Recipes updates archive

Older entries rotated out of RECIPES_CHANGELOG.md. Newest first, same format.

2026-07-25 (Molecular and Cellular Biology directed pass)

Added

  • Quantify a western blot from a scanned image (Problem class: Data analysis; Evidence: Reported) — rung-2 Western Blot Quantification skill recipe: uncompressed blot scan + lane map → analyze_pixel_distribution/find_roi_from_image band detection → two-step normalization (target → loading control → control lane, total-protein preferred) → replicate aggregation → committed quantify_blot.py + pinned env + blot01_quant.csv/replicate_summary.csv + roi_overlay.png + provenance.json (skill commit, versions, image sha256, ROI/background/normalization method). Fills the previously-uncovered western-blot catalog skill; cross-linked to the microscopy cell-segmentation recipe as its densitometry counterpart. Reported — integrated-density quantification with total-protein-over-housekeeping normalization is field-standard and directly evidenced (Maloy et al., Anal. Biochem. 2022; Kirshner & Gibbs, Mol. Cell. Endocrinol. 2018; Maestri et al., PLOS ONE 2025); the agent path is not separately benchmarked. Fully open; Laptop.
  • Track single particles and measure their diffusion coefficient (Problem class: Data analysis; Evidence: Reported) — rung-2 trackpy skill recipe: video-microscopy stack + pixel size / frame interval / diameter → Crocker–Grier detection (tp.locate) → trajectory linking (tp.link + tp.filter_stubs) → MSD fit (tp.emsd/tp.imsd) for diffusion coefficient D and anomalous exponent → committed track_and_diffuse.py + pinned env + trajectories.csv/msd.csv/diffusion.csv + tracks_overlay.png + provenance.json (skill commit, trackpy/pims versions, pixel size, frame interval, linking params, stack sha256). Fills the previously-uncovered trackpy catalog skill; cross-linked to the microscopy cell-segmentation recipe as its motion-across-frames counterpart. Reported — locate→link→MSD→diffusion-mode is the canonical SPT chain (Kusumi et al., Biophys. J. 1993; Taylor et al., Methods Mol. Biol. 2019; Vu et al., BBA Biomembranes 2021); the agent path is not separately benchmarked. Fully open; Laptop.

Verified (no changes)

  • 8 aging recipes spot-checked (oldest last_verified first), all catalog links resolve and none flagged; last_verified bumped to 2026-07-25: Predict the regulatory effect of a non-coding variant, Detect somatic copy-number variants from tumor sequencing, Predict hospital readmission from EHR data, Segment an organ or tumor in a medical image with nnU-Net, Fit a survival model to censored clinical outcomes, Segment and quantify cells in a microscopy image, Convert raw analytical instrument data to Allotrope ASM JSON, Parse FCS flow-cytometry files for downstream immunophenotyping.

2026-07-25 (Integrative Structural and Computational Biology directed pass)

Added

  • Build a protein interaction network and rank hub genes (Problem class: Data analysis; Evidence: Reported) — rung-2 STRING skill recipe: gene-symbol list (DE hits / module / screen set) + species + high-confidence cutoff (combined score ≥0.7) → STRING REST network retrieval (results/string_edges.tsv) → deterministic pinned rank_hubs.py (networkx degree/betweenness/closeness centrality replicating CytoHubba “Degree”, Louvain modules) → results/hub_genes.csv / top_hubs.csv + ppi_network.png → synthesis anchored only to the tables → provenance.json (STRING skill commit, STRING database version + query date, species, score cutoff, gene-list sha256, library versions, model id). Fills the previously-uncovered STRING catalog tool; complements the term-overlap run-functional-enrichment-on-a-gene-list (find central genes vs. name pathways) and takes a gene set from map-disease-to-genes-and-pathways (all cross-linked). Reported — the DEGs→STRING PPI→CytoHubba hub-ranking chain is one of the most heavily documented systems-biology patterns (Hasan et al., Sci. Rep. 2023; Jiang et al., Int. Immunopharmacol. 2024; Mao et al., Sci. Rep. 2025); the exact Claude+STRING-skill composition is not separately benchmarked. Fully open; Laptop.

Verified (no changes)

  • 3 Integrative Structural and Computational Biology recipes spot-checked, last_verified bumped to 2026-07-25: Predict an antibody–antigen complex to map an epitope (Boltz plugin + gget catalog links and both benchmark sources resolve), Superpose two protein structures and quantify where they differ (PyMOL / AlphaFold / PDB catalog links resolve), and Set up a protein molecular dynamics simulation in GROMACS from a PDB ID (molecule-mcp bundle link resolves).

2026-07-25 (Immunology and Microbiology directed pass)

Added

  • Scan a protein for candidate CD4 T-cell (helper) epitopes (Problem class: Experimental design; Evidence: Validated) — rung-2 MHC Class II Prediction skill recipe: antigen FASTA + HLA-DR/DQ/DP allele list → 15-mer sliding-window scoring with NetMHCIIpan-4.3 (+ optional MixMHC2pred-2.0) → register-aware 9-mer core extraction → class II %Rank cutoffs (≤1% strong / ≤5% weak) → committed scan_classII_epitopes.py + pinned env + epitopes_classII.csv + provenance.json (predictor versions/data release, allele strings, window, input sha256, run date, model id). CD4/MHC-II sibling of the existing scan-protein-for-cd8-t-cell-epitopes recipe (cross-linked both ways); run both for a full vaccine-antigen analysis. Validated — NetMHCIIpan is the field-standard CD4 predictor and the exact predictor→shortlist step is used in current published vaccine design (Yakobi & Nwodo, Biochem. Biophys. Rep. 2025; coverage-optimized peptide sets T-cell-confirmed in Nilsson et al., Front. Immunol. 2021); class II remains materially less accurate than class I, and the agent path is not separately benchmarked. Fully open; Laptop.
  • Profile shotgun metagenome taxa with Kraken2 and Bracken (Problem class: Data analysis; Evidence: Reported) — rung-2 Kraken2 Metagenomic Classification skill recipe: quality-trimmed, host-depleted shotgun FASTQs + named reference database → Kraken2 (--confidence 0.2) read classification → Bracken species-level abundance re-estimation → merged bracken_species.tsv → committed profile_metagenome.sh + pinned environment.yml + provenance.json (kraken2/bracken/KrakenTools versions, database name + build/download date, confidence threshold, per-sample classified fraction, FASTQ sha256s, run date, model id). Shotgun-reads upstream of the 16S diversity recipe’s abundance-table analysis (cross-linked both ways); complements the AMR/virulence recipe. Reported — Kraken2→Bracken is the field-standard fast profiler; database+confidence-score dependence pinned per Liu et al., aBIOTECH 2024, applied at 2,101-metagenome scale in Li et al., Gastroenterology 2025; the agent path is not separately benchmarked. Fully open; Workstation with GPU (RAM/disk floor, not GPU).

Verified (no changes)

  • 2 Immunology recipes spot-checked, last_verified bumped to 2026-07-25: Compute 16S microbiome alpha/beta diversity (scikit-bio skill + Biomni links resolve; added shotgun cross-link) and Scan a protein for candidate CD8 T-cell epitopes (MHC Binding Prediction / Epitope Prediction links resolve; added CD4 sibling cross-link).

2026-07-25 (Chemistry directed pass)

Added

  • Train a QSAR model from your own assay data and predict untested compounds (Problem class: Data analysis; Evidence: Proposed) — rung-2 ChemLint MCP recipe: assay CSV (smiles + measured activity) + untested library → standardize/QC + activity-cliff/outlier flagging → ECFP4 + MACCS featurization → scaffold-based train/test split → Random Forest / gradient-boosting with 5-fold CV, hyperparameter tuning, and Y-randomization chance-correlation control → Tanimoto applicability-domain gate → committed qsar_train.py + pinned requirements.txt + model.pkl + validation report + predictions/scored_library.csv + provenance.json (ChemLint commit, RDKit/scikit-learn versions, descriptor set, algorithm, split seed, Y-scramble result, AD cutoff, input sha256, run date, model id). Ligand-based counterpart to the structure-based rank-compound-library-by-predicted-affinity; predictive sibling of analyze-sar-of-a-compound-series and bespoke-endpoint sibling of benchmark-admet-property-with-pytdc (all cross-linked). Proposed — no documented Claude+ChemLint attempt; grounded on a current field-standard exemplar (MACCS-RF QSAR with CV, applicability domain, Y-randomization, SHAP for SMYD3 modulators, Alzahrani et al., Mol. Divers. 2026) and the ChEMBL-anchored ML-QSAR repurposing pattern (Kamboj et al., CSBJ 2022, 10-fold CV r 0.80–0.92). Fully open; Laptop.

Verified (no changes)

  • 3 Chemistry recipes spot-checked, all current, last_verified bumped to 2026-07-25: Filter a virtual screening hit list (MedChem/Datamol/RDKit links + PAINS/BRENK/Lipinski/Veber sources resolve; K-Dense marketplace live), Identify an unknown compound from an MS/MS spectrum (matchms skill + pyOpenMS/PubChem links resolve), Estimate pharmacokinetic properties of a small molecule (RDKit/MedChem/ChEMBL links resolve).

2026-07-19 (Drug Repurposing and Discovery directed pass)

Added

  • Analyze the SAR of a measured compound series (Problem class: Data analysis; Evidence: Reported) — rung-2 SAR Analysis skill recipe: a CSV of assayed analogs (smiles + IC50/Ki) → optional datamol standardization → maximum-common-substructure scaffold detection + R-group decomposition (RDKit rdFMCS/rdRGroupDecomposition) → substituent-vs-pIC50 table per R-position → single-change activity-cliff flagging ( ΔpIC50 ≥ 1) → committed sar_analysis.py + pinned requirements.txt + sar_table.csv/activity_cliffs.csv + aligned-structure HTML report + provenance.json (RDKit/pandas versions, skill commit, MCS threshold, cliff cutoff, input sha256, run date, model id). Retrospective analyze-what-you-measured sibling of the forward enumerate-analogs recipe (cross-linked both ways). Reported — R-group decomposition and matched-pair single-substituent comparison are field-standard for lead optimization (Raut & Dixit, RSC Med. Chem. 2025; Ding et al., Curr. Med. Chem. 2020; Kombo & LaMarche, J. Med. Chem. 2025) and the BixBench-evaluated skill calls validated RDKit routines; the Claude-driven path is not separately benchmarked. Fully open; Laptop.

Verified (no changes)

2026-07-19 (Translational Medicine directed pass)

Added

  • Diagnose a rare disease from patient phenotypes (Problem class: Knowledge synthesis; Evidence: Reported) — rung-2 ToolUniverse Rare Disease Diagnosis skill recipe: patient HPO terms (+ optional candidate variants) captured in a committed case.yaml → skill federates HPO/Orphanet/OMIM/DisGeNET disease+gene lookup → gene prioritization (MARRVEL, ClinGen validity, GTEx tissue expression) → per-variant ACMG interpretation (ClinVar, gnomAD, EVE/SpliceAI) → tiered (T1–T4) reports/*.md + raw results/*.json audit trail + provenance.json (tooluniverse version, skill commit, per-database query dates, ClinVar/gnomAD release accessions, input sha256, model id). Phenotype-first sibling of interpret-clinical-variant (cross-linked both ways). Reported — knowledge-grounded LLM layers over the same MARRVEL/ClinVar/gnomAD tool stack give +12–15 pp Recall@1 for phenotype-driven gene prioritization (LA-MARRVEL, Lee et al., arXiv 2511.02263, 2025-11 / rev 2026-03), and tool grounding is what carries the workflow (MARRVEL-MCP 95% vs 33% without tools, bioRxiv 2025-11-28); rung-1 baseline fails (database-free GPT-4 ~16%, Kim et al., 2024). The exact ToolUniverse-skill composition is not separately benchmarked. Fully open; Laptop.

Verified (no changes)

  • 1 recipe spot-checked, last_verified bumped to 2026-07-19: interpret-clinical-variant — BioMCP catalog page resolves and not flagged, MARRVEL-MCP source URL still loads; added cross-link to the new rare-disease recipe.

2026-07-19 (Neuroscience directed pass)

Added

  • Build a resting-state functional-connectivity matrix from preprocessed fMRI (Problem class: Data analysis; Evidence: Reported) — rung-2 Nilearn skill recipe: fMRIPrep-preprocessed BOLD + confounds TSV → named atlas (Schaefer 2018) NiftiLabelsMasker (standardize/detrend/band-pass) → explicit fMRIPrep confound strategy (motion + cosines + aCompCor + non-steady-state) → ROI time-series with confounds regressed → Pearson + partial-correlation ROI-to-ROI matrices → network-ordered heatmap + FD/scrubbing QC → committed build_connectome.py + pinned requirements.txt + connectivity_pearson.csv/connectivity_partial.csv + figure + provenance.json (Nilearn version, atlas name/release, confound columns, band-pass/TR, connectivity kind, input sha256s, run date, model id). fMRI counterpart to the EEG ERP-extraction recipe (cross-linked). Reported — Nilearn is the field-standard connectivity library (Abraham et al., Front. Neuroinform. 2014) and the atlas→confound-regression→correlation workflow is canonical (Kumar et al., PLoS Comput. Biol. 2020) and reused in current rs-FC studies (Messina et al., Neurology 2026; Dai et al., J. Affect. Disord. 2026); the Claude+Nilearn-skill assembly is not separately benchmarked. Fully open; Laptop.

