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

Small Molecule Binder Discovery Checklist

Pre-delivery verification checklist for binder discovery reports.

Report Quality Checklist

Structure & Format

• [ ] Report file created with correct naming: [TARGET]_binder_discovery_report.md
• [ ] All 9 main sections present
• [ ] Executive summary completed (not [Researching...])
• [ ] Data sources section populated

Phase 1: Target Validation

• [ ] UniProt accession documented
• [ ] ChEMBL target ID obtained (or "Not in ChEMBL" noted)
• [ ] Ensembl gene ID resolved
• [ ] Druggability assessed from ≥2 sources
• [ ] Druggability scorecard included
• [ ] Binding site information present (or "No structural data available")
• [ ] Target class identified (kinase, GPCR, enzyme, etc.)

Phase 1.4: Structure Prediction (NVIDIA NIM)

• [ ] NVIDIA_API_KEY availability checked and documented
• [ ] Structure prediction method stated (AlphaFold2/ESMFold/PDB)
• [ ] pLDDT confidence scores reported (if NvidiaNIM_alphafold2 used)
• [ ] Confidence breakdown table included (Very High/Confident/Low/Very Low)
• [ ] Key binding residue pLDDT values documented
• [ ] Mean pLDDT interpretation provided
• [ ] Structure source attributed (NvidiaNIM_alphafold2, NvidiaNIM_esmfold, or PDB)

Phase 2: Known Ligands

• [ ] ChEMBL bioactivity data queried
• [ ] Activity statistics table included (count, potency distribution)
• [ ] Top 10 most potent compounds listed with IC50/Ki values
• [ ] Chemical probes identified (or "None available" stated)
• [ ] SAR insights summarized (scaffolds, key modifications)
• [ ] Approved drugs for target listed (or "None approved")

Phase 3: Structure

• [ ] PDB structures listed with resolution (or "No experimental structure")
• [ ] Best structure for docking identified and justified
• [ ] Binding pocket described (residues, volume, character)
• [ ] Key interaction points documented
• [ ] AlphaFold structure noted if no experimental

Phase 3.5: Docking Validation (NVIDIA NIM)

• [ ] Reference compound selected for docking validation
• [ ] Reference compound docked (NvidiaNIM_diffdock or NvidiaNIM_boltz2)
• [ ] Docking method documented (DiffDock vs Boltz2)
• [ ] Best pose confidence score reported
• [ ] Steric clash assessment included
• [ ] Validation status stated (✓ binding pocket captured or ✗ issues)
• [ ] Docking tool source attributed

Phase 4: Compound Expansion

• [ ] ≥3 diverse seed compounds used for similarity search
• [ ] Similarity threshold documented (typically 70-85%)
• [ ] Total compounds from similarity search reported
• [ ] Substructure search completed (if scaffolds identified)
• [ ] Cross-database mining results included
• [ ] Deduplication performed and counts reported
• [ ] Source attribution for each search

Phase 4.4: De Novo Generation (NVIDIA NIM) - Optional

• [ ] Seed scaffolds identified from top actives
• [ ] Masked SMILES designed (if NvidiaNIM_genmol used)
• [ ] Mask positions and purpose documented
• [ ] Number of molecules generated reported
• [ ] Lipinski pass rate documented
• [ ] Mean QED score reported
• [ ] Unique scaffolds count noted
• [ ] Top generated compounds listed with SMILES, QED, LogP
• [ ] Generation tool attributed (NvidiaNIM_genmol or NvidiaNIM_molmim)
• [ ] Generated compounds passed to ADMET filtering

Phase 5: ADMET Filtering

• [ ] Physicochemical filters applied (Lipinski, QED)
• [ ] Filter thresholds documented
• [ ] Bioavailability predictions included
• [ ] Toxicity predictions included (AMES, hERG, DILI minimum)
• [ ] CYP interaction flags noted
• [ ] Structural alert check completed
• [ ] Filter funnel table with pass/fail counts at each stage
• [ ] Common failure reasons documented

