From mims-harvard-tooluniverse
Discovers novel small molecule binders for protein targets using structure-based and ligand-based approaches. Generates reports with candidate compounds, ADMET profiles, and synthesis feasibility for virtual screening and hit-to-lead.
npx claudepluginhub joshuarweaver/cascade-data-analytics --plugin mims-harvard-tooluniverseThis skill uses the workspace's default tool permissions.
Systematic discovery of novel small molecule binders using 60+ ToolUniverse tools across druggability assessment, known ligand mining, similarity expansion, ADMET filtering, and synthesis feasibility.
Conducts multi-round deep research on GitHub repos via API and web searches, generating markdown reports with executive summaries, timelines, metrics, and Mermaid diagrams.
Dynamically discovers and combines enabled skills into cohesive, unexpected delightful experiences like interactive HTML or themed artifacts. Activates on 'surprise me', inspiration, or boredom cues.
Generates images from structured JSON prompts via Python script execution. Supports reference images and aspect ratios for characters, scenes, products, visuals.
Systematic discovery of novel small molecule binders using 60+ ToolUniverse tools across druggability assessment, known ligand mining, similarity expansion, ADMET filtering, and synthesis feasibility.
LOOK UP DON'T GUESS - Always retrieve actual data from tools before drawing conclusions. Do not assume druggability, binding sites, or compound properties based on target class alone.
KEY PRINCIPLES:
Before any tool call, reason about the target's structural biology:
Is the binding site a well-defined pocket (small molecule accessible) or a flat protein-protein interface (needs peptide/macrocycle)? This determines your screening strategy.
Use this reasoning to select phases and warn the user about challenges before executing a full workflow.
DO NOT show search process or tool outputs to the user. Instead:
Create the report file FIRST - Before any data collection:
[TARGET]_binder_discovery_report.md[Researching...] in each sectionProgressively update the report - As you gather data, update each section immediately.
Output separate data files:
[TARGET]_candidate_compounds.csv - Prioritized compounds with SMILES, scores[TARGET]_bibliography.json - Literature references (optional)Every piece of information MUST include its source:
Example: *Source: ChEMBL via ChEMBL_get_target_activities (CHEMBL203)*
Phases in order:
get_tool_info)CRITICAL: Verify tool parameters before calling unfamiliar tools.
tool_info = tu.tools.get_tool_info(tool_name="ChEMBL_get_target_activities")
Common parameter corrections (verify with get_tool_info if uncertain):
OpenTargets_*: ensemblId (camelCase); ADMETAI_*: smiles must be a listNvidiaNIM_alphafold2: sequence not seq; NvidiaNIM_genmol: SMILES must contain [*{min-max}]NvidiaNIM_boltz2: polymers=[{"molecule_type": "protein", "sequence": "..."}]Resolve all IDs upfront and store for downstream queries:
1. UniProt_search(query=target_name, organism="human") -> UniProt accession
2. MyGene_query_genes(q=gene_symbol, species="human") -> Ensembl gene ID
3. ChEMBL_search_targets(query=target_name, organism="Homo sapiens") -> ChEMBL target ID
4. GtoPdb_get_targets(query=target_name) -> GtoPdb ID (if GPCR/channel/enzyme)
Use multi-source triangulation:
OpenTargets_get_target_tractability_by_ensemblID(ensemblId) - tractability bucketDGIdb_get_gene_druggability(genes=[gene_symbol]) - druggability categoriesOpenTargets_get_target_classes_by_ensemblID(ensemblId) - target classGPCRdb_get_protein + GPCRdb_get_ligands + GPCRdb_get_structuresTheraSAbDab_search_by_target(target=target_name)Decision Point: If no tractability data and binding site reasoning suggests PPI or disordered region, explicitly warn the user before proceeding.
ChEMBL_search_binding_sites(target_chembl_id)get_binding_affinity_by_pdb_id(pdb_id) for co-crystallized ligandsInterPro_get_protein_domains(accession) for domain architectureRequires NVIDIA_API_KEY. Two options:
NvidiaNIM_alphafold2(sequence, algorithm="mmseqs2") - high accuracy, 5-15 minESMFold_predict_structure(sequence) - fast (~30s), max 1024 AApLDDT guidance: >=90 very high confidence, 70-90 confident, <70 use with caution. Low pLDDT in the putative binding region undermines docking reliability.
Priority order for bioactivity data:
ChEMBL_get_target_activities - curated, SAR-readyBindingDB_get_ligands_by_uniprot - direct Ki/Kd with literature linksGtoPdb_search_ligands - pharmacology focus (GPCRs, channels)PubChem_search_assays_by_target_gene - HTS screens, novel scaffoldsOpenTargets_get_chemical_probes_by_target_ensemblID - validated probesKey steps:
BindingDB_get_targets_by_compoundTools:
PDB_search_similar_structures(query=uniprot, type="sequence") - find PDB entriesget_protein_metadata_by_pdb_id(pdb_id) - resolution, methodget_binding_affinity_by_pdb_id(pdb_id) - co-crystal ligand affinitiesget_ligand_smiles_by_chem_comp_id(chem_comp_id) - ligand SMILES from PDBemdb_search(query) - cryo-EM structures (prefer for GPCRs, ion channels)alphafold_get_prediction(qualifier) - AlphaFold DB fallbackIf PDB + SDF available: use get_diffdock_info(protein=PDB, ligand=SDF, num_poses=10).
