). Flags the disclosure caveat as the real access bar — query SMILES leave the machine to commercial vendors, and no offline substitute can exist

Updated

  • Score a drug-combination screen for synergy — cross-linked to the new dose-response recipe, which is where the single-agent potency values it requires now come from
  • Filter virtual screening hits — cross-linked to the new sourcing recipe as the downstream procurement step for the survivors

Flagged

  • None.

Verified (no changes)

  • Recheck slot not scheduled this run (recipe_recheck: no); no aging-recipe verification performed.

2026-08-02

Added

  • Handle missing endpoint data in a confirmatory trial (Problem class: Data analysis; Evidence: Proposed) — opens the clinical-biostatistics leg of the cookbook, which was empty: protocol design was covered, primary-endpoint analysis was not. Rung 2 on the Missing Data Sensitivity skill. The value is ordering: estimand written down before any method is chosen (a strategy named per intercurrent event, since discontinuation and rescue usually need different ones), Little’s MCAR test reported but explicitly barred from licensing MAR, MMRM in R mmrm with Kenward–Roger df rather than statsmodels.mixedlm (which has no such correction and is not confirmatory), the rbmi reference-based family run together rather than one pick, both Rubin’s and frequentist variances side by side, and the tipping point in residual-SD units. Provenance carries the RNG seed — multiple imputation is stochastic, and the sensitivity table is the one a reviewer will try to reproduce. Grounded in White, Joseph & Best 2020 (JR/CR/CIR as a causal model whose varied assumption is the tipping point), Cro, Carpenter & Kenward 2019 (information anchoring), Cro et al. 2020, and von Hippel (required imputations grow quadratically with FMI, not linearly).
  • Define a phenotype cohort in an OMOP CDM database (Problem class: Data analysis; Evidence: Proposed) — first recipe to touch the OMOP/OHDSI leg, pairing the OMOPHub MCP server with pyomop at rung 3. Neither alone works: without a vocabulary service the model supplies concept ids from memory and a wrong-but-plausible id produces a query that runs cleanly and returns the wrong cohort; without a database client you have a concept set and nobody to count. The recipe pins the vocabulary release first, forbids stating a concept_id not retrieved from the server, keeps include_descendants as a per-concept decision, commits the codes you excluded as well as those you kept, runs check_sql before run_sql against a read-only role, and demands an attrition table rather than a count. Evidence: Kashyap et al., JAMIA 2020 — OHDSI phenotype classifiers degraded across sites (recall −0.08 domestically, −0.18 internationally), concluding the definition should travel rather than the fitted artifact; EHRSQL and its 2024 shared task — confident answers to unsupported questions are the deployment-critical failure in EHR text-to-SQL, hence execution-based verification.

Updated

  • Recipes curator state — banked three Translational Medicine follow-ups: cross-ancestry association lookup across the FinnGen / BioBank Japan / eQTL Catalogue connectors (with a caveat to settle first — Claude Science connectors have no addressable MCP URL, so a recipe built only on them cannot produce the committed artifact the reproducibility rule requires), an adaptive-trial-design page, and a CONSORT trial-report page that must justify itself against the new missing-data recipe before it is written.

Flagged

  • None.

Verified (no changes)

  • No aging-recipe recheck this slot (recipe_recheck: no); the run went to discovery.

2026-08-02

Neuroscience directed pass. Two new recipes, both closing previously-uncovered GA skills in the largest catalog section (96 Neuroscience-tagged tools, 12 recipes touching the area before today). Both cover the animal-experiment layer the cookbook had skipped: the behavior the neural data gets regressed against, and the imaging route to a population-activity matrix.

