Recipe landscape

This cookbook pairs concrete life-science problems with recommended assemblies of the components catalogued elsewhere on this site. Each recipe is the answer to “I have this problem — what’s the simplest viable assembly that has evidence behind it?” The recipes are deliberately biased toward the lowest rung of the simplicity ladder that actually solves the problem.

Coverage by problem class

Literature triage

  • Triage a stack of new preprints in your field — One skill or MCP · Reported · Fully open · Laptop. Use the bio-research plugin in Claude Code to pull last-week bioRxiv and PubMed hits, rank by relevance, and produce a one-paragraph readout per paper.

Hypothesis generation

  • Enumerate analogs around a lead compound for SAR expansion — One skill or MCP · Proposed · Fully open · Laptop. Use the Datamol skill in Claude Code to enumerate standardized, drug-like analogs around a lead SMILES — tautomers, stereoisomers, and fragment-substituted variants — as a ranked SAR-expansion table.

Experimental design

  • Scan a therapeutic antibody for glycosylation sites — One skill or MCP · Proposed · Fully open · Laptop. Use the Glycoengineering skill to scan an antibody sequence for N-glycosylation sequons and O-glycosylation hotspots before committing to a cell-line or developability campaign.

  • Set up a protein molecular dynamics simulation in GROMACS from a PDB ID — One skill or MCP · Proposed · Fully open · Workstation with GPU. Drive the GROMACS Copilot component of molecule-mcp to take a PDB ID through topology, solvation, ion neutralisation, minimisation, NVT/NPT equilibration, and a short production MD, with RMSD/RMSF/Rg analysis on the resulting trajectory.

Data analysis

  • Analyze an existing MD trajectory for stability, flexibility, and contacts — One skill or MCP · Proposed · Fully open · Workstation with GPU. Drive the MDAnalysis skill (MDTraj as backup) to take a finished GROMACS/AMBER/NAMD trajectory through RMSD/RMSF/Rg, contact maps, H-bonds, and PCA without writing the analysis script by hand.

  • Assemble a tissue reference atlas from the CELLxGENE Census — One skill or MCP · Reported · Fully open · Workstation with GPU. Use the cellxgene-census Claude skill to slice a tissue- or disease-specific AnnData from the CZ CELLxGENE Census, optionally fetching the precomputed scVI embedding for reference mapping or downstream integration.

  • Benchmark an ADMET property with PyTDC — One skill or MCP · Reported · Fully open · Laptop. Drive the PyTDC Claude skill to load a Therapeutics Data Commons ADMET dataset with its leaderboard split, train or score a baseline model, and emit the standard TDC metrics so a new method can be compared head-to-head against the leaderboard.

  • Build a phylogenetic tree from a set of sequences — One skill or MCP · Proposed · Fully open · Laptop. Use the phylogenetics Claude skill to take a FASTA of homologous sequences through MAFFT alignment, IQ-TREE 2 maximum-likelihood inference with bootstrap support, and an annotated tree figure.

  • Call peaks and find enriched motifs from ChIP-seq or ATAC-seq — Multi-tool harness · Proposed · Fully open · Laptop. Chain the MACS3 peak-calling skill into the HOMER motif skill to turn aligned ChIP-seq/ATAC-seq BAMs into a called-peak set with nearest-gene annotation and enriched TF motifs.

  • Compute 16S microbiome alpha/beta diversity from a BIOM table — One skill or MCP · Proposed · Fully open · Laptop. Use the scikit-bio Claude skill to take a BIOM feature table and sample metadata through Shannon/Simpson/Faith’s PD, UniFrac, PCoA, and PERMANOVA in one chat.

  • Compute HRV from an ECG recording — One skill or MCP · Proposed · Fully open · Laptop. Use the NeuroKit2 Claude Skill to clean an ECG trace, detect R-peaks, and return time-domain, frequency-domain, and non-linear HRV indices from a single conversation.

  • Compute a bacterial pan-genome from a set of genome assemblies — Multi-tool harness · Proposed · Fully open · Workstation with GPU. Annotate a panel of bacterial assemblies with Bakta, then cluster genes into core/accessory partitions with Roary to get a pan-genome and presence/absence matrix.

  • Dock a ligand library into a target structure with DiffDock — One skill or MCP · Proposed · Fully open · Workstation with GPU. Drive DiffDock from Claude Code to generate per-ligand binding poses against a PDB or AlphaFold target, filter by confidence, and emit a ranked pose set ready for MM/PBSA rescoring or visualization.

  • Filter a virtual screening hit list with drug-likeness rules and structural alerts — One skill or MCP · Reported · Fully open · Laptop. Use the MedChem skill in Claude Code to cascade Lipinski / Veber / PAINS / BRENK filters over a SMILES hit list and emit a triaged CSV with per-rule flags and a final keep/drop column.