If only sequence + SMILES: use NvidiaNIM_boltz2(polymers=[...], ligands=[...]).
Dock a known reference inhibitor first to validate the binding pocket geometry before running candidates.
Use 3-5 diverse actives as seeds, similarity threshold 70-85%:
ChEMBL_search_similar_molecules(molecule=SMILES, similarity=70)PubChem_search_compounds_by_similarity(smiles, threshold=0.7)ChEMBL_search_substructure(smiles=core_scaffold)STITCH_get_chemical_protein_interactions(identifier=gene, species=9606)GenMol - scaffold hopping with masked regions:
NvidiaNIM_genmol(smiles="...core...[*{3-8}]...tail...[*{1-3}]...", num_molecules=100, temperature=2.0, scoring="QED")
MolMIM - controlled analog generation:
NvidiaNIM_molmim(smi=reference_smiles, num_molecules=50, algorithm="CMA-ES")
Apply sequentially (all tools accept smiles=[list]):
ADMETAI_predict_physicochemical_properties - Lipinski violations <= 1, QED > 0.3, MW 200-600ADMETAI_predict_bioavailability - oral bioavailability > 0.3ADMETAI_predict_toxicity - AMES < 0.5, hERG < 0.5, DILI < 0.5ADMETAI_predict_CYP_interactions - flag CYP3A4 inhibitorsChEMBL_search_compound_structural_alerts - no PAINSInclude a filter funnel summary in the report showing pass/fail counts at each stage.
Composite score: docking confidence (40%) + ADMET score (30%) + similarity to known active (20%) + novelty (10%, not in ChEMBL + novel scaffold bonus).
Evidence tiers for candidates:
Deliver top 20 candidates with: Rank, ID, SMILES, docking score, ADMET score, overall score, source, evidence tier.
PubMed_search_articles(query="[TARGET] inhibitor SAR") - peer-reviewedEuropePMC_search_articles(query, source="PPR") - preprints (not peer-reviewed)openalex_search_works(query) - citation analysisTarget ID: ChEMBL_search_targets -> GtoPdb_get_targets -> "Not in databases"
Druggability: OpenTargets tractability -> DGIdb druggability -> target class proxy
Bioactivity: ChEMBL -> BindingDB -> GtoPdb -> PubChem BioAssay -> "No data"
Structure: PDB -> EMDB (membrane) -> NvidiaNIM_alphafold2 -> NvidiaNIM_esmfold -> AlphaFold DB -> "None"
Similarity: ChEMBL similar -> PubChem similar -> "Search failed"
Docking: get_diffdock_info -> NvidiaNIM_boltz2 -> similarity-based scoring
Generation: NvidiaNIM_genmol -> NvidiaNIM_molmim -> similarity search only
Literature: PubMed -> EuropePMC (preprints) -> OpenAlex
GPCR data: GPCRdb_get_protein -> GtoPdb_get_targets
When ToolUniverse tools return limited compound sets, access chemical databases directly:
import requests, pandas as pd
# PubChem batch property retrieval (up to 100 CIDs per call)
cids = "2244,5988,3672"
url = f"https://pubchem.ncbi.nlm.nih.gov/rest/pug/compound/cid/{cids}/property/MolecularWeight,XLogP,TPSA,HBondDonorCount,HBondAcceptorCount/JSON"
props = pd.DataFrame(requests.get(url).json()["PropertyTable"]["Properties"])
# ChEMBL bioactivity bulk download for a target
target_id = "CHEMBL203" # EGFR
url = f"https://www.ebi.ac.uk/chembl/api/data/activity.json?target_chembl_id={target_id}&pchembl_value__gte=5&limit=1000"
activities = requests.get(url).json()["activities"]
df = pd.DataFrame(activities)[["molecule_chembl_id", "canonical_smiles", "pchembl_value", "standard_type"]]
# Lipinski Rule of 5 filtering (no RDKit needed)
lipinski = props[(props["MolecularWeight"] <= 500) & (props["XLogP"] <= 5) &
(props["HBondDonorCount"] <= 5) & (props["HBondAcceptorCount"] <= 10)]
# SDF download from PubChem (for docking input)
sdf_url = f"https://pubchem.ncbi.nlm.nih.gov/rest/pug/compound/cid/{cids}/SDF"
sdf_content = requests.get(sdf_url).text
See tooluniverse-data-wrangling skill for format cookbook and pagination patterns.
AlphaFold2: 5-15 min (async, max ~2000 AA). ESMFold: ~30 sec (max 1024 AA). DiffDock: ~1-2 min/ligand. Boltz2: ~2-5 min. GenMol/MolMIM: ~1-3 min.
Always check: import os; nvidia_available = bool(os.environ.get("NVIDIA_API_KEY"))
For large expansions (>500 compounds): batch in chunks of 100, prioritize top candidates for docking.