Added

  • Track animal pose in behavioral video (Problem class: Data analysis; Evidence: Proposed) — rung-2 DeepLabCut skill recipe. The ordering is the point: run a SuperAnimal zero-shot probe with a labeled preview video first and skip the labeling project entirely if it tracks cleanly — SuperAnimal covers 45+ species with no new labels and is 10–100× more data-efficient than transfer learning when fine-tuned (Ye et al., Nat. Commun. 2024). If labeling is needed, label for diversity across animals, lighting and camera position rather than volume: 25%-participant models held up for adult joint angles but failed on clinical parameters and toddlers where 75% models reached good-to-excellent agreement against Vicon (Verhoeven et al., J. Biomech. 2025). Output is a committed run_pose.py emitting a keypoint table whose low-likelihood frames are NaN, never interpolated, a pose_qc.csv (% frames masked, longest masked run per bodypart), kinematics in mm, and a provenance.json recording the model snapshot and test error. Accuracy ceiling is documented — 0.228 cm median error, 2.0% of range of motion, against XROMM (Moore et al., J. Exp. Biol. 2022). Fully open, Workstation with GPU.
  • Extract neuronal traces from a calcium imaging movie (Problem class: Data analysis; Evidence: Proposed) — rung-2 Calcium Imaging Analysis Guide skill recipe. Every rule in the skill branches on the preparation, so the recipe makes you declare it first (CNMF-E for one-photon/miniscope where out-of-focus background dominates, CNMF/sparse-NMF/Cellpose for two-photon; non-rigid for awake behaving, rigid for anesthetized). Neuropil subtraction goes before dF/F, not after; ROI footprints get inspected over the mean/max projection before the trace heatmap, because segmentation failures are obvious there and invisible downstream; rejected_rois.csv carries a reason string per rejection and is treated as part of the result. Deconvolution is a separate, optional, differently-named output labelled an estimated firing rate — the distinction that ground-truth work makes concrete: CASCADE, trained on >35 h from 298 neurons, infers absolute rates and beats model-based algorithms (Rupprecht et al., Nat. Neurosci. 2021). Component evidence anchored on CaImAn’s near-human ROI detection against multi-labeler annotations on nine two-photon datasets (Giovannucci et al., eLife 2019). Fully open with a copyleft note (Suite2p GPL-3.0, CaImAn GPL-2.0, conda-only), Workstation with GPU where RAM and NVMe are the real limits.

Both are Proposed rather than Reported: no documented attempt at driving either skill on a real dataset is known, and both skills’ text is AI-generated and not expert-reviewed. Each recipe says so and tells the reader which numbers to re-derive — for the calcium page, the neuropil coefficient specifically, since it shifts every correlation in the dataset.

Updated

  • None.

Flagged

  • None.

Verified (no changes)

  • No recheck this slot (recipe_recheck: no) — the full run went to the directed pass.

2026-08-01

Directed pass on Molecular and Cellular Biology. Two GA bioSkills single-cell skills were catalogued but had no recipe; both now do. Aging-recipe verification was skipped this slot (recipe_recheck: no).

Added

  • Analyze a Perturb-seq CRISPR screen for perturbation effects — Data analysis · One skill or MCP · Proposed · Fully open · Workstation with GPU. Rung-2 recipe on the Perturb-seq Analysis skill: mixture-model posterior guide assignment with the flat threshold demoted to a disagreement check, Mixscape escaper classification that reports a high escaper fraction as low editing efficiency rather than as a null result, and calibrated testing at the replication unit that actually exists — the transfection, not the cell. Per-cell rank_genes_groups is explicitly forbidden as the primary test, because low-MOI association testing produces excess false positives from sparsity, confounding and model misspecification (Barry et al., Genome Biol 2024). Effect sizes come from E-distance in a single fixed PCA embedding reused across all perturbations (Peidli et al., Nat Methods 2024), and any perturbation-prediction model must clear an additive baseline first — five foundation models plus two other deep-learning models failed to (Ahlmann-Eltze et al., Nat Methods 2025).
  • Build a clonal lineage tree from CRISPR scar recorders — Data analysis · One skill or MCP · Proposed · Fully open · Workstation with GPU. Rung-2 recipe on the Lineage Tracing skill, built around the fact that a single-solver tree looks authoritative and carries no indication of how much signal was behind it. The assay is declared up front because a static barcode groups clones but licenses no branching order; the filter log is treated as part of the result rather than as a diagnostic, since the upstream pipeline itself changes relative method performance (Dai & Molloy 2026); state frequencies serve as the homoplasy early warning; and three solvers are kept as separate Newick outputs with pairwise Robinson–Foulds and triplets-correct agreement instead of one “best” tree (Jones et al., Genome Biol 2020). CoSpar supplies clonal fate bias, with scVelo/CellRank disagreement recorded rather than resolved (Wang et al., Nat Biotechnol 2022). Provenance records the Cassiopeia commit SHA, because 2.0+ installs from a git ref and there is no version to pin.