  • Fit a survival model to censored clinical outcomes — One skill or MCP · Proposed · Fully open · Laptop. Use the scikit-survival Claude Skill to fit Cox and Random Survival Forest models to censored time-to-event data and report concordance with proper validation.

  • Identify an unknown compound from an MS/MS spectrum — One skill or MCP · Proposed · Fully open · Laptop. Use the matchms skill in Claude Code to clean an unknown MS/MS spectrum and rank candidate identities by spectral similarity against a reference library.

  • Infer a gene-regulatory network from single-cell RNA-seq — One skill or MCP · Reported · Fully open · Workstation with GPU. Use the Arboreto Claude skill to run GRNBoost2 on a QC’d single-cell AnnData and recover a ranked TF–target edge list — the first step of a SCENIC regulon pipeline.

  • Infer cell-cell communication from single-cell RNA-seq — One skill or MCP · Proposed · Fully open · Laptop. Use the LIANA-MCP server to run multiple ligand-receptor methods on an annotated AnnData object, aggregate their ranks, and plot the consensus communication network.

  • Integrate multiple single-cell RNA-seq datasets across batches — One skill or MCP · Reported · Fully open · Workstation with GPU. Use the scvi-tools skill in Claude Code to fit scVI (or scANVI when labels are available) on a concatenated AnnData of multiple batches, returning a batch-corrected latent space and integrated UMAP.

  • Parse FCS flow-cytometry files for downstream immunophenotyping — One skill or MCP · Proposed · Fully open · Laptop. Use the FlowIO Claude skill to parse FCS 2.0/3.0/3.1 files into tidy DataFrames with channel categorisation and batch metadata extraction, ready for downstream immunophenotyping or QC.

  • Predict gene-knockout phenotypes with flux balance analysis — One skill or MCP · Proposed · Fully open · Laptop. Use the COBRApy Claude Skill to load a genome-scale metabolic model and screen single/double gene deletions for growth phenotypes and essential genes.

  • Profile ChIP-seq or ATAC-seq signal around genomic features — One skill or MCP · Proposed · Fully open · Laptop. Use the deepTools Claude Skill to turn aligned ChIP-seq/ATAC-seq BAMs into normalized bigWig tracks and TSS/peak-centered profile and heatmap figures.

  • Run a GWAS on case-control genotype data — One skill or MCP · Proposed · Fully open · Workstation with GPU. Drive PLINK2 from QC through PCA-adjusted logistic-regression association on a case-control cohort, returning Manhattan-ready summary stats.

  • Run bulk RNA-seq differential expression from a counts matrix — One skill or MCP · Reported · Fully open · Laptop. Use the PyDESeq2 Claude Skill to fit a negative-binomial GLM from a counts matrix and sample-condition table, returning ranked log2 fold-change and BH-adjusted p-values.

  • Run first-pass QC on a single-cell RNA-seq dataset — One skill or MCP · Reported · Fully open · Laptop. Use Anthropic’s single-cell-rna-qc skill to MAD-filter a 10x .h5 or AnnData .h5ad file and emit standard QC figures before downstream Scanpy or scvi-tools work.

  • Run functional enrichment on a gene list — One skill or MCP · Reported · Fully open · Laptop. Use the gget Claude skill to run a gene list through Enrichr against GO, KEGG, Reactome, and disease libraries, then ask Claude to summarise the enriched terms back to the biology with explicit citations.

  • Score point mutations for functional impact with a protein language model — One skill or MCP · Proposed · Fully open · Workstation with GPU. Use the ESM skill to compute zero-shot masked-marginal log-likelihood-ratio scores for a list of point mutations, ranking them by predicted functional effect without any labelled data.

  • Sort spikes from a Neuropixels recording end-to-end — One skill or MCP · Reported · Fully open · Workstation with GPU. Use the K-Dense neuropixels-analysis skill in Claude Code to take a raw SpikeGLX or Open Ephys Neuropixels recording through preprocessing, drift correction, and Kilosort4 to curated single units.

Knowledge synthesis

  • Build a pharmacogenomic dosing report from a patient’s diplotypes — One skill or MCP · Proposed · Fully open · Laptop. Turn a patient’s PGx diplotypes plus a medication list into CPIC/DPWG genotype-to-dosing recommendations with cited guideline sources.

  • Build a target dossier from gene name to structure to cancer dependency — Multi-tool harness · Proposed · Fully open · Laptop. Combine the Open Targets plugin, UniProt MCP, AlphaFold MCP, and the DepMap skill to produce a one-page dossier on a candidate gene — disease evidence, protein annotation, predicted structure, and cancer-cell-line essentiality.