Phase 6: Candidate Docking & Prioritization

• [ ] All candidates docked (NvidiaNIM_diffdock or NvidiaNIM_boltz2)
• [ ] Docking scores compared to reference compound
• [ ] Scoring methodology explained (docking + ADMET + novelty)
• [ ] All scoring dimensions documented with weights
• [ ] ≥20 candidates ranked (or all available if fewer)
• [ ] Top 20 table includes: ID, SMILES, Docking vs Ref, ADMET score, overall score
• [ ] Synthesis feasibility assessed (SA score or commercial availability)
• [ ] Scaffold diversity noted (number of distinct scaffolds)
• [ ] Evidence tier assigned to each candidate (T0-T5 scale)

Phase 7: Recommendations

• [ ] ≥3 immediate actions listed
• [ ] Experimental validation plan outlined
• [ ] Backup strategies identified
• [ ] Timeline/priority suggestions included

Phase 8: Data Gaps

• [ ] All data gaps aggregated in one section
• [ ] Reason for each gap documented
• [ ] Alternative approaches suggested for gaps

Phase 10: Methods Summary

• [ ] Methods summary table included
• [ ] All tools used listed with purpose
• [ ] Key tools documented:
• [ ] Sequence retrieval (UniProt_search)
• [ ] Structure prediction (NvidiaNIM_alphafold2/esmfold if used)
• [ ] Docking validation (NvidiaNIM_diffdock/boltz2 if used)
• [ ] Known ligands (ChEMBL_get_target_activities)
• [ ] Similarity search (ChEMBL_search_similar_molecules)
• [ ] De novo generation (NvidiaNIM_genmol/molmim if used)
• [ ] ADMET filtering (ADMETAI_predict_*)
• [ ] Candidate docking (NvidiaNIM_diffdock/boltz2 if used)

Data Output Files

• [ ] [TARGET]_candidate_compounds.csv created with prioritized list
• [ ] CSV includes: Rank, ID, SMILES, Similarity, ADMET_Score, Overall_Score, Source
• [ ] All candidates have valid SMILES

Citation Requirements

Every Section Must Include

• [ ] Source attribution for data (database name + tool used)
• [ ] Specific identifiers used (ChEMBL ID, PDB ID, etc.)
• [ ] Date of data retrieval (for dynamic databases)

Format Examples

*Source: ChEMBL via `ChEMBL_get_target_activities` (CHEMBL203)*
*Source: PDB via `get_protein_metadata_by_pdb_id` (1M17)*
*Source: ADMET-AI via `ADMETAI_predict_toxicity`*
*Source: NVIDIA NIM via `NvidiaNIM_alphafold2` (pLDDT: 90.94)*
*Source: NVIDIA NIM via `NvidiaNIM_diffdock` (confidence: 0.906)*
*Source: NVIDIA NIM via `NvidiaNIM_genmol` (100 molecules)*

Evidence Grading

All Candidates Must Have

• [ ] Evidence tier assigned (★★★★ to ○○○)
• [ ] Rationale for tier documented
• [ ] Docking comparison to reference (if docking performed)

Tier Definitions Applied

Docking-Enhanced Grading

• [ ] If docking performed, apply upgrade/downgrade rules
• [ ] Docking > reference → T0 (★★★★)
• [ ] Docking within 5% → T2 (★★☆)
• [ ] Docking >20% worse → downgrade one tier

Quantified Minimums

Final Review

Before Delivery

• [ ] No [Researching...] placeholders remaining
• [ ] All tables properly formatted
• [ ] No empty sections (use "Not available" with explanation)
• [ ] Executive summary synthesizes key findings
• [ ] Recommendations are actionable and specific
• [ ] Report is self-contained (user doesn't need external context)

Common Issues to Avoid

• [ ] Missing evidence tiers on candidates
• [ ] Unattributed data (no source citation)
• [ ] Empty ADMET predictions (tool may have failed silently)
• [ ] Inconsistent compound identifiers
• [ ] Missing SMILES in candidate list
• [ ] Recommendations without supporting data

Small Molecule Binder Discovery Examples

Detailed workflow examples for common discovery scenarios.