Notes

  • Proteomics remains the largest Molecular and Cellular Biology hole — MaxQuant and PRIDE are catalogued and uncovered, and the leg is likely rung 3. Filed as the first pick for the next slot, along with the open question of whether doublet detection and Harmony should become recipes or become steps inside the existing QC and integration recipes.

2026-08-01

Directed pass on Integrative Structural and Computational Biology. Two of the 51 ISCB-tagged catalog tools that no recipe had yet reached — the bioSkills structural-biology pair — now have pages, filling the two steps that sit in front of every structure-based drug-design workflow the cookbook already covers: is this structure trustworthy, and where on it do I aim?

Added

  • Detect and rank druggable pockets on a protein structure (Problem class: Data analysis; Evidence: Proposed) — rung-2 Binding Site Detection skill. Enumerate cavities with fpocket, rank ligandability with P2Rank, and keep the two as separate columns reconciled by centroid proximity and lining-residue overlap rather than averaged into one score — that design is what the largest benchmark to date supports (13 predictors over ~30,000 complexes; fpocket rescored by PRANK or DeepPocket reaches 60% recall versus 39% for the weakest method, and stronger scoring alone is worth up to +14% recall and +30% precision — Utgés & Barton, J. Cheminform. 2024). Anchors against known sites via UniProt features or a holo PDB ligand before trusting rank 1, emits a docking_box.json that drops straight into Vina/smina/DiffDock, and escalates to mdpocket over an MD ensemble only when the apo surface is genuinely flat. The apo caveat is cited as measured fact, not folklore (CryptoBench, 1,107 apo–holo pairs), as is the degradation on membrane targets (Pliushcheuskaya & Künze 2025). Fully open, Laptop.
  • Vet a PDB structure before you build on it (Problem class: Data analysis; Evidence: Proposed) — rung-2 Structure Validation skill. The recipe’s point is that global metrics do not license a local measurement: it reads the R-free gap rather than R-work, z-scores B-factors within the structure rather than against an absolute cutoff, and then answers the per-residue question — occupancy, altlocs, truncated side chains, local rotamer status — for the residues you actually care about. Method-specific reading for cryo-EM (FSC 0.143 global versus 0.5 map-model, with local resolution worse than either) and NMR ensembles. Criteria are MolProbity’s (Williams et al., Protein Sci. 2018) and remain the archive’s own direction of travel (wwPDB IHMValidation, 2026). Flagged caveat: Phenix is free for academic use only, and the recipe requires you to record whether MolProbity ran or the coarse Bio.PDB fallback did. Fully open, Laptop.

Updated

  • Triage an AlphaFold model for structure-based drug design — cross-linked to both new recipes. Its step 3 asks the reader to supply a pocket residue list “from a pocket-prediction tool”; the pocket recipe is now that tool. Its step 4 sends readers to the PDB; the vetting recipe is now what to do when they get there.

Verified (no changes)

  • recipe_recheck: no this slot — no aging-recipe verification pass was run.

2026-08-01

Added

  • Predict checkpoint-blockade response for a tumor from its biomarker profile (Problem class: Knowledge synthesis; Evidence: Proposed) — Immunology and Microbiology focus-day recipe closing the top-of-queue deferred item and completing the immuno-oncology trio alongside the neoantigen and CIViC-actionability recipes. Rung 2: the ToolUniverse Immunotherapy Response Prediction skill turns a per-sample biomarker table (TMB with assay, MMR/MSI with method, PD-L1 with antibody clone and scoring convention kept un-converted, HLA genotype, resistance genotypes) into per-axis component scores with T1–T4 evidence grades — never the integrated 0–100 score alone — plus a direction-disagreement flag as the human-review trigger, emitted through a committed .claude/commands/ici-response.md with ici_biomarkers.csv and a provenance.json recording each knowledgebase’s release date. Marked Proposed because the skill’s integrated score is not benchmarked against outcomes; the component axes are strong: TMB and the T-cell-inflamed GEP are independently predictive and only weakly correlated across >300 samples, 22 tumor types, 4 KEYNOTE trials (Cristescu et al., Science 2018) — the finding that motivates keeping the axes separate — with GEP development in Ayers et al., JCI 2017, TMB-H alone insufficient across 5,621 metastatic breast cancers (Sammons et al. 2023), and response heterogeneity within dMMR/MSI-H at 55.6% vs 32.8% in IMvigor210 (Kim et al., Cancers 2025).