Verified (no changes)

2026-07-18 (Molecular and Cellular Biology directed pass)

Added

  • Integrate multi-omics layers into interpretable factors with MOFA+ (Problem class: Data analysis; Evidence: Proposed) — rung-2 MOFA+ skill recipe: 2+ omics views on the same samples/cells (muon/MuData for single-cell multi-modal, AnnData dict for bulk) → per-view feature selection → build MOFA+ model → train (mofapy2, fixed factor count + seed) → per-factor per-view variance decomposition → factor–metadata association → top loadings per factor → committed mofa_run.py + pinned requirements.txt + mofa_model.hdf5 + variance_explained.csv/factor_metadata_assoc.csv/top_loadings_<view>.csv + figure + provenance.json (mofapy2/muon versions, factor count + seed, per-view feature counts, input sha256s, run date, model id). Cross-modality factorization complement to the within-modality scVI batch-integration recipe; feeds the functional-enrichment and TF/pathway-activity recipes (all cross-linked). Proposed — no documented Claude+MOFA+-skill attempt; grounded on the canonical method (Argelaguet et al., Genome Biology 2020) and a current applied exemplar (12-factor MOFA+ decomposition of 667 TCGA gliomas validated across n=1685 without retraining; Saleh et al., Cancers 2026). Fully open; Laptop.

Updated

  • infer-gene-regulatory-network-from-scrnaseq — fixed a broken feedback-footer URL (was pointing at the non-canonical goodb.github.io/github.com/goodb host; now scripps-ai-enablement); last_verified → 2026-07-18 (Arboreto/AnnData/Scanpy catalog pages resolve, method sources current).
  • assemble-reference-atlas-from-cellxgene-census, compute-hrv-from-ecg-recording — fixed the same broken feedback-footer host (goodb → scripps-ai-enablement); no other changes.

Verified (no changes)

  • 4 additional MCB recipes spot-checked (oldest-first), all current; last_verified bumped to 2026-07-18: run-bulk-rnaseq-differential-expression (PyDESeq2), run-functional-enrichment-on-a-gene-list (gget), infer-tf-and-pathway-activities-from-expression (decoupler-MCP), annotate-cell-types-in-single-cell-data (CellTypist/popV) — all linked catalog pages resolve, sources current.

2026-07-18 (Integrative Structural and Computational Biology directed pass)

Added

  • Design amino-acid sequences for a fixed protein backbone (Problem class: Experimental design; Evidence: Reported) — rung-3 two-model toolbelt: ProteinMPNN skill samples sequences for a target backbone .pdb (fixed catalytic/interface positions, temperature sweep) → ESMFold skill refolds every design → self-consistency gate (Cα-scRMSD < 2.0 Å AND mean pLDDT > 80) keeps only foldable candidates → ranked survivors → committed .claude/commands/mpnn-design.md + pinned skill envs + designs/<name>_mpnn.fasta + results/<name>_selfconsistency.csv + provenance.json (model/versions, sampling settings, cutoffs, backbone sha256, run date, model id). First recipe to compose the ProteinMPNN family; cross-links LigandMPNN/SolubleMPNN variants, AlphaFold2 as a stricter second gate, and the score-protein-variants-with-esm sibling. Reported — ProteinMPNN is the validated field-standard inverse-folding model (Dauparas et al., Science 2022), the design→refold→filter self-consistency routine is standard practice (Lin et al., Science 2023), and a ProteinMPNN redesign of a flavin-binding fluorescent protein was wet-lab confirmed (Nikolaev et al., Protein Sci. 2024); the Claude-skill assembly is not separately benchmarked. Fully open; Workstation with GPU.

Verified (no changes)

  • 2 recipes spot-checked (oldest-first, ISCB-focused), all current; last_verified bumped to 2026-07-18: predict-rna-secondary-structure-and-accessibility (ViennaRNA skill catalog page resolves; SciAgent-Skills repo live), infer-protein-function-from-structure (Foldseek skill catalog page resolves; Foldseek Search web service still up).

2026-07-18 (Immunology and Microbiology directed pass)

Added

  • Analyze a single-cell TCR repertoire alongside gene expression (Problem class: Data analysis; Evidence: Reported) — rung-2 scirpy Analysis skill recipe: 10x/AIRR single-cell VDJ + matching clustered .h5ad → pp.index_chains + barcode-matched modality pairing → chain_qc filtering of multichain doublets/orphan cells → exact-CDR3-nt clonotype definition (TCR-appropriate, not BCR distance clustering) → clonal expansion + per-cluster/per-condition diversity + repertoire overlap → clonality overlaid on the transcriptomic UMAP → committed sc_tcr.py + pinned requirements.txt + clonotypes.csv + diversity/overlap tables + figure + provenance.json (scirpy/scanpy/mudata versions, clonotype strategy + params, input sha256s, run date, model id). T-cell, transcriptome-integrated counterpart to the B-cell reconstruct-bcr-clonal-lineages recipe; downstream of qc-single-cell-rna-seq and annotate-cell-types-in-single-cell-data (all cross-linked). Reported — scirpy is the scverse-standard single-cell TCR tool with a published benchmark (Sturm et al., Bioinformatics 2020) and the exact multi-modal workflow is a 2025 methods protocol (Plattner, Sturm & Rieder, Methods Cell Biol. 2025); the agent-driven skill assembly is not separately benchmarked. Fully open; Laptop.

Verified (no changes)

  • 3 recipes spot-checked (oldest-first), all current; last_verified bumped to 2026-07-18: assemble-reference-atlas-from-cellxgene-census (Census 2025-11-08 LTS still current; all catalog links resolve), annotate-a-bacterial-genome (Bakta/Prokka catalog pages resolve; canonical sources stable).

2026-07-18 (Chemistry directed pass; composition report #55)

Added

  • Plan a synthetic route for a target molecule (Problem class: Experimental design; Evidence: Reported) — rung-2 CovaSyn MCP recipe: target SMILES → optional RDKit/Datamol canonicalization → covaplatform retrosynthesis call (N routes, max depth) → per-step precursor + transform-class capture with buyable-leaf flags → route scoring (shorter, fully-buyable first) → committed plan_route.py + pinned requirements.txt + routes.csv/route_summary.csv + provenance.json (CovaSyn version, model/suite id, building-block catalog snapshot, request params, input sha256, run date, model id). Downstream make-check for enumerate-analogs-around-a-lead (cross-linked) and filter-virtual-screening-hits, with ChemCrow as the rung-4 execute-and-iterate alternative. Reported — agentic tool-grounded retrosynthesis is benchmarked near expert level (LARC, Baker et al., arXiv 2508.11860, 2025: 72.9% on 48 constrained tasks; ChemCrow, Nat. Mach. Intell. 2024); the exact Claude+CovaSyn pairing is not separately benchmarked. Subscription required (CovaSyn freemium, credit-metered, cloud SMILES submission); Laptop.

Updated

  • Prioritize targets within a disease via Open Targets — processed composition report #55 (@goodb, outcome=worked); added an Alzheimer’s-prevention field report (SORL1 lead via a prevention-tuned re-weighting of the pillar fields against MONDO_0004975) that reinforces the direct-GraphQL fallback when the hosted MCP is rate-limited; last_verified 2026-07-12 → 2026-07-18.

2026-07-16 (Translational Medicine scope; user request #52)

Added

  • Interpret variants that gain or lose glycosylation sites (Problem class: Knowledge synthesis; Evidence: Reported) — answers user request #52 from the GlyGen team. Rung-3 two-MCP toolbelt: GlyGen MCP get_protein_summary/get_site_summary for glycosite ground truth → Swiss-Prot numbering harmonization (the Asn135↔Asn167 antithrombin trap) → LOG/GOG classification (N-X-S/T sequon destroy/create) → BioMCP variant_searcher/variant_getter for ClinVar + AlphaMissense joins (GlyGen’s own variants come from EBI/BioMuta, so these add coverage) → expression sanity-check → committed glyco_variants.py + pinned requirements.txt + glyco_candidates.csv + provenance.json (GlyGen release + endpoint, BioMCP version, ClinVar/AlphaMissense snapshot dates, input sha256, run date, model id) + optional IEEE-2791 BioCompute Object. Cross-linked to interpret-clinical-variant and scan-antibody-glycosylation-sites. Reported — the GlyGen team documents this exact use case and ships a variants.ipynb reference notebook (SERPINC1 LOG / IFNGR2 GOG worked examples; Mazumder et al., Research Square 2026-07-01; glygener/colab-notebooks); the Claude-driven GlyGen-MCP+BioMCP assembly is not separately benchmarked. Fully open (GlyGen MCP Beta, wrapper repo no-LICENSE caveat); Laptop.

2026-07-12 (Drug Repurposing and Discovery directed pass)

Added

  • Find drug-repurposing candidates by walking a biomedical knowledge graph (Problem class: Knowledge synthesis; Evidence: Proposed) — rung-2 offline PrimeKG skill recipe: resolve a disease node → get_neighbors for disease genes → one-hop protein_protein expansion → drug_protein candidate drugs → drop drugs already indicated for the disease → score by number of connecting genes (hop-weighted), captured as a committed kg_repurpose.py + pinned requirements.txt + candidates.csv + provenance.json (skill commit, PrimeKG data version, resolved node id, input sha256, run date, model id). Offline, no-license, graph-connectivity complement to the quantitative target-first scan-drug-repurposing-candidates recipe (now cross-linked). Proposed — no documented attempt at the skill-driven graph-walk assembly; grounded on the same PrimeKG substrate that COMIC used to recover 21/30 recent FDA repurposing pairs (9.55% over SOTA; Aamer et al., BMC Bioinformatics 2026) and CellAwareGNN reports AUPRC 0.826 on (Zhang et al. 2026), with the honest caveat that the skill exposes neighbor lookups (interpretable heuristic), not a trained TxGNN-class predictor. Fully open; Laptop.

Updated

  • Prioritize targets within a disease via Open Targets — verified; last_verified 2026-06-20 → 2026-07-12 (linked catalog pages resolve and are unflagged; known-issue note on Open Targets MCP handshake still current).
  • Scan approved drugs for repurposing candidates against a disease — verified; last_verified 2026-06-28 → 2026-07-12; added See also cross-link to the new knowledge-graph repurposing recipe.

2026-07-12 (Translational Medicine directed pass)

Added

  • Triage GWAS lead SNPs to candidate drug targets (Problem class: Knowledge synthesis; Evidence: Reported) — rung-2 GWAS-MCP recipe: a version-controlled leads.csv of rsIDs → per-variant get_variant_info/annotate_snps consequence + nearest gene → get_eqtl_data eQTL-implicated gene(s) → query_gwas_catalog co-reported traits → get_drug_targets/search_open_targets tractability + clinical precedent, captured as a committed triage_gwas_leads.py + pinned requirements.txt + targets_triage.csv + provenance.json (gwas-mcp version, per-database snapshot dates, input sha256, run date, model id). Variant-first complement to the disease-first prioritize-targets recipe and upstream of build-target-dossier; scoped to the lookup-and-annotate layer (colocalization/MR flagged out of scope). Reported — reproduces the lookup layer of a quantitatively validated GWAS→target workflow (Lessard et al., BMC Genomics 2024: approved-target enrichment RR 2.58 vs 1.75 nearest-gene, >85% MR-directionality match), resting on the genetic-support premise (Nelson et al., Nat. Genet. 2015; King et al., PLoS Genet. 2019); the Claude+GWAS-MCP assembly is not separately benchmarked. Fully open; Laptop.

Updated

  • Build a target dossier — added a 2026-06 known-issue note to the Open Targets install step (hosted MCP endpoint fails the initialize handshake with JSON-RPC -32602, #43) with the direct-GraphQL / ToolUniverse workaround; last_verified → 2026-07-12.

Verified (no changes)

  • 5 aging recipes spot-checked, all current (last_verified → 2026-07-12): draft-phase23-clinical-trial-protocol, screen-polypharmacy-for-drug-interactions, profile-cancer-cohort-genomics-with-cbioportal, run-gwas-on-case-control-genotypes, compute-bacterial-pangenome-from-assemblies. Every linked catalog page resolves; source URLs load.