Example 1: Well-Characterized Target (EGFR)

User Query: "Find novel small molecule binders for EGFR"

Phase 1: Target Validation

# Step 1.1: Resolve identifiers
uniprot_result = tu.tools.UniProt_search(query="EGFR human", organism="human", limit=5)
# → P00533 (EGFR_HUMAN)

gene_result = tu.tools.MyGene_query_genes(q="EGFR", species="human")
# → ENSG00000146648, NCBI: 1956

chembl_result = tu.tools.ChEMBL_search_targets(query="EGFR", organism="Homo sapiens")
# → CHEMBL203

# Step 1.2: Assess druggability
tractability = tu.tools.OpenTargets_get_target_tractability_by_ensemblID(
    ensemblId="ENSG00000146648"
)
# → Small molecule: bucket 1 (approved drugs exist)

druggability = tu.tools.DGIdb_get_gene_druggability(genes=["EGFR"])
# → Categories: Kinase, Clinically Actionable, 45 drugs

Report Section 1.2:

### 1.2 Druggability Assessment

| Factor | Assessment | Score | Source |
|--------|------------|-------|--------|
| Target class | Receptor tyrosine kinase | ★★★ | ChEMBL |
| Tractability bucket | 1 (approved drugs) | ★★★ | Open Targets |
| Known drugs | 15 approved | ★★★ | DGIdb |
| Binding site | ATP pocket (well-characterized) | ★★★ | PDB |

**Overall Druggability**: ★★★ Highly druggable

*Sources: Open Targets, DGIdb, ChEMBL*

Phase 2: Known Ligand Mining

# Step 2.1: Get all bioactivity data
activities = tu.tools.ChEMBL_get_target_activities(
    target_chembl_id="CHEMBL203",
    limit=500
)
# → 4,847 activity records

# Step 2.2: Filter to potent compounds
potent = [a for a in activities['activities'] 
          if a['standard_type'] in ['IC50', 'Ki'] 
          and a['standard_value'] and float(a['standard_value']) < 100]
# → 312 compounds with IC50/Ki < 100 nM

# Step 2.3: Get molecule details for top actives
top_actives = []
for activity in potent[:20]:
    mol = tu.tools.ChEMBL_get_molecule(
        molecule_chembl_id=activity['molecule_chembl_id']
    )
    top_actives.append({
        'chembl_id': activity['molecule_chembl_id'],
        'ic50': activity['standard_value'],
        'smiles': mol['molecule_structures']['canonical_smiles'],
        'max_phase': mol['max_phase']
    })

# Step 2.4: Get chemical probes
probes = tu.tools.OpenTargets_get_chemical_probes_by_target_ensemblID(
    ensemblId="ENSG00000146648"
)

Report Section 2:

### 2.1 ChEMBL Bioactivity Summary

**Total Activity Records**: 4,847
- IC50: 2,345 | Ki: 1,234 | Kd: 456 | EC50: 812

**Potency Distribution**:
| Potency Range | Count | % |
|---------------|-------|---|
| < 10 nM | 89 | 1.8% |
| 10-100 nM | 223 | 4.6% |
| 100-1000 nM | 567 | 11.7% |
| > 1000 nM | 3,968 | 81.9% |

### 2.2 Top 10 Most Potent Compounds

| Rank | Compound | ChEMBL ID | IC50 (nM) | Phase | Scaffold |
|------|----------|-----------|-----------|-------|----------|
| 1 | Osimertinib | CHEMBL3353410 | 0.5 | 4 | Pyrimidine-amine |
| 2 | Dacomitinib | CHEMBL2110732 | 0.8 | 4 | Quinazoline |
| 3 | Afatinib | CHEMBL1173655 | 1.2 | 4 | Quinazoline |
| 4 | Erlotinib | CHEMBL553 | 2.0 | 4 | Quinazoline |
| 5 | Gefitinib | CHEMBL939 | 5.0 | 4 | Quinazoline |
| ... | ... | ... | ... | ... | ... |

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

Phase 3: Structure Analysis

# Step 3.1: Find PDB structures
pdb_list = tu.tools.PDB_search_similar_structures(
    query="P00533",
    type="sequence"
)
# → 156 structures

# Step 3.2: Get metadata for key structures
key_pdbs = ["1M17", "4HJO", "5UG9", "6JX4"]
for pdb_id in key_pdbs:
    metadata = tu.tools.get_protein_metadata_by_pdb_id(pdb_id=pdb_id)
    affinity = tu.tools.get_binding_affinity_by_pdb_id(pdb_id=pdb_id)

Report Section 3:

