mims-harvard/tooluniverse-binder-discovery

Small Molecule Binder Discovery Strategy

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:

1. Report-first approach - Create report file FIRST, then populate progressively
2. Target validation FIRST - Confirm druggability before compound searching
3. Multi-strategy approach - Combine structure-based and ligand-based methods
4. ADMET-aware filtering - Eliminate poor compounds early
5. Evidence grading - Grade candidates by supporting evidence
6. Actionable output - Provide prioritized candidates with rationale
7. English-first queries - Always use English terms in tool calls. Respond in the user's language

---

Binding Site Reasoning (Start Here)

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.

• Enzymes with active sites (proteases, kinases, ATPases): deep, well-defined pockets. Classic small molecule territory. Prioritize co-crystal structure search and known inhibitor scaffold analysis.
• GPCRs and ion channels: transmembrane pockets. Structure often available; start with GPCRdb and GtoPdb for known pharmacology.
• Nuclear receptors: deep hydrophobic pockets. Excellent small molecule tractability; ligand-based methods are well-powered.
• Protein-protein interfaces: flat, large contact surface. Small molecules rarely compete effectively unless there is a "hot spot" cavity. Check whether any allosteric pockets exist before committing to small molecule strategy. Warn the user if no pocket is found.
• Intrinsically disordered regions: essentially no small molecule approach. Redirect to peptide or degrader strategies.
• Scaffolding / adaptor proteins: assess co-crystal structures for unexpected pockets before declaring undruggable.

Use this reasoning to select phases and warn the user about challenges before executing a full workflow.

---

Critical Workflow Requirements

1. Report-First Approach (MANDATORY)

DO NOT show search process or tool outputs to the user. Instead:

1. Create the report file FIRST - Before any data collection:

   • File name: [TARGET]_binder_discovery_report.md
   • Initialize with all section headers from the template (see REPORT_TEMPLATE.md)
   • Add placeholder text: [Researching...] in each section

2. Progressively update the report - As you gather data, update each section immediately.

3. Output separate data files:

   • [TARGET]_candidate_compounds.csv - Prioritized compounds with SMILES, scores
   • [TARGET]_bibliography.json - Literature references (optional)

2. Citation Requirements (MANDATORY)

Every piece of information MUST include its source:

Example: *Source: ChEMBL via ChEMBL_get_target_activities (CHEMBL203)*

---

Workflow Overview

Phases in order:

• Phase 0: Tool verification (check parameter names with get_tool_info)
• Phase 1: Target validation — resolve IDs, assess druggability, identify binding sites, predict structure if needed
• Phase 2: Known ligand mining — ChEMBL, BindingDB, GtoPdb, PubChem BioAssay, chemical probes; SAR analysis
• Phase 3: Structure analysis — PDB co-crystals, EMDB (membrane targets), binding pocket characterization
• Phase 3.5: Docking validation — dock reference inhibitor to validate pocket geometry
• Phase 4: Compound expansion — similarity/substructure search (seeds: 3-5 diverse actives) + de novo generation
• Phase 5: ADMET filtering — physicochemical, bioavailability, toxicity, CYP, structural alerts
• Phase 6: Candidate docking and prioritization — score and rank top 20
• Phase 6.5: Literature evidence — PubMed, EuropePMC, OpenAlex
• Phase 7: Report synthesis and delivery

---

Phase 0: Tool Verification

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 list
• NvidiaNIM_alphafold2 (requires NVIDIA_API_KEY env var; free key at build.nvidia.com): sequence not seq; NvidiaNIM_genmol (requires NVIDIA_API_KEY env var; free key at build.nvidia.com): SMILES must contain [*{min-max}]
• NvidiaNIM_boltz2 (requires NVIDIA_API_KEY env var; free key at build.nvidia.com): polymers=[{"molecule_type": "protein", "sequence": "..."}]

---

Phase 1: Target Validation

1.1 Identifier Resolution

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_search_targets(query=target_name) -> GtoPdb ID (if GPCR/channel/enzyme)

1.2 Druggability Assessment

Use multi-source triangulation:

• OpenTargets_get_target_tractability_by_ensemblID(ensemblId) - tractability bucket
• DGIdb_get_gene_druggability(genes=[gene_symbol]) - druggability categories
• OpenTargets_get_target_classes_by_ensemblID(ensemblId) - target class
• For GPCRs: GPCRdb_get_protein + GPCRdb_get_ligands + GPCRdb_get_structures
• For antibody landscape: TheraSAbDab_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.