2026-07-12 (Neuroscience directed pass)

Added

  • Extract event-related potentials from EEG epochs (Problem class: Data analysis; Evidence: Reported) — rung-2 MNE-Python (EEG) skill recipe: raw continuous EEG + event markers → montage/reference → band-pass + notch → ICA artifact removal → epoching → per-condition evoked averaging, captured as a committed preprocess_erp.py + pinned env + *_ave.fif/erp_counts.csv + provenance.json (MNE version, filter cutoffs, ICA method + dropped-component indices + seed, rejection threshold, per-condition kept/rejected counts, input sha256, run date, model id). Cookbook’s first EEG-analysis recipe; sibling to the single-signal HRV recipe, and can run on public EEG discovered via the DANDI recipe. Reported — MNE-Python is the field-standard M/EEG toolbox (Gramfort et al., Front. Neurosci. 2013) and the filter→ICA→epoch→average workflow is reused across recent reproducible-pipeline packages built on it (EEG-Pype, PLoS Comput. Biol. 2026; osl-ephys, Front. Neurosci. 2025); the agent-driven assembly is not separately benchmarked. Fully open; Laptop.

Verified (no changes)

2026-07-11 (Molecular and Cellular Biology directed pass)

Added

  • Quantify bulk RNA-seq FASTQ into a gene-level counts matrix (Problem class: Data analysis; Evidence: Proposed) — rung-3 toolbelt: fastp trim → Salmon decoy-aware quasi-mapping (--gcBias --seqBias --validateMappings) → tximport gene-level aggregation via tx2gene → committed quantify_rnaseq.py + pinned requirements.txt + fastp JSON QC + counts.csv/coldata.csv + provenance.json (transcriptome release, index flags, FASTQ sha256, tool versions, run date, model id). Fills the FASTQ→counts gap the bulk RNA-seq DE recipe assumed away, now cross-linked as its downstream companion; alignment-based alternative noted via STAR+featureCounts. Proposed — grounded on Salmon bias/decoy-aware benchmarks (Patro et al., Nat. Methods 2017; Srivastava et al., Genome Biol. 2020) and tximport aggregation (Soneson et al., F1000Res 2015); the agent-orchestrated chain is not separately benchmarked. Fully open; Laptop.

Updated

  • Run bulk RNA-seq differential expression from a counts matrix — added upstream cross-link to the new FASTQ→counts recipe in See also.

Verified (no changes)

  • 2 recipes spot-checked (profile-chipseq-atacseq-signal-around-features, call-peaks-and-motifs-from-chipseq-atacseq), all current — linked deepTools/MACS3/HOMER catalog pages resolve and are unflagged, source repos and DOIs load; last_verified bumped to 2026-07-11.

2026-07-11 (Integrative Structural and Computational Biology directed pass)

Added

  • Predict a protein–protein complex to map the binding interface (Problem class: Hypothesis generation; Evidence: Reported) — rung-2 Boltz plugin recipe: two partner FASTA sequences → boltz-structure-and-binding multi-chain co-folding (wide sampling) → local 5 Å inter-chain contact recomputation → committed .claude/commands/predict-ppi-interface.md + interface_from_boltz.py + pinned env + interface.csv (consensus interface residues) + provenance.json (plugin version, boltz-api job ids/date, model id, input sha256). General non-antibody PPI counterpart to the antibody–antigen complex recipe; closes the long-deferred “AlphaFold-Multimer complex interface” candidate now that Boltz is catalogued. Reported — Boltz-2 matches AF3 on PDB 2024–2025 complexes (Passaro et al., bioRxiv 2025); AF-Multimer/AF3 are field-standard multimer baselines (Hou et al., Nat. Commun. 2025); the Claude-plugin assembly is not separately benchmarked. Subscription required (hosted Boltz API); Laptop.

Updated

  • Score point mutations for functional impact with a protein language model — last_verified bumped to 2026-07-11; ESM/gget catalog pages resolve and are unflagged, sources still current.
  • Predict gene-knockout phenotypes with flux balance analysis — last_verified bumped to 2026-07-11; COBRApy catalog page and Biomni system page resolve and are unflagged, sources still current.

Verified (no changes)

  • 2 additional recipes (ESM variant scoring, FBA knockout) spot-checked with the two above; linked catalog/system pages resolve.

2026-07-11 (Immunology and Microbiology directed pass)

Added

  • Scan a protein for candidate CD8 T-cell epitopes (Problem class: Experimental design; Evidence: Validated) — rung-2 MHC Binding Prediction skill recipe: antigen FASTA + HLA class I alleles → tile to 8–11-mers → MHCflurry (+optional NetMHCpan-4.1/MixMHCpred) presentation + %Rank scoring → committed scan_epitopes.py + pinned env + epitopes.csv + provenance.json (predictor versions/model release, allele list, input sha256, run date, model id). Cookbook’s first MHC-I epitope recipe. Validated — NetMHCpan/MHCflurry captured >half of major epitopes in the top 277 of 767,788 candidates in a proteome-wide benchmark (Paul et al., PLoS Comput. Biol. 2020); SOTA reconfirmed 2026 (Mecklenbräuker et al., Mol. Cell. Proteomics). Fully open; Laptop.
  • Reconstruct B-cell clonal lineages from AIRR-seq (Problem class: Data analysis; Evidence: Reported) — rung-2 Immcantation BCR Analysis skill recipe: AIRR rearrangement table → shazam data-derived clonal threshold → scoper clonal families → SHM/BASELINe selection → dowser germline-rooted lineage trees → committed bcr_lineages.R + pinned Immcantation env + clones.tsv/trees + provenance.json (package versions, IMGT germline release, derived threshold, input sha256, run date, model id). Cookbook’s first BCR clonal-analysis recipe. Reported — Immcantation is the documented AIRR-seq clonal-analysis standard with an active supporting methods literature (Abdollahi et al., BMC Bioinformatics 2023; Zhang et al., Front. Immunol. 2022); the agent-orchestrated assembly is not separately benchmarked. Fully open; Laptop.

Updated

  • Scan a therapeutic antibody for glycosylation sites — last_verified bumped to 2026-07-11; Glycoengineering/gget/Adaptyv catalog pages resolve and are unflagged, sources still current.
  • Infer cell-cell communication from single-cell RNA-seq — last_verified bumped to 2026-07-11; LIANA-MCP catalog page resolves and is unflagged, sources still current.

Flagged

  • None.

Verified (no changes)

  • 2 recipes spot-checked (antibody-glycosylation, cell-cell communication), all current.

2026-07-11 (Chemistry directed pass)

Added

  • Prepare the correct protonation state of a ligand before docking (Problem class: Experimental design; Evidence: Proposed) — rung-2 Rowan skill recipe: ligand SMILES → optional local Datamol standardization → Rowan submit_macropka_workflow (pH 0–14) → dominant microspecies at pH 7.4 + governing macro-pKa → committed prepare_protonation.py + pinned env + prepared_ligands.csv + provenance.json (Rowan skill/workflow ids, pH, run date, input/output sha256, model id). Fills the ligand-prep gap upstream of the DiffDock docking, affinity-ranking, and GROMACS MD recipes, distinguished from the pH-blind standardization in the enumerate-analogs recipe. Proposed — grounded on Rowan’s documented pKa/macro-pKa workflows and the established impact of protonation/tautomeric state on docking enrichment (Kim et al., J. Comput. Aided Mol. Des. 2013); the agent-orchestrated prep assembly is not separately benchmarked. Fully open (free Rowan tier; cloud submission — data-residency caveat); Laptop.

Updated

  • Analyze an existing MD trajectory for stability, flexibility, and contacts — last_verified bumped to 2026-07-11; linked MDAnalysis/MDTraj catalog pages resolve and are unflagged, sources still current.

Flagged

  • None.

Verified (no changes)

  • 1 recipe spot-checked (MD-trajectory analysis), current.

2026-07-05 (Drug Repurposing and Discovery directed pass)

Added

  • Validate a drug target with a GO/NO-GO score before committing bench work (Problem class: Knowledge synthesis; Evidence: Proposed) — rung-2 Drug Target Validation skill over the ToolUniverse MCP: a gene + disease → four-gate scoring (disease association 30 / druggability 25 / safety 20 / clinical precedent 15 / validation evidence 10) → committed .claude/commands/validate-target.md slash command + targets/validation_scores.csv + per-target cited cards + provenance.json (tooluniverse version, skill tag, Open Targets/ChEMBL release labels, resolved accessions, run date, model id). Fills the single-target GO/NO-GO gap between the prioritize-targets recipe (disease-in, ranked list) and the target-dossier recipe (gene-in, free-form). Proposed — grounded on the ToolUniverse ecosystem paper’s hypercholesterolemia case study (Gao et al., arXiv:2509.23426, 2025) and the field-standard genetic-evidence/precedence-tractability-doability-safety framework (Nelson et al., Nat. Genet. 2015; Ochoa et al., NAR 2023); the composite-score assembly is not separately benchmarked. Fully open; Laptop.

Verified (no changes)

  • Enumerate analogs around a lead compound for SAR expansion — linked catalog pages (datamol, rdkit-skill) resolve and are unflagged; textbook cheminformatics sources current. Bumped last_verified to 2026-07-05.
  • Identify an unknown compound from an MS/MS spectrum — linked catalog pages (matchms, pyopenms, pubchem) resolve and are unflagged; matchms/SimMS sources current. Bumped last_verified to 2026-07-05.
  • Estimate pharmacokinetic properties of a small molecule — linked catalog pages (rdkit-skill, medchem, chembl) resolve and are unflagged; ChEMBL/ChemCrow/PharmaBench sources current. Bumped last_verified to 2026-07-05.
  • Triage an AlphaFold model for structure-based drug design — linked catalog pages (alphafold, pdb, uniprot) resolve and are unflagged; AlphaFold DB / pLDDT-benchmark sources current. Bumped last_verified to 2026-07-05.

2026-07-05 (Translational Medicine directed pass)

Added

  • Assemble a public cancer imaging cohort from NCI Imaging Data Commons (Problem class: Knowledge synthesis; Evidence: Reported) — rung-2 Imaging Data Commons skill recipe: cohort criteria (modality / body part / accompanying RTSTRUCT-SEG / license) → idc-index DuckDB metadata query → license-filtered series-level cohort_manifest.csv + committed build_cohort.py + pinned env + provenance.json (idc-index version, IDC release version, query date, manifest sha256, patient/series counts, model id) → manifest-based DICOM download. The medical-imaging counterpart to the DANDI discovery recipe and cellxgene-census atlas recipe; the upstream cohort step feeding the nnU-Net segmentation recipe via the DICOM-to-BIDS recipe. Emphasizes the per-collection CC-BY/CC-NC license gate. Reported — IDC and its idc-index client are the documented public-cancer-imaging cohort tooling (Fedorov et al., RadioGraphics 2023; Fedorov et al., Cancer Res. 2021); the Claude-skill assembly is not separately benchmarked. Fully open; Laptop.
  • Extract structured variables from free-text clinical notes (Problem class: Data analysis; Evidence: Validated) — rung-1 (Claude Code alone) recipe: a folder of de-identified clinical notes + a versioned codebook.md → per-cell extraction with evidence_quote and found/negated/not_mentioned status into records.jsonl → flattened registry.csv + committed extract_registry.py + pinned env + provenance.json (model id, codebook sha256, note count, gold-subset accuracy) + a clinician-abstracted gold-set accuracy check. Cookbook’s first note-extraction recipe; the upstream step to the harmonize-clinical-terms recipe (extract → then map terms to codes) and cross-linked to the readmission-prediction recipe. Validated — Claude 3.5 Sonnet extracted structured binary variables from ILD clinic notes at 96.2% accuracy, matching three-physician consensus (Chen et al., J. Med. Internet Res. 2026); oncologic-history extraction reached F1 = 0.983 (Bhayana et al., Radiology 2025). Fully open; Laptop.

Verified (no changes)

  • Match a patient summary to recruiting clinical trials — linked catalog pages (biomcp, clinicaltrials-gov-mcp, clinical-trial-protocol) and the Biomni system page resolve and are unflagged; TrialGPT / MatchMiner-AI sources current. Bumped last_verified to 2026-07-05.
  • Build a pharmacogenomic dosing report from a patient’s diplotypes — linked catalog pages (clinpgx-database, ddinter-database) resolve and are unflagged; CPIC guideline sources current. Bumped last_verified to 2026-07-05.

2026-07-04 (Molecular and Cellular Biology directed pass, later slot)

Added

  • Find selective genetic dependencies for a cancer context with DepMap (Problem class: Hypothesis generation; Evidence: Reported) — rung-2 DepMap skill recipe: a cancer context (lineage / driver mutation / fusion / expression state) → copy-number-corrected group contrast over the Chronos CRISPRGeneEffect matrix → pan-essential-filtered, FDR-corrected dependencies.csv (selective effect, p, fdr, copy_number_flag) + committed find_dependencies.py + pinned env + provenance.json (DepMap release label + matrix sha256s, context/background definition, thresholds, run date, model id). First recipe to use the catalogued DepMap skill as a standalone tool (context-in / genes-out); distinguished from the target-dossier recipe (gene-in, uses DepMap as one of four lookups) and the prioritize-targets recipe (disease-in, Open Targets associations). Cross-linked to the sgRNA-design recipe for validation. Reported — DepMap-mining for selective dependencies is routine, with recent peer-reviewed instances (Schneider et al., Cancer Res. 2024; Phillips et al., Nat. Commun. 2025; Iyer et al., EMBO J. 2025); the Claude-skill assembly is not separately benchmarked. Fully open; Laptop.