### 3.1 Available Structures

**Total PDB Structures**: 156
- With small molecule ligand: 89
- Apo structures: 34
- With peptide/protein: 33

| PDB ID | Resolution | Ligand | Affinity | Conformation |
|--------|------------|--------|----------|--------------|
| 1M17 | 2.6 Å | Erlotinib | Ki=0.4 nM | Active |
| 4HJO | 2.1 Å | Lapatinib | Ki=3 nM | Inactive |
| 5UG9 | 1.9 Å | Osimertinib | Ki=0.5 nM | Active/C797S |
| 6JX4 | 2.2 Å | Compound X | IC50=12 nM | Active |

**Best Structure for Novel Binder Design**: 5UG9 (highest resolution, relevant mutation)

Phase 4: Compound Expansion

# Step 4.1: Select diverse seeds
seeds = [
    ("CHEMBL553", "COc1cc2ncnc(Nc3ccc(C#C)c(c3)C#C)c2cc1OCCOCCOC"),  # Erlotinib scaffold
    ("CHEMBL3353410", "COc1cc(N2CCC(CC2)n2cc(nc2c2ccc(NC(=O)C=C)c(Nc3nc(ccn3)c3cccnc3)c2)C)ccc1NC(=O)C=C"),  # Osimertinib
    ("CHEMBL1779", "Fc1ccc(Nc2ncnc3cc(OCCCN4CCOCC4)c(OC)cc23)cc1Cl"),  # Gefitinib-like
]

# Step 4.2: Similarity search
all_similar = []
for seed_id, seed_smiles in seeds:
    similar = tu.tools.ChEMBL_search_similar_molecules(
        molecule=seed_smiles,
        similarity=75
    )
    all_similar.extend(similar['molecules'])
# → 892 similar compounds

# Step 4.3: PubChem expansion
pubchem_similar = tu.tools.PubChem_search_compounds_by_similarity(
    smiles=seeds[0][1],
    threshold=0.75
)
# → 456 additional compounds

# Step 4.4: Substructure search for quinazoline
substructure = tu.tools.ChEMBL_search_substructure(
    smiles="c1ccc2ncncc2c1"  # Quinazoline core
)
# → 234 quinazoline-containing compounds

# Step 4.5: Deduplicate
unique_candidates = deduplicate_by_smiles(all_similar + pubchem_similar + substructure)
# → 1,247 unique candidates

Phase 5: ADMET Filtering

# Step 5.1: Batch ADMET prediction
smiles_list = [c['smiles'] for c in unique_candidates]

# Physicochemical
physchem = tu.tools.ADMETAI_predict_physicochemical_properties(smiles=smiles_list)
# Filter: Lipinski ≤ 1, QED > 0.3
passed_physchem = [c for c in physchem if c['Lipinski_violations'] <= 1 and c['QED'] > 0.3]
# → 987 pass

# Bioavailability
bioavail = tu.tools.ADMETAI_predict_bioavailability(smiles=[c['smiles'] for c in passed_physchem])
# Filter: Oral bioavailability > 0.3
passed_bioavail = [c for c in bioavail if c['Bioavailability_Ma'] > 0.3]
# → 834 pass

# Toxicity
toxicity = tu.tools.ADMETAI_predict_toxicity(smiles=[c['smiles'] for c in passed_bioavail])
# Filter: AMES < 0.5, hERG < 0.5
passed_tox = [c for c in toxicity if c['AMES'] < 0.5 and c['hERG'] < 0.5]
# → 723 pass

# Structural alerts
final_candidates = []
for c in passed_tox:
    alerts = tu.tools.ChEMBL_search_compound_structural_alerts(smiles=c['smiles'])
    if not alerts.get('pains', []):
        final_candidates.append(c)
# → 678 pass all filters

Report Section 5:

### 5.4 Filter Summary

| Filter Stage | Input | Passed | Failed | Pass Rate |
|--------------|-------|--------|--------|-----------|
| Initial candidates | 1,247 | - | - | - |
| Physicochemical | 1,247 | 987 | 260 | 79% |
| Drug-likeness (QED) | 987 | 892 | 95 | 90% |
| Bioavailability | 892 | 834 | 58 | 93% |
| Toxicity (AMES/hERG) | 834 | 723 | 111 | 87% |
| Structural alerts | 723 | 678 | 45 | 94% |
| **Final Candidates** | **1,247** | **678** | **569** | **54%** |