  • Discover NWB recordings on DANDI and prepare them for sorting — Multi-tool harness · Reported · Fully open · Laptop. Use the Neurosift Tools MCP to find Neuropixels (or other extracellular) NWB sessions on DANDI matching your hypothesis, inspect them in-place, and stage a download list for downstream spike sorting.

  • Estimate pharmacokinetic properties of a small molecule — Multi-tool harness · Reported · Fully open · Laptop. Combine RDKit physchem descriptors, MedChem rule-based ADMET flags, and measured ChEMBL bioactivity / DMPK endpoints to build a defensible PK estimate for one compound without an ML predictor in the loop.

  • Interpret a clinical variant from a natural-language query — One skill or MCP · Proposed · Fully open · Laptop. Use BioMCP to convert a free-text variant query (“BRAF V600E”, “rs113993960”, “NM_004985.5:c.35G>A”) into a one-page clinical report with ClinVar significance, population frequency, in-silico predictions, and linked literature.

  • Match a patient summary to recruiting clinical trials — One skill or MCP · Reported · Fully open · Laptop. Use BioMCP (or the standalone ClinicalTrials.gov MCP) to take a free-text patient summary and return a ranked list of currently-recruiting trials with eligibility rationale.

  • Prioritize targets within a disease via Open Targets — One skill or MCP · Reported · Fully open · Laptop. Drive the Open Targets MCP plugin to rank candidate targets within a disease across the four prioritisation pillars (precedence, tractability, doability, safety) and emit a cited target shortlist.

  • Profile a cancer cohort’s genomics with cBioPortal — One skill or MCP · Reported · Fully open · Laptop. Use the cBioPortal Claude Skill to query TCGA-scale studies for mutation frequency, oncoprints, TMB, and survival by genomic subgroup in plain language.

  • Profile a compound’s polypharmacology from ChEMBL bioactivity data — One skill or MCP · Reported · Fully open · Laptop. Use the ChEMBL connector in Claude Code to pull every measured activity for one compound, group by target and assay, and surface the off-target profile a medicinal chemist needs before committing to a chemotype.

  • Scan adverse-event reports for a drug-safety signal — One skill or MCP · Proposed · Fully open · Laptop. Use the OpenFDA MCP server to pull FAERS adverse-event reports and label warnings for a drug, rank the top reported reactions, and assemble a pharmacovigilance snapshot.

  • Scan approved drugs for repurposing candidates against a disease — Multi-tool harness · Proposed · Subscription required · Laptop. Given a disease, use Open Targets to rank associated targets, ChEMBL and PubChem for approved-drug bioactivity, and DrugBank for mechanism / indication / interaction context to produce a ranked shortlist of repurposing candidates.

  • Screen a polypharmacy medication list for drug-drug interactions — One skill or MCP · Reported · Fully open · Laptop. Ground Claude in the DDInter skill to screen a multi-drug medication list for pairwise interactions with severity and mechanism, instead of asking the model from memory.

  • Triage an AlphaFold model for structure-based drug design — One skill or MCP · Proposed · Fully open · Laptop. Pull a UniProt AlphaFold prediction, surface pLDDT regions and pocket-residue confidence, and produce a go/no-go verdict on whether the model is fit for downstream docking or modelling — all from one MCP server.

Manuscript prep

  • Draft a Phase 2/3 clinical-trial protocol from an indication brief — One skill or MCP · Reported · Fully open · Laptop. Use the Anthropic clinical-trial-protocol plugin to expand a short indication / endpoint brief into an FDA/NIH-compliant Phase 2/3 protocol draft, with regulatory pathway, competitive landscape, and sample-size calculation already wired in.

Workflow automation

  • Convert raw analytical instrument data to Allotrope ASM JSON — One skill or MCP · Reported · Fully open · Laptop. Drop a raw instrument file (cell counter, plate reader, HPLC, qPCR) onto Claude Code and get back valid Allotrope Simple Model JSON, a flattened CSV for LIMS, and a standalone Python parser script.

  • Organize a raw DICOM dataset into a BIDS layout — One skill or MCP · Proposed · Fully open · Laptop. Drive HeuDiConv or dcm2bids from Claude Code via the K-Dense BIDS skill to convert a raw DICOM dump into a validated BIDS dataset ready for BIDS-Apps and OpenNeuro submission.

How evidence is distributed

The cookbook mixes three evidence levels:

  • Validated — at least one peer-reviewed paper or independent benchmark reports the assembly working.
  • Reported — a preprint, blog post, or case study documents someone running the assembly.
  • Proposed — rational composition from the catalog; no documented attempt is yet known.

Use the front-matter on each recipe to filter by the bar you need.

Gaps

Problem classes with no recipes yet are listed above as _No recipes yet._. Subject areas that are under-covered are noted in the curator’s internal state file as Deferred candidates and surfaced here when they cluster into a pattern.