2026-07-04 (Integrative Structural and Computational Biology directed pass)

Added

  • Superpose two protein structures and quantify where they differ (Problem class: Data analysis; Evidence: Proposed) — rung-2 PyMOL skill recipe: two coordinate files (AlphaFold model vs experimental PDB, or apo vs holo) → cealign/super superposition → global RMSD + per-residue Cα-deviation perres.csv + deviation-coloured overlay PNG + .pse session + committed .claude/commands/superpose.md + uv-pinned env + provenance.json (PyMOL version, method + cutoff, both input sha256s, chain selectors, aligned-atom count, RMSD, fetch date/accession, model id). First recipe to use the catalogued PyMOL skill; cross-linked to the Foldseek function recipe (database-search counterpart) and the AlphaFold triage recipe (confidence-side counterpart). Proposed — no LLM-driven superposition workflow is documented; grounded on peer-reviewed PyMOL cealign (Shindyalov & Bourne, Protein Eng. 1998) and AlphaFold model-vs-experiment RMSD validation (Jumper et al., Nature 2021). Fully open; Laptop.

Verified (no changes)

  • Sort spikes from a Neuropixels recording end-to-end — linked catalog pages (neuropixels-analysis, neurosift, aind-data) resolve and are unflagged; SpikeAgent/SpikeInterface sources current. Corrected the K-Dense SKILL.md source URL path (scientific-skills/ → skills/) to match the catalog page. Bumped last_verified to 2026-07-04.
  • Organize a raw DICOM dataset into a BIDS layout — linked catalog pages (bids, openneuro) resolve and are unflagged; BIDS-spec and component DOIs current. Bumped last_verified to 2026-07-04.

2026-07-04 (Immunology and Microbiology directed pass, later slot)

Added

  • Identify a bacterial isolate from its 16S rRNA sequence (Problem class: Data analysis; Evidence: Reported) — rung-2 BLAST (Bio-MCP) recipe: a colony-PCR Sanger 16S read (or assembled 16S contig) → blastn against a pinned curated 16S database (NCBI 16S_ribosomal_RNA) → coverage-filtered, identity-ranked hits.csv with a threshold-based assignment column (98.7% species / 94.5% genus) + committed identify_16s.py + pinned env + provenance.json (BLAST+ version, 16S DB release + sha256, query sha256, cutoffs, run date, model id). Single-isolate identification counterpart to the community-level 16S diversity recipe; cross-linked to the resistome screen (same BLAST-MCP tool), the bacterial-genome-annotation recipe (consumes the genus hint), and the phylogenetic-tree recipe. Reported — 98.7%/94.5% identity thresholds are the community standard (Kim et al., IJSEM 2014; Yarza et al., Nat. Rev. Microbiol. 2014); the BLAST-MCP assembly is not separately benchmarked. Fully open; Laptop.

Verified (no changes)

  • Build a phylogenetic tree from a set of sequences — linked catalog pages (phylogenetics, ETE Toolkit, Nextflow) and Biomni system page all resolve; sources current. Bumped last_verified to 2026-07-04.

2026-07-04 (Immunology and Microbiology directed pass)

Added

  • Screen a bacterial genome for resistance and virulence genes (Problem class: Data analysis; Evidence: Reported) — rung-2 BLAST (Bio-MCP) recipe: an annotated genome’s protein FASTA → makeblastdb + blastp against pinned CARD and VFDB references → identity/coverage-filtered best-hit-per-database → resistance_hits.csv + virulence_hits.csv + committed screen_resistome.py + pinned env + provenance.json (BLAST+ version, CARD release, VFDB download date, FASTA sha256s, cutoffs, input sha256). Picks up the AMR/virulence step the bacterial-genome-annotation recipe explicitly punts to the CLI, cross-linked to it and the pan-genome recipe. Reported — blast+ AMR/virulence detection validated at >95% sensitivity/specificity on a 131-isolate reference collection (Bogaerts et al., Microb. Genom. 2021); the BLAST-MCP assembly is not separately benchmarked. Fully open; Laptop.

Flagged

  • Missing component: dedicated resistome caller (RGI / AMRFinderPlus) — surfaced for the catalog curator; the new AMR recipe is homology-only (no point-mutation resistance models) until such a tool is Claude-installable.

Verified (no changes)

2026-06-28 (Drug Repurposing and Discovery directed pass)

Added

  • Rank a compound library against a target by predicted binding affinity (Problem class: Data analysis; Evidence: Reported) — rung-2 Boltz plugin recipe driving the hosted boltz-small-molecule-screen skill: a target sequence/PDB + a SMILES library → MedChem/Datamol pre-filter (upstream recipe) → structure-and-affinity screen on the hosted Boltz API (no local GPU) → screen_ranked.csv (affinity, binder probability, structure confidence) + committed rank_screen.py + .claude/commands/boltz-affinity-screen.md + pinned env + provenance.json (Boltz model id, job IDs, submission date, input sha256, target accession), ranking by the classifier score rather than fine affinity gaps. Cookbook’s first GPU-free structure-based affinity-screening recipe, cross-linked to the DiffDock recipe (GPU pose-level counterpart) and the virtual-screening hit filter (upstream). Reported — Boltz-2 approaches FEP accuracy (Pearson 0.62 on FEP+, doubles MF-PCBA average precision; Passaro et al., bioRxiv 2025) but an independent eval (Wan et al., arXiv:2603.05532, 2026) finds it a good binder classifier yet weak quantitative ranker; the Claude-plugin assembly is not separately benchmarked. Subscription required; Laptop.

Updated

  • Scan approved drugs for repurposing candidates against a disease — spot-checked; all catalog/system links resolve, tags consistent; last_verified → 2026-06-28.
  • Profile a compound’s polypharmacology from ChEMBL bioactivity data — spot-checked; all catalog links resolve, tags consistent; last_verified → 2026-06-28.

Verified (no changes)

  • No recipes are over the 30-day last_verified window (oldest is 2026-06-03); spot-checked the oldest Drug Repurposing recipes, all current.

2026-06-28 (Translational Medicine directed pass)

Added

  • Register longitudinal medical scans to a common frame (Problem class: Data analysis; Evidence: Reported) — rung-2 SimpleITK skill recipe: a baseline + follow-up CT/MRI pair → CenteredTransformInitializer → two-stage rigid (Euler3D, Mattes MI, multi-resolution pyramid) → deformable (B-spline) registration → resampled warped.nii.gz + persisted transform.tfm + propagated label mask (nearest-neighbour, Dice reported) + committed register_scans.py + pinned env + provenance.json (SimpleITK version, transform/metric/optimizer settings, control-point spacing, input sha256s), with a checkerboard/difference-overlay QC step. Cookbook’s first image-registration recipe, cross-linked to the nnU-Net segmentation recipe (produces the masks it propagates), the DICOM-to-BIDS recipe (upstream conversion), and the survival recipe (downstream). Reported — SimpleITK/ITK are peer-reviewed and field-standard (Yaniv et al., J. Digit. Imaging 2018); the Claude-skill assembly is not independently benchmarked. Fully open; Workstation with GPU.
  • Tile and stain-normalize a whole-slide image for ML (Problem class: Data analysis; Evidence: Reported) — rung-2 histolab skill recipe: a gigapixel H&E WSI (or folder) → tissue masking → fixed-size tile extraction at a chosen magnification → Macenko/Reinhard stain normalization against a committed reference → a tiles/ folder + manifest.csv (slide_id, tile_path, level, mpp, x, y, tissue_pct) + committed tile_wsi.py + pinned env + provenance.json (histolab/OpenSlide versions, tile size, level, tissue threshold, normalizer + reference sha256, slide sha256), with a contact-sheet QC step. Cookbook’s first digital-pathology WSI-preprocessing recipe, cross-linked to the pathml skill (heavier toolkit) and the microscopy segmentation recipe (cell-level counterpart). Reported — histolab is the peer-reviewed reproducible-preprocessing library (Marcolini et al., SoftwareX 2022); the Claude-skill assembly is not independently benchmarked. Fully open; Workstation with GPU.
  • Harmonize free-text clinical terms to standard codes (Problem class: Knowledge synthesis; Evidence: Proposed) — rung-2 Medical Terminologies MCP recipe: a column of free-text diagnoses/drugs/labs → grounded lookup against ICD-11/RxNorm/ATC/LOINC → ranked candidates.csv (all hits per term) → a curated crosswalk.csv (term, chosen code/system/concept, match_type, n_candidates) + committed harmonize_terms.py + pinned env + provenance.json (MCP version, per-service terminology release dates, run date, model id, input sha256), with an expert-review pass on needs_review rows. Cookbook’s first clinical-terminology harmonization recipe, cross-linked to the PyHealth skill (in-pipeline code utilities) and the pharmacogenomic dosing report (closest documented LLM-driving-clinical-references workflow). Proposed — grounding rests on authoritative WHO/RxNorm/LOINC services; no documented attempt of this exact LLM+MCP assembly is known. Fully open; Laptop.

Verified (no changes)

  • No recipes are over the 30-day last_verified window (oldest is 2026-06-03); spot-checked the oldest, all current.

2026-06-27 (Molecular and Cellular Biology directed pass)

Added

  • Segment and quantify cells in a microscopy image (Problem class: Data analysis; Evidence: Reported) — rung-2 Cellpose skill recipe: a fluorescence/brightfield image (or folder) → pretrained generalist instance segmentation (cyto3/nuclei) → per-cell label masks + a tidy cells.csv (count, area, eccentricity, centroid, per-channel mean intensity via scikit-image regionprops), with a committed segment_and_quantify.py + pinned env + provenance.json (cellpose version, model, diameter, flow/cellprob thresholds, input sha256) and an explicit mask-QC step (over-merge/over-split). Cookbook’s first cell-microscopy segmentation/quantification recipe, cross-linked to the scikit-image skill (classical fallback) and the nnU-Net medical-image recipe (radiology counterpart). Reported — Cellpose is field-defining (Stringer et al., Nature Methods 2021; Cellpose3, 2025); the Claude-skill assembly is not independently benchmarked. Fully open; Laptop.
  • Design CRISPR sgRNAs for a gene knockout (Problem class: Experimental design; Evidence: Reported) — rung-2 sgRNA Design skill recipe: a gene + genome build → tiered guide selection (validated Addgene libraries → CRISPick pre-computed picks → de novo Rule Set 2 / Azimuth on-target + CFD off-target scoring) → a committed guides.csv (sequence, PAM, exon, strand, on/off-target, source tier) + pinned env + provenance.json (genome build, skill + source DB versions), with a pre-order exon/PAM/off-target sanity check. Cookbook’s first CRISPR guide-design recipe, cross-linked to the plasmid-verification recipe. Reported — scoring rests on the field-standard Rule Set 2 / CFD models (Doench et al., Nature Biotechnology 2016); the Claude-skill assembly is not independently benchmarked. Fully open; Laptop.

2026-06-27 (Integrative Structural and Computational Biology directed pass)

Added

  • Infer the function of an uncharacterized protein from its 3D structure (Problem class: Knowledge synthesis; Evidence: Reported) — rung-2 Foldseek Structural Search skill recipe: a .pdb/.cif coordinate file → hosted Foldseek search against annotated structure DBs (afdb-swissprot, pdb100, afdb50) → ranked hit table + a banded (STRONG/SUGGESTIVE/NO-CONFIDENT-HIT) function call read off the top hits’ annotations, with a committed .claude/commands/foldseek-function.md, uv-pinned skill env, and provenance.json (Foldseek API date + DB version strings + input sha256). The structure-side fallback for when sequence search (BLAST/HMMER/InterPro) returns nothing; cookbook’s first structural-homology / function-inference recipe, cross-linked to the AlphaFold-triage recipe (structure-quality counterpart) and the ESM variant-scoring recipe. Reported — Foldseek is peer-reviewed and field-defining (van Kempen et al., Nature Biotechnology 2023: 4–5 orders of magnitude faster than DALI/TM-align at 86–133% of their sensitivity), and the skill is a maintained Google DeepMind release; the Claude-skill assembly is not independently benchmarked. Fully open; Laptop.

Verified (no changes)

  • All recipes within the 30-day last_verified window (oldest 2026-06-03); no aging verification due this run.