Updated

  • Prioritize tumor neoantigens for a personalized cancer vaccine — cross-linked to the new checkpoint-blockade recipe (the ICI-eligibility read, where TMB summarizes the neoantigen supply this recipe enumerates).
  • Annotate tumor somatic variants with clinical actionability evidence — cross-linked to the new checkpoint-blockade recipe as the immunotherapy-eligibility read of the same tumor profile.

Flagged

  • None.

Verified (no changes)

  • recipe_recheck: no this slot — aging-recipe verification skipped; the run went to discovery in the Immunology and Microbiology focus area.

2026-08-01

Directed pass on Immunology and Microbiology. Two recipes shipped, both closing previously-uncovered GA bioSkills immunoinformatics skills.

Added

  • Prioritize tumor neoantigens for a personalized cancer vaccine (Problem class: Experimental design; Evidence: Validated) — rung-2 recipe on the Neoantigen Prediction skill. Takes a VEP-annotated somatic VCF, four-digit HLA genotype, and tumor expression through pVACseq/pVACfuse, then applies the filters that actually decide the shortlist — cancer cell fraction, transcript expression, agretopicity, reference-proteome similarity, and HLA loss-of-heterozygosity (routed to a separate rejected_loh.tsv rather than silently dropped) — into a tiered neoantigens.tsv with per-exclusion reason strings and a pinned provenance.json. Keeps the pVACview/IGV manual review as a mandatory step. Validated by clinical use: ImmunoNX has supported 185+ patients across 11 trials (322 → 78 candidates on HCC1395), and a pVACtools-designed TNBC DNA vaccine induced neoantigen-specific T-cell responses in 14/18 patients.
  • Annotate TCR antigen specificity by clustering and database lookup (Problem class: Data analysis; Evidence: Validated) — rung-2 recipe on the TCR-Epitope Binding skill. Clusters CDR3 sequences within a cohort with tcrdist3 and GLIPH2 (kept as separate columns, because no single clusterer dominates), matches VDJdb/IEDB/McPAS-TCR with confidence filtering, and propagates hits to clusters only as candidate annotations with a supporting-member fraction. Explicit that supervised per-pair prediction on unseen epitopes is not defensible. Grounded in a nine-method benchmark over 190,670 TCRs and 2,313 epitopes and in GLIPH2 antigen discoveries in lung cancer and tuberculosis.

Updated

  • Analyze a single-cell TCR repertoire alongside gene expression — added a See also link to the new TCR-specificity recipe, which now covers the antigen-annotation handoff the page previously described only as a next step.

Verified (no changes)

  • No aging-recipe recheck this slot (recipe_recheck: no).

2026-07-27 (user request #74)

Added

  • Localize a fiber/probe implant tip in an Allen CCF subregion from 2D histology (Problem class: Data analysis; Neuroscience; Evidence: Proposed) — QUINT-replacement screening recipe for @goodb (#74). The earlier blocked note is now resolved: DeepSlice and brainglobe-atlasapi are pip-installable, so they are carried as recipe dependencies rather than uncatalogued components, and the recipe ships as a rung-3 multi-tool harness — DeepSlice batch coronal cutting-plane prediction → scikit-image skill per-section fiber-tip detection → tip pixel mapped through the QuickNII-compatible anchoring into CCF → brainglobe-atlasapi allen_mouse_25um voxel→structure lookup → placement table with subregion, hit/miss-vs-target verdict, and a distance-to-boundary margin (marginal calls flagged, not binarized), captured as committed scripts + pinned env + provenance. Screening-grade, not publication-grade (a single 2D section inherits AP error from plane prediction; the full QUINT GUI with VisuAlign nonlinear correction remains the alternative for the definitive figure). Proposed — component-level evidence strong (DeepSlice, Nat Commun 2023; brainglobe-atlasapi, JOSS 2020; QUINT, Front. Neuroinform. 2019), exact three-component assembly not separately attempted. License correction: DeepSlice is GPL-3.0-only per PyPI (not MIT as previously banked) — recipe stays Fully open, copyleft flagged. Fully open, Laptop.