**Top Failure Reasons**:
1. MW > 600 Da: 156 compounds (12.5%)
2. hERG liability (>0.5): 78 compounds (6.3%)
3. Low bioavailability (<0.3): 58 compounds (4.7%)
4. PAINS alerts: 32 compounds (2.6%)

Phase 6: Final Prioritization

Report Section 6.3:

### 6.3 Top 20 Candidate Compounds

| Rank | ID | Evidence | Sim. | ADMET | Novelty | Score | Rationale |
|------|-----|----------|------|-------|---------|-------|-----------|
| 1 | CPD-001 | ★★★ | 0.87 | 4.6 | Novel R-group | 4.5 | 87% to osimertinib, clean ADMET |
| 2 | CPD-002 | ★★☆ | 0.82 | 4.4 | Untested | 4.3 | Quinazoline analog, available |
| 3 | CPD-003 | ★★☆ | 0.79 | 4.5 | Novel core | 4.2 | Pyridine replacement |
| 4 | CPD-004 | ★★☆ | 0.81 | 4.2 | Untested | 4.1 | Improved solubility |
| 5 | CPD-005 | ★☆☆ | 0.76 | 4.3 | Novel scaffold | 4.0 | New chemotype |
| ... | ... | ... | ... | ... | ... | ... | ... |

**Scaffold Diversity**: 8 distinct scaffolds in top 20
**Commercial Availability**: 14/20 available (Enamine, MolPort)
**Estimated Hit Rate**: 20-35% based on similarity to approved drugs

*Scoring: Evidence (25%) + Similarity (25%) + ADMET (25%) + Novelty (15%) + Availability (10%)*

---

Example 2: Novel Target (Limited Data)

User Query: "Find small molecules for SLC7A11 (xCT transporter)"

Key Differences for Novel Targets

1. Phase 2 will yield limited data - Few or no known ligands 2. Rely more on structure-based - AlphaFold if no experimental structure 3. Consider similar targets - Related transporters may have ligands 4. Lower evidence tiers - Most candidates will be ★☆☆ or ☆☆☆

Modified Workflow

# Phase 1: Target validation
# SLC7A11: P60880 (UniProt), ENSG00000151012, limited ChEMBL data

# Phase 2: Known ligands - limited
activities = tu.tools.ChEMBL_get_target_activities(target_chembl_id="CHEMBL4523582")
# → Only 23 activity records, best IC50 = 2.5 µM

# Phase 2b: Check related targets (SLC7 family)
related_targets = ["SLC7A1", "SLC7A5", "SLC7A8"]  # Similar transporters
for target in related_targets:
    activities = tu.tools.ChEMBL_search_targets(query=target)
    # Look for ligands that might cross-react

# Phase 3: Structure - AlphaFold only
structure = tu.tools.alphafold_get_prediction(accession="P60880")
# → AF-P60880-F1 (predicted, pLDDT varies)

# Phase 4: Expand from weak actives AND related target ligands
# Use lower similarity threshold (60-70%) to find more chemotypes

# Phase 5-6: Same ADMET filtering and prioritization

Report Note for Novel Targets:

### Data Gaps & Limitations

| Gap | Impact | Mitigation |
|-----|--------|------------|
| Limited bioactivity data (23 records) | Low confidence in actives | Used related SLC7 family ligands |
| No experimental structure | Binding site uncertain | AlphaFold structure (moderate confidence) |
| Best known IC50 = 2.5 µM | Weak starting point | Prioritize diverse chemotypes |

**Recommended Strategy**: Broader screening campaign with multiple scaffolds
**Estimated Hit Rate**: 5-10% (lower confidence than well-characterized targets)

---

Example 3: Lead Optimization

User Query: "Find analogs of compound X (CHEMBL12345) for target Y with improved ADMET"

Focus on Narrow Expansion

# Step 1: Get reference compound
ref_compound = tu.tools.ChEMBL_get_molecule(molecule_chembl_id="CHEMBL12345")
ref_smiles = ref_compound['molecule_structures']['canonical_smiles']
ref_admet = tu.tools.ADMETAI_predict_toxicity(smiles=[ref_smiles])
# Known issue: hERG = 0.72 (liability)

# Step 2: Tight similarity search (85-95%)
similar = tu.tools.ChEMBL_search_similar_molecules(molecule=ref_smiles, similarity=85)
# → 156 close analogs