2026-06-27 (Immunology and Microbiology directed pass)

Added

  • Annotate and verify an engineered plasmid construct (Problem class: Experimental design; Evidence: Reported) — rung-2 pLannotate skill recipe taking a plasmid FASTA/GenBank through a local BLAST run against curated genetic-parts databases (Addgene, fpbase, Swiss-Prot, Rfam) → an annotated GenBank + feature table + HTML map, with the fragment column surfaced as the construct-verification signal and a committed annotate_plasmid.py + pinned env + provenance.json (pLannotate DB version + input sha256). Immunology and Microbiology / Molecular and Cellular Biology focus-day recipe; cookbook’s first engineered-plasmid recipe, cross-linked to the bacterial-genome annotation recipe (chromosomal counterpart) and explicitly contrasted with Prokka/Bakta, which don’t recognize engineered parts. Reported — pLannotate is the established engineered-plasmid annotator (McGuffie & Barrick, Nucleic Acids Res. 2021: annotates recombinant/synthetic/engineered elements and reports incomplete fragments that genome pipelines miss); the Claude-skill assembly is not independently benchmarked. Fully open; Laptop.

2026-06-27

Added

  • Predict an antibody–antigen complex to map an epitope (Problem class: Experimental design; Evidence: Reported) — rung-2 Boltz plugin recipe taking antibody/nanobody chains + an antigen sequence through a hosted-API boltz-structure-and-binding co-fold with a deliberately wide sample ensemble, then a local 4.5 Å CDR-contact recomputation per model (committed epitope_from_boltz.py + pinned env + provenance.json capturing boltz-api job IDs and submission date) → a consensus epitope across the top models rather than trusting top-1 confidence. Immunology and Microbiology focus-day recipe; cookbook’s first co-folding/complex-prediction recipe, cross-linked to the AlphaFold-triage recipe (single-chain counterpart) and the antibody-glycosylation recipe. Reported — Boltz-2/1x are independently benchmarked on Ab/VHH–antigen complexes (Gupta et al., Protein Science 2026, SNAC-DB: success rates ≤25%, 1000-sample oracle 50.5% vs near-flat confidence ranking; Ünsal et al., Brief. Bioinform. 2026, AntiConf), but the Claude-plugin assembly is not independently benchmarked. Subscription required (hosted Boltz API); Laptop.

Verified (no changes)

  • 4 recipes spot-checked, all current (last_verified bumped to 2026-06-27): set-up-protein-md-simulation-in-gromacs, convert-instrument-data-to-allotrope-asm, compute-16s-microbiome-diversity, parse-fcs-flow-cytometry-files. All linked catalog/system pages resolve and are unflagged; source citations stable.

2026-06-21

Added

  • Score a drug-combination screen for synergy (Problem class: Data analysis; Evidence: Proposed) — rung-2 ToolUniverse Drug Synergy skill recipe taking user-supplied single-agent + combination effect data (on one consistent scale) through model selection by data shape (DrugSynergy_calculate_bliss/_hsa/_loewe/_zip/_ci) → synergy score → synergy/additive/antagonism classification via the standard ±10 thresholds (CI < 1 inverse), with scale-mixing and dose-dependence footguns surfaced. Drug Repurposing and Discovery focus-day recipe; cookbook’s first combination-synergy recipe, cross-linked to the polypharmacology recipe and the drug-repurposing scan recipe. Proposed — no documented Claude-driven ToolUniverse synergy assembly; grounded in the field-standard reference models from Ianevski et al., Nucleic Acids Research 2022 (SynergyFinder 3.0) and the skill’s SKILL.md. Laptop.
  • Detect somatic copy-number variants from tumor sequencing (Problem class: Data analysis; Evidence: Reported) — rung-2 CNVkit skill recipe taking tumor WES/targeted-panel BAMs through pooled-reference construction → coverage binning + bias correction → CBS segmentation → gene-level amplification/deletion calls (with stated log2 thresholds) → scatter/diagram QC plots and SEG/VCF export, with purity/ploidy and matched-normal caveats surfaced. Translational Medicine focus-day recipe; cookbook’s first somatic-CNV recipe, paired with the cBioPortal cohort recipe for cohort-level context. Reported — CNVkit is the field-standard engine for copy-number from targeted/exome data (Talevich et al., PLOS Comput. Biol. 2016); the Claude-skill assembly is not independently benchmarked.
  • Predict hospital readmission from EHR data (Problem class: Data analysis; Evidence: Proposed) — rung-2 PyHealth skill recipe taking a credentialed EHR extract (MIMIC-IV/eICU/OMOP) through the dataset → 30-day-readmission task → RETAIN/Transformer sequence model → patient-level-split AUROC/AUPRC + calibration vs a logistic-regression baseline, with data-use-agreement and cross-institution-transfer caveats surfaced. Translational Medicine focus-day recipe; cookbook’s first EHR clinical-prediction recipe, complementing the survival-model recipe. Proposed — no documented LLM-driven PyHealth assembly; grounded in Yang et al., KDD 2023 and PyHealth 2.0, arXiv:2601.16414 (2026). Institutional access (PhysioNet credentialed datasets).
  • Segment an organ or tumor in a medical image with nnU-Net (Problem class: Data analysis; Evidence: Reported) — rung-2 nnU-Net skill recipe taking labeled CT/MRI volumes through dataset-fingerprint planning → auto-configured preprocessing/architecture → 5-fold cross-validated training → best-config selection → held-out mask prediction + volume QC, with the nnU-Net folder/_0000 data contract and the GPU/multi-day-per-fold cost surfaced. Translational Medicine focus-day recipe; cookbook’s first medical-image-segmentation recipe, chained off the DICOM-to-BIDS recipe upstream and feeding the survival-model recipe downstream. Reported — nnU-Net is field-defining (Isensee et al., Nature Methods 2021: match/beat specialized solutions on 23 challenges, no manual tuning); the Claude-skill assembly is not independently benchmarked. Workstation with GPU.

Verified (no changes)

  • 4 Translational Medicine recipes spot-checked and refreshed to 2026-06-21 (last_verified bumped): interpret-clinical-variant, match-patient-to-clinical-trials, scan-adverse-events-for-drug-safety-signal, fit-survival-model-to-clinical-outcomes — all linked catalog pages resolve and source URLs current.
  • benchmark-admet-property-with-pytdc spot-checked and refreshed to 2026-06-21 — pytdc/molfeat/datamol catalog pages resolve, TDC ADMET_Group leaderboard URL loads (22 datasets incl. caco2_wang).

2026-06-20

Added

  • Annotate cell types in a single-cell dataset (Problem class: Data analysis; Evidence: Reported) — rung-2 CellTypist skill recipe taking a QC’d/clustered AnnData through reference-model logistic-regression annotation with majority_voting over clusters → per-cell + per-cluster labels + confidence → a canonical-marker sanity check, with an optional rung-3 escalation to the popV consensus skill (8 classifiers + agreement score) when single-method confidence is poor. Molecular and Cellular Biology focus-day recipe; cookbook’s first cell-type-annotation recipe, chained off the scRNA-seq QC recipe. Reported — CellTypist (Domínguez Conde et al., Science 2022) and popV (Ergen et al., Nat. Genet. 2024) are peer-reviewed; the Claude-skill assembly is not independently benchmarked.
  • Predict the regulatory effect of a non-coding variant (Problem class: Knowledge synthesis; Evidence: Reported) — rung-2 AlphaGenome skill recipe taking a single chr:pos:ref>alt variant through tissue-ontology resolution → sequence-to-function scoring across expression / accessibility / histone marks / splicing / TF binding → ranked modalities + ISM motif logo + splicing-disruption analysis, with tissue choice surfaced as the key judgment call. Molecular and Cellular Biology focus-day recipe; cookbook’s first non-coding/regulatory variant recipe, cross-linked to the coding-variant clinical-variant recipe. Reported — AlphaGenome is peer-reviewed (Nature 2026: matches/exceeds best external models on 25/26 variant-effect evaluations); Subscription required (signup-gated free research-preview API key).
  • Predict RNA secondary structure and target-site accessibility (Problem class: Data analysis; Evidence: Proposed) — rung-2 ViennaRNA skill recipe taking an RNA sequence through MFE folding + partition function → centroid/ensemble metrics → RNAplfold target-window accessibility (mean/min unpaired probability) → optional RNAduplex guide-strand check → a ranked, parameter-pinned design card for siRNA/sgRNA/ASO design or riboswitch analysis. Integrative Structural and Computational Biology focus-day recipe; cookbook’s first RNA-secondary-structure recipe. Proposed — no documented LLM-driven ViennaRNA assembly; grounded in Lorenz et al., Algorithms Mol. Biol. 2011 and the established role of RNAplfold accessibility in reagent-efficacy prediction.
  • Infer transcription-factor and pathway activities from expression (Problem class: Data analysis; Evidence: Proposed) — rung-2 decoupler-MCP recipe taking an annotated AnnData through tf_activity (CollecTRI/ULM) and pathway_activity (PROGENy/MLM) footprint inference → between-condition ranked activity tables → a grounded summary with a positive-control check. Immunology and Microbiology focus-day recipe; cookbook’s first footprint/activity-inference recipe, kept distinct from over-representation enrichment and de-novo GRN inference. Proposed — no documented LLM-driven decoupler-MCP assembly; grounded in Badia-i-Mompel et al., Bioinform. Adv. 2022, Schubert et al., Nat. Commun. 2018 (PROGENy), and Müller-Dott et al., NAR 2023 (CollecTRI).
  • Map a disease to its implicated genes and pathways (Problem class: Knowledge synthesis; Evidence: Reported) — rung-3 chain of two existing Reported recipes: Open Targets target ranking (overall association score) → gget/Enrichr functional enrichment, with DisGeNET as a positive control and a grounded synthesis. Canonicalized from composition report #43 (knee OA, EFO_0004616). Reported — #43 documents the chain running end-to-end on a laptop in under a minute with the disease recovered as a DisGeNET positive control.
  • Annotate a single bacterial genome assembly (Problem class: Data analysis; Evidence: Reported) — rung-2 Bakta skill recipe taking one assembled bacterial/archaeal genome through database-pinned annotation (CDS, rRNA/tRNA/ncRNA, CRISPR arrays, replicon features) → GFF3/GenBank/protein-FASTA → a feature-count sanity check, with the replicon-completeness and database-version footguns surfaced and an optional CARD/VFDB AMR overlay noted. Immunology and Microbiology focus-day recipe; the single-isolate counterpart to the multi-genome pan-genome recipe. Reported — single-isolate Bakta annotation is the field-standard opening move (Santhosh et al., BMC Genomics 2025; Schwengers et al., Microb. Genom. 2021).

Updated

  • Run functional enrichment on a gene list — fixed the gget install block (per #41): removed the non-existent K-Dense-AI/claude-scientific-skills marketplace, replaced with the catalog’s npx skills add + manual-HTTPS-clone paths; added a Field-reports note.
  • Prioritize targets within a disease via Open Targets — added a known-issue note (per #43) that the hosted Open Targets MCP endpoint is failing its initialize handshake (JSON-RPC -32602); documented the direct GraphQL API and ToolUniverse OpenTargets_* tools as the working path, plus a Field-reports entry.

Verified (no changes)

  • 6 aging recipes spot-checked (linked catalog tools resolve and unflagged, canonical sources resolve), last_verified bumped to 2026-06-20: Infer a gene-regulatory network from single-cell RNA-seq, Run first-pass QC on a single-cell RNA-seq dataset, Triage a stack of new preprints, Run bulk RNA-seq differential expression, Compute HRV from an ECG recording, Discover NWB recordings on DANDI.