2026-07-26 (Drug Repurposing and Discovery directed pass)

Added

  • Predict ADMET properties for a compound series with an ML predictor (Problem class: Data analysis; Evidence: Reported) — rung-2 ADMETlab MCP server recipe: SMILES CSV → wash_molecule standardization → batched predict_admet/fetch_admet_csv over ~119 ML endpoints (CYP1A2/2C9/2C19/2D6/3A4, hERG, HIA, F20/F30, Caco-2, logS, CL, t½, PPB, BBB, Ames) with per-endpoint uncertainty → committed predict_admet.py + pinned requirements.txt + series_admet.csv (red/amber/green triage per endpoint, n_red rollup) + provenance.json (MCP server version, ADMETlab 3.0 model version, UTC query date, input sha256, model id). Fills the long-standing ML-ADMET-predictor gap now that the ADMETlab MCP is catalogued (verified 2026-07-26) — the ML-prediction layer between rule-based filter-virtual-screening-hits and descriptor/analog estimate-pk-properties, and distinct from model-benchmarking benchmark-admet-property-with-pytdc (all cross-linked). Reported — ADMETlab 3.0 is peer-reviewed (multi-task DMPNN, >400k entries, 119 endpoints, per-prediction uncertainty; Fu et al., NAR 2024); the exact Claude+ADMETlab-MCP assembly is not separately benchmarked. Cloud-side data-residency caveat noted. Fully open; Laptop.

Verified (no changes)

  • Aging-recipe sweep: oldest last_verified across the cookbook is 2026-07-04 (≤30 days), so no recipes were due for re-verification this run.

2026-07-26 (Translational Medicine directed pass)

Added

  • Annotate tumor somatic variants with clinical actionability evidence (Problem class: Knowledge synthesis; Evidence: Reported) — rung-2 CIViC connector recipe: tumor somatic variant list (gene + HGVS, variant type) → CIViC clinical-evidence match per variant → committed annotate_actionability.py + pinned env + actionability.csv (evidence type Predictive/Prognostic/Diagnostic/Predisposing/Oncogenic, disease, therapy/outcome, A–E evidence level, citation, CIViC evidence id) + suggested AMP/ASCO/CAP tier (I/II/III) + provenance.json (CIViC release date, connector identity, input sha256, versions). Fills the previously-uncovered precision-oncology actionability gap; the somatic/therapy-oriented counterpart to the germline interpret-clinical-variant recipe, upstream of match-patient-to-clinical-trials, and CNV sibling of detect-somatic-cnvs-from-tumor-sequencing (all cross-linked). Reported — CIViC is among the most-used MTB annotation knowledgebases (Borchert et al., Brief. Bioinform. 2021; Lutz et al., Target. Oncol. 2025) with a quantified programmatic-match step (38.6% of 59,437 GENIE tumors matched CIViC Level A/B, >1,200 variants/s; Wagner et al., JCO CCI 2020) and documented evidence-type/level schema (Krysiak et al., NAR 2023); the exact Claude+CIViC assembly is not separately benchmarked and requires human review before clinical use. Institutional access; Laptop.

Verified (no changes)

  • 7 aging recipes spot-checked and last_verified bumped to 2026-07-26: Design CRISPR sgRNAs for a gene knockout and Annotate and verify an engineered plasmid construct (SciAgent-Skills repo resolves, sgRNA/pLannotate catalog pages GA + security-cleared, canonical sources intact); Tile and stain-normalize a whole-slide image for ML, Rank a compound library against a target by predicted binding affinity, Profile a compound’s polypharmacology from ChEMBL bioactivity data, Register longitudinal medical scans to a common frame, and Harmonize free-text clinical terms to standard codes (all linked catalog pages — histolab, pathml, Boltz, gget, ChEMBL, SimpleITK, Medical Terminologies MCP, PyHealth, icd-10-codes, PubChem, DrugBank, Open Targets, ToolUniverse — resolve and none flagged).

2026-07-26 (Neuroscience directed pass)

Added

Verified (no changes)

  • 1 aging recipe spot-checked and last_verified bumped to 2026-07-26: Discover NWB recordings on DANDI and prepare them for sorting (all catalog links — Neurosift Tools, OpenNeuro, neuropixels-analysis, AIND Data — resolve and none flagged; canonical journal sources intact).