# Step 3: Predict ADMET for all
all_smiles = [m['smiles'] for m in similar['molecules']]
admet_results = tu.tools.ADMETAI_predict_toxicity(smiles=all_smiles)

# Step 4: Filter for improved hERG
improved = [m for m, a in zip(similar['molecules'], admet_results) 
            if a['hERG'] < 0.5]  # Improved from 0.72
# → 34 analogs with improved hERG

# Step 5: Check if activity maintained
# Query ChEMBL for any existing bioactivity on these compounds

Report Focus:

### Lead Optimization Summary

**Reference Compound**: CHEMBL12345
- IC50 = 45 nM (target Y)
- hERG liability = 0.72 (CONCERN)
- CYP3A4 inhibitor = 0.65 (moderate)

**Optimization Goal**: Reduce hERG while maintaining potency

### Improved Analogs

| Analog | Sim. | hERG | CYP3A4 | Modification | Activity Status |
|--------|------|------|--------|--------------|-----------------|
| CPD-A | 0.91 | 0.34 | 0.42 | N-methyl→N-H | Untested |
| CPD-B | 0.88 | 0.28 | 0.51 | Remove basic N | Untested |
| CPD-C | 0.92 | 0.41 | 0.38 | Add hydroxyl | IC50=67 nM (similar) |

**Recommended**: CPD-C (proven activity retention, improved hERG)

---

Example 4: NVIDIA NIM-Enhanced Workflow (CDK4 Kinase)

User Query: "Discover novel CDK4 inhibitors using all available structure prediction and docking"

Check NVIDIA NIM Availability

import os
nvidia_available = bool(os.environ.get("NVIDIA_API_KEY"))
print(f"NVIDIA NIM tools: {'Available' if nvidia_available else 'Unavailable'}")

Phase 1: Target Validation + Structure Prediction

# Step 1.1: Standard identifier resolution
uniprot_result = tu.tools.UniProt_search(query="CDK4 human", organism="human", limit=5)
# → P11802 (CDK4_HUMAN)

gene_result = tu.tools.MyGene_query_genes(q="CDK4", species="human")
# → ENSG00000135446

chembl_result = tu.tools.ChEMBL_search_targets(query="CDK4", organism="Homo sapiens")
# → CHEMBL3116

# Step 1.2: Get protein sequence
uniprot_details = tu.tools.UniProt_get_entry(accession="P11802")
cdk4_sequence = uniprot_details['sequence']

# Step 1.3: Check for existing PDB structures
pdb_structures = tu.tools.PDB_search_similar_structures(query="P11802", type="sequence")
# → 25 structures, best: 2W9Z (2.0 Å, with palbociclib)

# Step 1.4: Predict structure with NVIDIA NIM for comparison
if nvidia_available:
    predicted_structure = tu.tools.NvidiaNIM_alphafold2(
        sequence=cdk4_sequence,
        algorithm="mmseqs2",
        relax_prediction=False
    )
    # → PDB content with pLDDT scores
    # Mean pLDDT: 88.5 (high confidence)

Report Section 1.4:

### 1.4 Structure Prediction Quality

**Method**: AlphaFold2 via NVIDIA NIM
**Mean pLDDT**: 88.5 (high confidence)

| Confidence Level | Range | Fraction | Interpretation |
|------------------|-------|----------|----------------|
| Very High | ≥90 | 65.2% | Highly reliable |
| Confident | 70-90 | 28.1% | Reliable |
| Low | 50-70 | 5.8% | Use caution |
| Very Low | <50 | 0.9% | Unreliable |

**Key Binding Residue Confidence** (ATP pocket):
| Residue | Function | pLDDT |
|---------|----------|-------|
| K35 | ATP binding | 92.4 |
| E51 | Salt bridge | 89.1 |
| D99 | Catalytic | 94.2 |
| D158 | DFG motif | 91.8 |

**Recommendation**: Use experimental structure 2W9Z (2.0 Å) for docking,
predicted structure validates binding pocket geometry.