2026-06-14

Added

  • Screen a polypharmacy medication list for drug-drug interactions (Problem class: Knowledge synthesis; Evidence: Reported) — rung-2 DDInter skill recipe taking a medication list through per-drug ID resolution → pairwise DDInter queries → a cited severity/mechanism/management table with explicit “clean” lines, plus an optional rung-3 DailyMed + ClinPGx overlay on the major pairs. Drug Repurposing and Discovery focus-day recipe; cookbook’s first DDI-screening recipe. Reported — Domián et al., Explor. Res. Clin. Soc. Pharm. 2025 documents that ungrounded LLMs over-flag/hallucinate DDIs (Copilot 1,813 vs a 204-interaction reference on 57 real patients), establishing that screening must be anchored to a curated DDI database — the assembly this recipe recommends.
  • Run a GWAS on case-control genotype data (Problem class: Data analysis; Evidence: Proposed) — rung-2 PLINK2 skill recipe taking a PLINK/VCF genotype set through sample + variant QC (call rate, MAF, HWE-in-controls) → LD pruning → genotype PCA → PCA-adjusted logistic-regression --glm association with a lambda_GC inflation check, handing genome-wide-significant loci to the GWAS Catalog skill for annotation. Translational Medicine focus-day recipe; cookbook’s first GWAS recipe. Proposed — no documented LLM-driven PLINK2 assembly; grounded in Chang et al., GigaScience 4:7 (2015) and the canonical QC tutorial (Marees et al., Int. J. Methods Psychiatr. Res. 27:e1608 (2018)).
  • Build a pharmacogenomic dosing report from a patient’s diplotypes (Problem class: Knowledge synthesis; Evidence: Proposed) — rung-2 ClinPGx skill recipe taking star-allele diplotypes plus a medication list through diplotype→metabolizer-phenotype translation (CPIC PostgREST API) → per-drug CPIC/DPWG dosing recommendation lookup → a cited drug gene phenotype recommendation table, with explicit “no actionable guidance” flagging and a DDInter phenoconversion overlay noted. Translational Medicine focus-day recipe; cookbook’s first pharmacogenomic-dosing recipe, distinct from the germline-pathogenicity variant-interpretation recipe. Proposed — no documented LLM-driven ClinPGx/CPIC assembly; grounded in the CPIC guideline corpus (Amstutz et al., Clin. Pharmacol. Ther. 2018; Molden & Jukić, Front. Pharmacol. 2021).
  • Profile a cancer cohort’s genomics with cBioPortal (Problem class: Knowledge synthesis; Evidence: Reported) — rung-2 cBioPortal skill recipe taking a study + gene set through study/profile lookup → per-gene mutation+CNA alteration frequency and co-occurrence/mutual-exclusivity → TMB summary → a Kaplan-Meier overall-survival split by mutation status, with cohort-denominator caveats enforced. Translational Medicine focus-day recipe; cookbook’s first cohort-level cancer-genomics recipe, cross-linked to the gene-centric target dossier, single-variant variant-interpretation, and adjusted-modelling survival recipe. Reported — the cBioPortal-backed AI-HOPE conversational-agent family documents the assembly class (AI-HOPE-WNT, Front. Artif. Intell. 2025, recapitulating WNT-EOCRC survival p=0.0167/0.0007; AI-HOPE-TP53, Cancers 2025).

Verified (no changes)

  • Build a target dossier and Draft a Phase 2/3 clinical-trial protocol — linked catalog tools and key sources re-checked, last_verified bumped to 2026-06-14.
  • Assemble a tissue reference atlas from the CELLxGENE Census — linked catalog tools (cellxgene-census, scvi-tools, scanpy, anndata) and Census/scvi-hub sources re-checked, last_verified bumped to 2026-06-14.

2026-06-13

Added

  • Infer cell-cell communication from single-cell RNA-seq (Problem class: Data analysis; Evidence: Proposed) — rung-2 LIANA-MCP recipe taking an annotated AnnData object through ls_ccc_method → multi-method communicate (CellPhoneDB/Connectome/NATMI/SingleCellSignalR) → rank_aggregate consensus ligand-receptor tetrads → circle_plot/ccc_dotplot, consuming the annotated object from the scRNA-seq QC recipe. Molecular and Cellular Biology focus-day recipe; cookbook’s first cell-cell-communication recipe. Proposed — no documented LLM-driven LIANA-MCP assembly; grounded in Dimitrov et al., Nat. Commun. 13:3735 (2022), a 2026 consensus-LIANA application (Wei et al., PLOS ONE 2026), and the method-disagreement benchmark (Xie et al., Biomolecules 13:1211 (2023)).
  • Call peaks and find enriched motifs from ChIP-seq or ATAC-seq (Problem class: Data analysis; Evidence: Proposed) — rung-3 toolbelt chaining the MACS3 skill (callpeak, narrow/broad mode → narrowPeak BED) into the HOMER skill (annotatePeaks.pl nearest-gene context + findMotifsGenome.pl de-novo/known motif enrichment). Molecular and Cellular Biology focus-day recipe; the binding-site/motif companion to the deepTools signal-profiling recipe, which deliberately stops before peak calling. Proposed — no documented LLM-driven MACS3→HOMER assembly; grounded in the field-standard pipeline (Zhang et al., Genome Biol. 9:R137 (2008); Heinz et al., Mol. Cell 38:576 (2010)).
  • Analyze an existing MD trajectory for stability, flexibility, and contacts (Problem class: Data analysis; Evidence: Proposed) — rung-2 MDAnalysis skill recipe taking a finished GROMACS/AMBER/NAMD trajectory through a load-and-sanity-check → aligned RMSD/RMSF/Rg → interface contact map + H-bond occupancy → backbone PCA battery, with the MDTraj skill as the DSSP/Ramachandran fallback. Integrative Structural and Computational Biology focus-day recipe; the post-simulation-analysis companion to the GROMACS setup recipe. Proposed — no documented LLM-driven MDAnalysis-skill assembly; grounded in Michaud-Agrawal et al., J. Comput. Chem. 32:2319 (2011), McGibbon et al., Biophys. J. 109:1528 (2015), and class-level agentic-MD evidence (MDCrow, Mach. Learn. Sci. Technol. 2025).
  • Scan a therapeutic antibody for glycosylation sites (Problem class: Experimental design; Evidence: Proposed) — rung-2 Glycoengineering skill recipe taking heavy/light-chain sequences through N-X-S/T sequon detection (flagging Fc Asn-297 vs unintended variable-domain sites) → O-glycosylation hotspot prediction → a parent-vs-variant sequon diff, with optional minimal site-knockout edit suggestions. Immunology and Microbiology focus-day recipe; cookbook’s first antibody-developability / glycosylation recipe. Proposed — no documented LLM-driven glycoengineering-skill assembly; grounded in 2026 Fc-glycan/ADCC literature (Shuang et al., mAbs 2026; Illés 2026) and the galactosylation-as-CQA reference (Klingler et al., Biotechnol. Bioeng. 2024).
  • Compute a bacterial pan-genome from a set of genome assemblies (Problem class: Data analysis; Evidence: Proposed) — rung-3 toolbelt chaining the Bakta skill (identical per-genome annotation → GFF3) into the Roary skill (CD-HIT/BLAST/MCL clustering → core/soft-core/shell/cloud partition, gene_presence_absence.csv, and a core_gene_alignment.aln that feeds the phylogenetics recipe). Immunology and Microbiology focus-day recipe; cookbook’s first comparative-genomics / pan-genome recipe. Proposed — no documented LLM-driven Bakta→Roary assembly; grounded in the field-standard pipeline (Page et al., Bioinformatics 2015; Schwengers et al., Microb. Genom. 2021) and a 2025 27,884-genome application (Sholeh et al., Mol. Genet. Genomics 2025).

Verified (no changes)

  • 35 recipes spot-checked; all last_verified dates within the 30-day window, no aging recipes due.

2026-06-11

Added

  • Profile ChIP-seq or ATAC-seq signal around genomic features (Problem class: Data analysis; Evidence: Proposed) — rung-2 deepTools skill recipe taking aligned ChIP-seq/ATAC-seq BAMs through bamCoverage BPM-normalized bigWig generation → multiBamSummary + plotCorrelation replicate QC → computeMatrix + plotHeatmap/plotProfile TSS/peak-centered visualization, with upstream BAM handling via the pysam skill. Molecular and Cellular Biology focus-day recipe; cookbook’s first ChIP-seq/ATAC-seq coverage-profiling recipe. Proposed — no documented LLM-driven deepTools workflow; grounded in Ramírez et al., NAR 44:W160 (2016) plus class-level Biomni.
  • Predict gene-knockout phenotypes with flux balance analysis (Problem class: Data analysis; Evidence: Proposed) — rung-2 COBRApy skill recipe taking a genome-scale SBML model through baseline FBA sanity-check → genome-wide single_gene_deletion essentiality ranking → focused double_gene_deletion synthetic-lethality screen, with an explicit growth-ratio essentiality threshold. Molecular and Cellular Biology focus-day recipe; cookbook’s first constraint-based metabolic-modelling recipe. Proposed — no documented LLM-driven COBRApy workflow; grounded in Ebrahim et al., BMC Syst. Biol. 7:74 (2013) and Orth et al., Nat. Biotechnol. 28:245 (2010), plus class-level Biomni.

Verified (no changes)

  • 33 recipes spot-checked; all last_verified dates within the 30-day window, no aging recipes due.

2026-06-10

Added

  • Score point mutations for functional impact with a protein language model (Problem class: Data analysis; Evidence: Proposed) — rung-2 ESM skill recipe taking a wild-type protein sequence (optionally fetched by UniProt accession via the gget skill) and a list of substitutions through masked-marginal log-likelihood-ratio scoring → a ranked tolerated/deleterious CSV, with a wt-marginal one-pass variant for full single-mutation landscapes. Integrative Structural and Computational Biology focus-day recipe; cookbook’s first zero-shot variant-effect / protein-fitness recipe and the database-free complement to the clinical-variant interpretation recipe. Proposed — no documented LLM-driven ESM-skill scoring assembly; grounded in the canonical zero-shot method Meier et al., NeurIPS 2021, the ProteinGym benchmark, and 2025 directed-evolution use Zhang et al., Nat. Commun. 2025.

Verified (no changes)

  • 31 recipes spot-checked; all last_verified dates within the 30-day window, no aging recipes due.

2026-06-09

Added

Updated

  • Estimate pharmacokinetic properties of a small molecule — promoted Proposed → Reported on the first field report (issue #12). A user ran the full three-layer assembly through to a finished PK card and captured it in a standalone pk_card.py, verified across caffeine, ibuprofen, quercetin, and terfenadine. Added a Field reports subsection under Evidence and refreshed last_verified to 2026-06-09.

Verified (no changes)

User requests

  • #12 @goodb — resolved. This entry had been stuck open since 2026-05-27 because the responder emitted no machine-readable trailer, so the request content lived only in the GitHub issue body — which the sandboxed curator agent (no gh/shell) could not read, leaving it “un-actionable” on every retry. Fixed at the source: the recipes.yml / curate.yml workflows now pre-fetch open user-request issue bodies into .request-bodies/<NN>.md before the agent runs, the responder fallback now rebuilds a structured queue entry from the issue-form fields, and RECIPE_AGENT.md / AGENT.md point the agent at the pre-fetched files instead of a gh issue view it can’t run.

2026-06-08

Added

  • Identify an unknown compound from an MS/MS spectrum (Problem class: Data analysis; Evidence: Proposed) — rung-2 matchms skill recipe taking experimental tandem-MS spectra plus a reference library (GNPS / MassBank / in-house .msp) through format import → peak cleaning and metadata harmonization → modified-cosine scoring with precursor-m/z gating → a ranked candidate-identity CSV, handing confirmed InChIKeys off to the PubChem MCP and the polypharmacology recipe. Chemistry focus-day recipe; cookbook’s first metabolomics / spectral-library-matching recipe. Proposed — no documented LLM-driven matchms workflow; grounded in the canonical library paper Huber et al., JOSS 5(52):2411 (2020) plus methodological anchors Onoprishvili et al., Bioinformatics (2025) (SimMS) and Xing et al., Anal. Chem. (2025) (enhanced reverse spectral search).

Verified (no changes)

  • Aging-recipe sweep: oldest last_verified is 2026-05-24 (15 days), within the 30-day window — no recipes due for re-verification this run.

User requests

  • #12 (@goodb) — still no gh permission to read the issue body from this run; left open for next-run retry.

2026-06-07

Added

  • Enumerate analogs around a lead compound for SAR expansion (Problem class: Hypothesis generation; Evidence: Proposed) — rung-2 Datamol skill recipe taking a lead SMILES through standardization → tautomer / stereoisomer enumeration → single-point fragment-substitution scan → ECFP4 Tanimoto + QED scoring → a deduplicated SAR-expansion CSV, with explicit handoff to the VS-hit-filtering developability gate and the polypharmacology bioactivity lookup. Drug Repurposing and Discovery focus-day recipe; cookbook’s first dedicated analog-enumeration / lead-optimisation recipe and the natural upstream of the existing hit-filtering recipe; cookbook’s second Hypothesis generation recipe. Proposed — no documented LLM-driven Datamol enumeration workflow; closest grounding is the K-Dense rdkit→datamol→medchem lead-optimisation workflow plus the underlying primitives Rogers & Hahn, JCIM 50:742 (2010) (ECFP/Tanimoto), Bickerton et al., Nat. Chem. 4:90 (2012) (QED), and Griffen et al., J. Med. Chem. 54:7739 (2011) (matched molecular pairs).

Updated

  • Nav orders rebalanced to keep alphabetical title ordering after the new addition. “Enumerate analogs…” inserted at 10; everything from “Estimate pharmacokinetic properties” downward shifted +1 (Estimate → 11, Filter VS hits → 12, Infer GRN → 13, Integrate single-cell → 14, Interpret variant → 15, Match patient → 16, Organize DICOM → 17, Parse FCS → 18, Prioritize targets → 19, Profile polypharmacology → 20, Run bulk RNA-seq → 21, Run first-pass QC → 22, Run functional enrichment → 23, Scan repurposing → 24, Set up MD → 25, Sort spikes → 26, Triage preprints → 27, Triage AlphaFold → 28, Fit survival → 29, Scan adverse events → 30).