1.3 Binding Site Analysis

• ChEMBL_search_binding_sites(target_chembl_id)
• get_binding_affinity_by_pdb_id(pdb_id) for co-crystallized ligands
• InterPro_get_protein_domains(accession) for domain architecture

1.4 Structure Prediction (NVIDIA NIM)

Requires NVIDIA_API_KEY. Two options:

• AlphaFold2: NvidiaNIM_alphafold2(sequence, algorithm="mmseqs2") - high accuracy, 5-15 min
• ESMFold: ESMFold_predict_structure(sequence) - fast (~30s), max 1024 AA

pLDDT guidance: >=90 very high confidence, 70-90 confident, <70 use with caution. Low pLDDT in the putative binding region undermines docking reliability.

---

Phase 2: Known Ligand Mining

Priority order for bioactivity data:

1. ChEMBL_get_target_activities - curated, SAR-ready
2. BindingDB_get_ligands_by_uniprot - direct Ki/Kd with literature links
3. GtoPdb_search_ligands - pharmacology focus (GPCRs, channels)
4. PubChem_search_assays_by_target_gene - HTS screens, novel scaffolds
5. OpenTargets_get_chemical_probes_by_target_ensemblID - validated probes

Key steps:

1. Filter to IC50/Ki/Kd < 10 uM; retrieve molecule details for top actives
2. Identify chemical probes and approved drugs
3. Analyze SAR: common scaffolds, key modifications
4. Check off-target selectivity: BindingDB_get_targets_by_compound

---

Phase 3: Structure Analysis

Tools:

• PDB_search_similar_structures(query=uniprot, type="sequence") - find PDB entries
• get_protein_metadata_by_pdb_id(pdb_id) - resolution, method
• get_binding_affinity_by_pdb_id(pdb_id) - co-crystal ligand affinities
• get_ligand_smiles_by_chem_comp_id(chem_comp_id) - ligand SMILES from PDB
• EMDB_search_structures(query) - cryo-EM structures (prefer for GPCRs, ion channels)
• alphafold_get_prediction(qualifier) - AlphaFold DB fallback

Phase 3.5: Docking Validation (NVIDIA NIM)

If 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.

---

Phase 4: Compound Expansion

4.1-4.3 Search-Based Expansion

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)

4.4 De Novo Generation (NVIDIA NIM)

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")

---

Phase 5: ADMET Filtering

Apply sequentially (all tools accept smiles=[list]):

1. Physicochemical: ADMETAI_predict_physicochemical_properties - Lipinski violations <= 1, QED > 0.3, MW 200-600
2. Bioavailability: ADMETAI_predict_bioavailability - oral bioavailability > 0.3
3. Toxicity: ADMETAI_predict_toxicity - AMES < 0.5, hERG < 0.5, DILI < 0.5
4. CYP: ADMETAI_predict_CYP_interactions - flag CYP3A4 inhibitors
5. Alerts: ChEMBL_search_compound_structural_alerts - no PAINS

Include a filter funnel summary in the report showing pass/fail counts at each stage.

---

Phase 6: Candidate Docking & Prioritization

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:

• T1 (3 stars): Experimental IC50/Ki < 100 nM
• T2 (2 stars): Docking within 5% of reference OR IC50 100-1000 nM
• T3 (1 star): >80% similarity to T1 compound
• T4 (0 stars): 70-80% similarity, scaffold match only
• T5 (no stars): Generated molecule, ADMET-passed, no docking

Deliver top 20 candidates with: Rank, ID, SMILES, docking score, ADMET score, overall score, source, evidence tier.

---

Phase 6.5: Literature Evidence

• PubMed_search_articles(query="[TARGET] inhibitor SAR") - peer-reviewed
• EuropePMC_search_articles(query, source="PPR") - preprints (not peer-reviewed)
• openalex_search_works(query) - citation analysis

---

Fallback Chains

Target ID:     ChEMBL_search_targets -> GtoPdb_search_targets -> "Not in databases"
Druggability:  OpenTargets tractability -> DGIdb druggability -> target class proxy
Bioactivity:   ChEMBL -> BindingDB -> GtoPdb -> PubChem BioAssay -> "No data"
Structure:     PDB -> EMDB (membrane) -> alphafold_get_prediction -> 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_search_targets

---

Programmatic Access (Beyond Tools)

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.

---

NVIDIA NIM Runtime Notes

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.

---

Reference Files

• WORKFLOW_DETAILS.md - Phase-by-phase procedures, code patterns, screening protocols
• TOOLS_REFERENCE.md - Complete tool reference with parameters and fallback chains
• REPORT_TEMPLATE.md - Report file template and evidence grading system
• EXAMPLES.md - End-to-end workflow examples (EGFR, novel target, lead optimization)
• CHECKLIST.md - Pre-delivery verification checklist