Verified (no changes)

  • 29 existing recipes spot-checked; none past the 30-day last_verified window (oldest is 2026-05-24, profile-compound-polypharmacology), so no re-verification was due this run.

2026-06-06

Added

  • Fit a survival model to censored clinical outcomes (Problem class: Data analysis; Evidence: Proposed) — rung-2 scikit-survival skill recipe taking a tidy covariate table plus a (time, event) outcome through structured-Surv encoding → Kaplan-Meier + log-rank → Cox PH (with a proportional-hazards check) → Random Survival Forest → cross-validated Harrell’s c-index → risk-group stratification. First Translational Medicine focus-day recipe of this run; cookbook’s first dedicated time-to-event / prognosis recipe. Proposed — no documented end-to-end LLM-driven sksurv workflow; closest grounding is the library reference Pölsterl, JMLR 21(212):1–6 (2020) and recent RSF-vs-nomogram prognosis studies Zhang et al., Transl. Cancer Res. (2026) and Liu et al., Medicine (2026).
  • Scan adverse-event reports for a drug-safety signal (Problem class: Knowledge synthesis; Evidence: Proposed) — rung-2 OpenFDA MCP recipe taking a drug name through generic-name resolution → FAERS top-reaction ranking → structured label / warning pull → label-vs-FAERS cross-check → an honest “reports, not rates” framing. Second Translational Medicine focus-day recipe of this run; promoted from the Deferred — next-run priority list; cookbook’s first pharmacovigilance recipe. Proposed — no documented attempt of this exact MCP assembly; openFDA/FAERS is the canonical public pharmacovigilance source and the server wraps it faithfully.

Verified (no changes)

  • 27 existing recipes spot-checked; none past the 30-day last_verified window (oldest is 2026-05-24), so no re-verification was due this run.

2026-06-05

Added

Updated

  • Nav orders rebalanced to keep alphabetical title ordering after the new addition and to fix a stale collision between Run first-pass QC and Run functional enrichment (both stamped 20). “Organize a raw DICOM dataset…” inserted at 16; everything from “Parse FCS…” downward shifted by +1, with Run first-pass QC at 21 and Run functional enrichment at 22: Parse FCS flow-cytometry files → 17, Prioritize targets → 18, Profile polypharmacology → 19, Run bulk RNA-seq DE → 20, Run first-pass QC → 21, Run functional enrichment → 22, Scan repurposing → 23, Set up protein MD → 24, Sort spikes → 25, Triage preprints → 26, Triage AlphaFold → 27.

Verified (no changes)

  • No aging recipes due — every last_verified date is within the 30-day window. The verification floor sits at 2026-05-24 (scan-drug-repurposing-candidates); next aging boundary is 2026-06-23.

User requests

  • #12 @goodb — still cannot access the issue body (no gh permission for the repo in this run); leaving open in recipes/curator-state.md for the next run with gh access.

2026-06-04

Added

  • Run functional enrichment on a gene list (Problem class: Data analysis; Evidence: Reported) — rung-2 gget skill recipe taking a list of gene symbols through gget enrichr against GO BP, KEGG, Reactome, MSigDB Hallmark, and DisGeNET → per-library CSV → grounded natural-language summary with explicit verification pass against the saved tables and a random-gene negative-control step. First Molecular and Cellular Biology focus-day recipe of this run; the cookbook’s first dedicated functional-enrichment / pathway-interpretation recipe and the natural downstream step after bulk RNA-seq DE. Reported evidence anchored in Wang et al., GeneAgent, Nature Methods 22:1677, 2025 — self-verification against Enrichr and curated databases lifts ROUGE-L on MSigDB from 0.239±0.038 (GPT-4) to 0.310±0.047 (GeneAgent) across 1,106 gene sets, with 84% of 15,848 claims database-supported and 92% of self-verification decisions correct on a 132-claim expert-judged sample; complementary anchors Hu et al., Nat. Methods 21:2353, 2024 and Joshi et al., llm2geneset (bioRxiv 2024-11-12).

Verified (no changes)

User requests

  • #12 @goodb — still cannot access the issue body (no gh permission in this run); leaving open in recipes/curator-state.md for the next run with gh access.

2026-06-03

Added

Updated

  • Nav orders rebalanced to keep alphabetical title ordering after the new addition. “Dock a ligand library…” inserted at 8; everything from “Draft Phase 2/3…” downward shifted by +1: Draft Phase 2/3 clinical-trial protocol → 9, Estimate PK → 10, Filter virtual screening → 11, Infer GRN → 12, Integrate single-cell → 13, Interpret clinical variant → 14, Match patient to trials → 15, Parse FCS flow-cytometry files → 16, Prioritize targets → 17, Profile polypharmacology → 18, Run bulk RNA-seq DE → 19, Run first-pass QC → 20, Scan repurposing → 21, Set up protein MD → 22, Sort spikes → 23, Triage preprints → 24, Triage AlphaFold → 25.

Verified (no changes)

  • No aging recipes due — every last_verified date is within the 30-day window. The recipe set’s verification floor sits at 2026-05-22 (integrate-single-cell-datasets, sort-spikes-from-neuropixels-recording); next aging boundary is 2026-06-21.

User requests

  • #12 @goodb — still cannot access the issue body (no gh permission for the repo in this run); leaving the request open in recipes/curator-state.md for the next run with gh access.

2026-06-02

Added

  • Compute 16S microbiome alpha/beta diversity from a BIOM table (Problem class: Data analysis; Evidence: Proposed) — rung-2 scikit-bio skill recipe taking a BIOM feature table + sample metadata + Newick tree through rarefaction → Shannon/Simpson/Faith’s PD → weighted/unweighted UniFrac → PCoA → PERMANOVA with explicit grouping-column and permutation-count flags. First Immunology and Microbiology focus-day recipe of this run; cookbook’s first dedicated microbiome / community-ecology recipe. Proposed because no documented end-to-end attempt of this exact assembly exists; closest class-level evidence is Huang et al. Biomni (bioRxiv 2025.05.30.656746) whose published benchmark includes microbiome disease-taxa bioinformatics across five datasets (HMP, MetaPhlAn2 human metagenomics, drinking-water OTU matrices) at ~4× over base-LLM accuracy.
  • Parse FCS flow-cytometry files for downstream immunophenotyping (Problem class: Data analysis; Evidence: Proposed) — rung-2 FlowIO skill recipe taking a directory of vendor-emitted FCS 2.0/3.0/3.1 files through FlowData parsing → per-file metadata harvest → scatter/fluorescence/time channel categorisation → optional log/gain transforms → concatenated long-format events Parquet, with explicit failure surfacing for partial-acquisition files. Second Immunology and Microbiology focus-day recipe; cookbook’s first cytometry / FCS recipe. Proposed because no documented end-to-end attempt of this exact assembly exists; closest class-level evidence is “Enhancing Clinical Workflow Efficiency in Flow Cytometry Reporting with LLMs” (PMC13053331, J. Clin. Immunol. 2026), which demonstrates pathologist-level accuracy of fine-tuned LLMs on the downstream report-generation step the parsed-events output feeds into.

Updated

  • Nav orders rebalanced to keep alphabetical title ordering after the two additions: Assemble Census atlas → 1, Benchmark ADMET → 2, Build target dossier → 3, Compute 16S microbiome diversity → 4 (new), Compute HRV → 5, Convert instrument data → 6, Discover NWB on DANDI → 7, Draft Phase 2/3 clinical-trial protocol → 8, Estimate PK → 9, Filter virtual screening → 10, Infer GRN → 11, Integrate single-cell → 12, Interpret clinical variant → 13, Match patient to trials → 14, Parse FCS flow-cytometry files → 15 (new), Prioritize targets → 16, Profile polypharmacology → 17, Run bulk RNA-seq DE → 18, Run first-pass QC → 19, Scan repurposing → 20, Set up protein MD → 21, Sort spikes → 22, Triage preprints → 23, Triage AlphaFold → 24.

Verified (no changes)

  • No aging recipes due — every last_verified date is within the 30-day window. The recipe set’s verification floor sits at 2026-05-22 (integrate-single-cell-datasets, sort-spikes-from-neuropixels-recording); next aging boundary is 2026-06-21.

User requests

  • #12 (claude:recipe-feedback) — remains in ## User requests (open); gh CLI is still not available in this run’s environment so the issue body cannot be inspected. Retry next run with gh access.

2026-06-01

Added

  • Convert raw analytical instrument data to Allotrope ASM JSON (Problem class: Workflow automation; Evidence: Reported) — rung-2 instrument-data-to-allotrope skill recipe taking a vendor-format file (cell counter, plate reader, HPLC, MS, qPCR) through auto-detect → allotropy native parse → ASM JSON-LD + flattened CSV + exportable Python parser, with strict-validation of the raw-vs-derived split before LIMS / data-lake handoff. First Chemistry focus-day recipe of this run; cookbook’s first workflow-automation recipe spanning the Anthropic life-sciences plugin family. Anchored in the Claude for Life Sciences launch (October 2025), the Anthropic Vi-CELL tutorial, and the underlying Benchling-Open-Source/allotropy reference parser.
  • Set up a protein molecular dynamics simulation in GROMACS from a PDB ID (Problem class: Experimental design; Evidence: Proposed) — rung-2 molecule-mcp recipe driving the GROMACS Copilot server end-to-end (topology → solvation → ion neutralisation → minimisation → NVT/NPT → 50 ns production → RMSD/RMSF/Rg) with explicit force-field / water-model / GPU-offload flags. Second Chemistry focus-day recipe; first cookbook entry exercising the GROMACS path of the molecule-mcp bundle. Proposed because no documented end-to-end attempt of this exact assembly exists; closest peer-reviewed class-level evidence is MDCrow (Campbell et al., Mach. Learn. Sci. Technol. 2025, DOI:10.1088/2632-2153/ae4b07) — OpenMM rather than GROMACS but same architecture — plus GROMACS-supporting follow-ons DynaMate (arXiv:2512.10034) and NAMD-Agent (arXiv:2507.07887), and the MDGym benchmark (arXiv:2605.08941) as a reality check (Claude Code / Codex / OpenHands all solve <21% of easy GROMACS/LAMMPS tasks).

Updated

  • Nav orders rebalanced to restore strict alphabetical title ordering after the two additions and to correct two prior off-by-many drifts (Benchmark ADMET was at 20 instead of 2; Prioritize Targets was at 19 instead of 14): Assemble Census atlas → 1, Benchmark ADMET → 2, Build target dossier → 3, Compute HRV → 4, Convert instrument data → 5 (new), Discover NWB on DANDI → 6, Draft a Phase 2/3 clinical-trial protocol → 7, Estimate PK → 8, Filter virtual screening → 9, Infer GRN → 10, Integrate single-cell → 11, Interpret clinical variant → 12, Match patient to trials → 13, Prioritize targets → 14, Profile polypharmacology → 15, Run bulk RNA-seq DE → 16, QC single-cell → 17, Scan repurposing → 18, Set up protein MD in GROMACS → 19 (new), Sort spikes → 20, Triage preprints → 21, Triage AlphaFold → 22.
  • recipes/curator-state.md — ## Missing components entry for “DeepChem (K-Dense Skill)” removed; DeepChem is now catalogued at catalog/tools/deepchem.md.

Verified (no changes)

  • No aging recipes due — every last_verified date is within the 30-day window. The recipe set’s verification floor sits at 2026-05-22 (integrate-single-cell-datasets, sort-spikes-from-neuropixels-recording); next aging boundary is 2026-06-21.

User requests

  • #12 (claude:recipe-feedback) — remains in ## User requests (open); gh CLI is still not available in this run’s environment so the issue body cannot be inspected. Retry next run with gh access.

2026-05-31

Added

Verified (no changes)

  • No aging recipes due — every last_verified date is within the 30-day window. The recipe set’s verification floor sits at 2026-05-22 (integrate-single-cell-datasets, sort-spikes-from-neuropixels-recording); next aging boundary is 2026-06-21.

User requests

  • #12 (claude:recipe-feedback) — remains in ## User requests (open); gh CLI is still not available in this run’s environment so the issue body cannot be inspected. Retry next run with gh access.

2026-05-30

Added

  • Draft a Phase 2/3 clinical-trial protocol from an indication brief (Problem class: Manuscript prep; Evidence: Reported) — rung-2 clinical-trial-protocol Anthropic Healthcare plugin recipe that walks an indication / endpoint paragraph through the four-waypoint flow — regulatory classification, ClinicalTrials.gov competitive landscape, sample-size calculation, FDA/NIH-template drafting — emerging with a reviewable draft Phase 2/3 protocol scaffold. First Translational Medicine focus-day recipe of the new run; resolves a previously deferred candidate. Evidence anchored in the Anthropic plugin tutorial (Claude for Healthcare launch, January 2026) and class-level validation in Markey et al. Clinical Trials 2025 (80% content relevance, >99% terminology accuracy with RAG), Shin et al. Clinical Pharmacology & Therapeutics 2026 (100% accuracy on disease/intervention/comparator extraction, 14/15 trials for sample-size identification), Hauptman et al. JMIR Dermatology 2026, and Maleki, arXiv 2404.05044 (2024).

Updated

  • Nav orders rebalanced across the recipe set to keep alphabetical ordering after the addition: Assemble Census atlas → 1, Build target dossier → 2, Compute HRV → 3, Discover NWB on DANDI → 4, Draft a Phase 2/3 clinical-trial protocol → 5 (new), Estimate PK → 6, Filter virtual screening → 7, Infer GRN → 8, Integrate single-cell → 9, Interpret clinical variant → 10, Match patient to trials → 11, Profile polypharmacology → 12, Run bulk RNA-seq DE → 13, QC single-cell → 14, Scan repurposing → 15, Sort spikes → 16, Triage preprints → 17, Triage AlphaFold → 18.

Verified (no changes)

  • No aging recipes due — every last_verified date is within the 30-day window. The recipe set’s verification floor sits at 2026-05-22 (integrate-single-cell-datasets, sort-spikes-from-neuropixels-recording); next aging boundary is 2026-06-21.

User requests

  • #12 (claude:recipe-feedback) — remains in ## User requests (open); gh CLI is still not available in this run’s environment so the issue body cannot be inspected. Retry next run with gh access.

2026-05-29 (second pass — Neuroscience directed)

Added

Updated

  • Nav orders rebalanced across the recipe set to keep alphabetical ordering after the addition: Assemble Census atlas → 1, Build target dossier → 2, Compute HRV → 3, Discover NWB on DANDI → 4, Estimate PK → 5, Filter virtual screening → 6, Infer GRN → 7, Integrate single-cell → 8, Interpret clinical variant → 9, Match patient to trials → 10, Profile polypharmacology → 11, Run bulk RNA-seq DE → 12, QC single-cell → 13, Scan repurposing → 14, Sort spikes → 15, Triage preprints → 16, Triage AlphaFold → 17.

Verified (no changes)

  • No aging recipes this run — every last_verified date is within the 30-day window. The recipe set’s verification floor sits at 2026-05-22 (integrate-single-cell-datasets, sort-spikes-from-neuropixels-recording); next aging boundary is 2026-06-21.

User requests

  • #12 (claude:recipe-feedback) — remains in ## User requests (open); gh CLI still unavailable in this run’s environment so the issue body cannot be inspected. Retry next run with gh access.

2026-05-29

Added

Updated

  • Nav orders rebalanced across the recipe set to keep alphabetical ordering after the addition: Assemble Census atlas → 1, Build target dossier → 2, Compute HRV → 3, Estimate PK → 4, Filter virtual screening → 5, Infer GRN → 6, Integrate single-cell → 7, Interpret clinical variant → 8, Match patient to trials → 9, Profile polypharmacology → 10, Run bulk RNA-seq DE → 11, QC single-cell → 12, Scan repurposing → 13, Sort spikes → 14, Triage preprints → 15, Triage AlphaFold → 16.

Verified (no changes)

  • 4 recipes spot-checked at the 30-day boundary and bumped to last_verified: 2026-05-29 — Triage preprints, QC single-cell, Build target dossier, Run bulk RNA-seq DE. All linked catalog tools (bio-research, pubmed, single-cell-rna-qc, pydeseq2, open-targets, uniprot, alphafold, depmap) remain present and unflagged.

User requests

  • #12 (claude:recipe-feedback) — remains in ## User requests (open); gh CLI is not available in this run’s environment so the issue body still cannot be inspected. Retry on the next run that has gh access.

2026-05-28

Added

Updated

  • Nav orders rebalanced across the recipe set to keep alphabetical ordering after the two additions: Assemble Census atlas → 1, Build target dossier → 2, Estimate PK → 3, Filter virtual screening → 4, Infer GRN → 5, Integrate single-cell → 6, Interpret clinical variant → 7, Match patient to trials → 8, Profile polypharmacology → 9, Run bulk RNA-seq DE → 10, QC single-cell → 11, Scan repurposing → 12, Sort spikes → 13, Triage preprints → 14, Triage AlphaFold → 15.

Missing components flagged to the catalog curator

  • pySCENIC wrapper (cisTarget + AUCell) — would unlock the full SCENIC pipeline downstream of the new GRN-inference recipe (motif filtering against cisTarget databases, per-cell regulon AUCell scoring).

Verified (no changes)

  • All 13 pre-existing recipes have last_verified within the 30-day window (oldest 2026-05-21); no aging verifications were due this run.

2026-05-27

Added

  • Estimate pharmacokinetic properties of a small molecule (Problem class: Knowledge synthesis; Evidence: Proposed) — rung-3 RDKit + MedChem + ChEMBL assembly producing a descriptor / rule-based / analog-anchored PK card for a single SMILES. Ships in response to user request #8. Closest documented analogues: ChemCrow (Bran et al., Nature Machine Intelligence 2024) and PharmaBench (Niu et al., Scientific Data 2024).
  • Triage an AlphaFold model for structure-based drug design (Problem class: Knowledge synthesis; Evidence: Proposed) — rung-2 AlphaFold MCP recipe producing a pLDDT-anchored go/refine/fall-back-to-PDB verdict on a UniProt accession. First Integrative Structural and Computational Biology-primary recipe. Evidence grounded in the EBI AlphaFold DB papers (Varadi 2022, Varadi 2024), the interface-pLDDT benchmark (Bryant 2022), and the AlphaFold-for-docking assessment (Karelina 2023).

Updated

  • Nav orders rebalanced across the recipe set to keep alphabetical ordering after the two additions: Estimate PK properties → 2, Filter virtual screening hits → 3, Integrate single-cell datasets → 4, Interpret clinical variant → 5, Match patient to trials → 6, Profile polypharmacology → 7, Run bulk RNA-seq DE → 8, QC single-cell RNA-seq → 9, Scan repurposing candidates → 10, Sort spikes → 11, Triage preprints → 12, Triage AlphaFold model → 13.

Missing components flagged to the catalog curator

  • ADMET-AI / AdmetLab 3.0 / Deep-PK wrapper — would let the new PK-properties recipe move from descriptor-and-analog estimation to defensible ML prediction for CYP / hERG / microsomal endpoints.
  • DeepChem (K-Dense Skill) — already flagged in the catalog curator’s state; would also strengthen the PK-properties recipe.
  • Co-folding / AlphaFold-Multimer / Boltz-2 wrapper — would unlock a complex-modelling companion to the AlphaFold triage recipe.

Verified (no changes)

  • All recipes have last_verified within the 30-day window; no aging verifications were due this run.

2026-05-25

Added

  • Filter a virtual screening hit list with drug-likeness rules and structural alerts (Problem class: Data analysis; Evidence: Reported) — rung-2 MedChem + Datamol cascade for Lipinski → Veber → PAINS → BRENK triage of SMILES hit lists. First Chemistry-primary recipe in the cookbook. Evidence anchored in the K-Dense lead-optimisation workflow and the foundational filter papers (Baell & Holloway PAINS 2010, Brenk 2008, Lipinski 2001, Veber 2002).
  • Profile a compound’s polypharmacology from ChEMBL bioactivity data (Problem class: Knowledge synthesis; Evidence: Reported) — rung-2 single-tool recipe over the ChEMBL connector. Second Chemistry-primary recipe and the compound-centric mirror of the existing target-dossier recipe. Evidence grounded in the Anthropic ChEMBL Connector tutorial and the ChEMBL curation paper (Mendez et al., NAR 2019).

Updated

  • Integrate multiple single-cell RNA-seq datasets across batches — nav_order 2 → 3 for alphabetical position after the new Filter recipe.
  • Interpret a clinical variant from a natural-language query — nav_order 3 → 4.
  • Match a patient summary to recruiting clinical trials — nav_order 4 → 5.
  • Run bulk RNA-seq differential expression from a counts matrix — nav_order 5 → 7 (after the new Profile recipe).
  • Run first-pass QC on a single-cell RNA-seq dataset — nav_order 6 → 8.
  • Scan approved drugs for repurposing candidates against a disease — nav_order 7 → 9.
  • Sort spikes from a Neuropixels recording end-to-end — nav_order 8 → 10.
  • Triage a stack of new preprints in your field — nav_order 9 → 11.

Verified (no changes)

  • 9 existing recipes spot-checked; all last_verified dates within the 30-day window, all linked catalog pages resolve.

2026-05-24

Added

Updated

  • Sort spikes from a Neuropixels recording end-to-end — nav_order 7 → 8 for alphabetical position.
  • Triage a stack of new preprints in your field — nav_order 8 → 9 for alphabetical position.

Verified (no changes)

  • 8 existing recipes spot-checked; all last_verified dates within the 30-day window, all linked catalog pages resolve.

2026-05-23

Added

  • Match a patient summary to recruiting clinical trials (Problem class: Knowledge synthesis; Evidence: Reported) — rung-2 BioMCP / cyanheads-ClinicalTrials.gov-MCP recipe; first Translational-Medicine-focused recipe in the cookbook. Evidence grounded in TrialGPT (Jin et al., Nature Communications 2024, 87.3% criterion-matching accuracy).
  • Interpret a clinical variant from a natural-language query (Problem class: Knowledge synthesis; Evidence: Proposed) — rung-2 BioMCP recipe; pairs with the trial-matching recipe for variant-driven enrollment. Closest analogous benchmark is MARRVEL-MCP (bioRxiv 2025-11).

Updated

  • Run bulk RNA-seq differential expression from a counts matrix — nav_order 3 → 5 for alphabetical position after the two new TM recipes.
  • Run first-pass QC on a single-cell RNA-seq dataset — nav_order 4 → 6 for alphabetical position.
  • Sort spikes from a Neuropixels recording end-to-end — nav_order 5 → 7 for alphabetical position.
  • Triage a stack of new preprints in your field — nav_order 6 → 8 for alphabetical position.

Verified (no changes)

  • 5 existing recipes spot-checked; all last_verified dates within the 30-day window, all linked catalog pages resolve.

2026-05-22

Added

  • Integrate multiple single-cell RNA-seq datasets across batches (Problem class: Data analysis; Evidence: Reported) — rung-2 recipe wrapping the Anthropic scvi-tools skill for scVI / scANVI batch integration; written in response to user request #7; evidence grounded in Hrovatin 2025 and scIB-E 2025 (source).
  • Sort spikes from a Neuropixels recording end-to-end (Problem class: Data analysis; Evidence: Reported) — rung-2 recipe wrapping the K-Dense neuropixels-analysis skill (SpikeInterface + Kilosort4); first Neuroscience-only recipe in the cookbook (source).

Updated

  • Run bulk RNA-seq differential expression from a counts matrix — nav_order shifted 2 → 3 for alphabetical position.
  • Run first-pass QC on a single-cell RNA-seq dataset — nav_order shifted 3 → 4 for alphabetical position.
  • Triage a stack of new preprints in your field — nav_order shifted 4 → 6 for alphabetical position.

Verified (no changes)

  • 4 existing recipes spot-checked (all linked catalog pages resolve; last_verified 2026-05-21 still within the 30-day window so no bumps).

2026-05-21

Added

  • Run first-pass QC on a single-cell RNA-seq dataset (Problem class: Data analysis; Evidence: Reported) — rung-2 recipe wrapping Anthropic’s single-cell-rna-qc skill for canonical scverse MAD-based filtering of 10x .h5 / AnnData .h5ad inputs (source).
  • Run bulk RNA-seq differential expression from a counts matrix (Problem class: Data analysis; Evidence: Reported) — rung-2 recipe wrapping the K-Dense PyDESeq2 skill for negative-binomial GLM differential expression, including pseudobulk single-cell handoff guidance (source).
  • Build a target dossier from gene name to structure to cancer dependency (Problem class: Knowledge synthesis; Evidence: Proposed) — first rung-3 toolbelt recipe composing Open Targets, UniProt, AlphaFold, and DepMap into a one-page target dossier; first Proposed-evidence entry in the cookbook (closest analogue).

Updated

  • Triage a stack of new preprints in your field — nav_order shifted from 1 to 4 to reflect alphabetical ordering after the three new Mol/Cell Bio additions; no content changes.

Verified (no changes)

  • 1 recipe spot-checked, current (triage-new-preprints, last_verified 2026-05-21).

2026-05-21 (initial seed)

Added

  • Section bootstrap — recipes/ section created with landing page, landscape page, and the all-recipes index; recipes/curator-state.md initialized; RECIPES_CHANGELOG.md (this file) created. Curator prompt and daily workflow added at RECIPE_AGENT.md and .github/workflows/recipes.yml.
  • Triage a stack of new preprints in your field (Problem class: Literature triage; Evidence: Reported) — first seed recipe demonstrating the schema and the lowest rung of the simplicity ladder (Claude Code alone + bioRxiv MCP) (source).