*Source: NVIDIA NIM via `NvidiaNIM_alphafold2`*

Phase 3.5: Docking Validation

# Step 3.5.1: Get reference compound (palbociclib) for validation
ref_compound = tu.tools.ChEMBL_get_molecule(molecule_chembl_id="CHEMBL1906")
palbociclib_smiles = ref_compound['molecule_structures']['canonical_smiles']

# Step 3.5.2: Get PDB structure content
pdb_content = tu.tools.get_pdb_structure_file(pdb_id="2W9Z", format="pdb")

# Step 3.5.3: Dock reference compound to validate binding pocket
if nvidia_available:
    # Option A: DiffDock (with PDB + SDF)
    validation_result = tu.tools.NvidiaNIM_diffdock(
        protein=pdb_content,
        ligand=palbociclib_sdf,  # SDF content
        num_poses=10
    )
    reference_confidence = validation_result['poses'][0]['confidence']
    # → 0.92 (excellent)
    
    # Option B: Boltz2 (from sequence + SMILES)
    boltz_result = tu.tools.NvidiaNIM_boltz2(
        polymers=[{"molecule_type": "protein", "sequence": cdk4_sequence}],
        ligands=[{"smiles": palbociclib_smiles}],
        sampling_steps=50,
        diffusion_samples=1
    )
    # → Complex structure with pTM=0.85, ipTM=0.78

Report Section 3.5:

### 3.5 Docking Validation Results

**Reference Compound**: Palbociclib (CHEMBL1906)
**Known IC50**: 11 nM

**DiffDock Validation**:
| Metric | Value | Interpretation |
|--------|-------|----------------|
| Best Pose Confidence | 0.92 | Excellent |
| Poses in ATP pocket | 10/10 | Correct binding mode |
| RMSD to crystal pose | 1.2 Å | Good agreement |

**Boltz2 Validation**:
| Metric | Value | Interpretation |
|--------|-------|----------------|
| pTM | 0.85 | High structure confidence |
| ipTM | 0.78 | Good interface prediction |
| Aggregate Score | 0.81 | Reliable complex |

**Validation Status**: ✓ Binding pocket correctly captured by both methods

*Source: NVIDIA NIM via `NvidiaNIM_diffdock`, `NvidiaNIM_boltz2`*

Phase 4.4: De Novo Molecule Generation

# Step 4.4.1: Identify seed scaffolds from top actives
# Palbociclib scaffold: pyrido[2,3-d]pyrimidin-7-one core
seed_smiles = "CC(=O)c1c(C)c2cnc(Nc3ccc(N4CCNCC4)cn3)nc2n(C2CCCC2)c1=O"

# Step 4.4.2: Design masked SMILES for scaffold hopping
# Vary the piperazine tail and add positions on pyridine
masked_smiles = "CC(=O)c1c(C)c2cnc(Nc3ccc([*{4-10}])c([*{1-3}])n3)nc2n(C2CCCC2)c1=O"

# Step 4.4.3: Generate with GenMol
if nvidia_available:
    genmol_result = tu.tools.NvidiaNIM_genmol(
        smiles=masked_smiles,
        num_molecules=100,
        temperature=2.0,
        scoring="QED"
    )
    # → 100 generated molecules with QED scores
    generated_mols = genmol_result['molecules']
    
    # Filter by basic criteria
    good_qed = [m for m in generated_mols if m['QED'] > 0.5]
    # → 78 pass QED filter

# Step 4.4.4: Alternative with MolMIM (close analogs)
if nvidia_available:
    molmim_result = tu.tools.NvidiaNIM_molmim(
        smi=seed_smiles,
        num_molecules=50,
        algorithm="CMA-ES"
    )
    # → 50 optimized analogs

Report Section 4.4:

### 4.4 De Novo Generation Results

**Seed Scaffold**: Pyrido[2,3-d]pyrimidin-7-one (from palbociclib)
**Method**: GenMol via NVIDIA NIM

**Masked Positions**:
- Position 1: Piperazine tail region `[*{4-10}]`
- Position 2: Pyridine substituent `[*{1-3}]`

| Metric | GenMol | MolMIM |
|--------|--------|--------|
| Molecules Generated | 100 | 50 |
| Passing QED > 0.5 | 78 (78%) | 42 (84%) |
| Mean QED Score | 0.62 | 0.68 |
| Unique Scaffolds | 8 | 3 |
| Mean LogP | 3.8 | 4.1 |

**Top Generated Compounds (GenMol)**:
| ID | SMILES (abbreviated) | QED | LogP | Modification |
|----|---------------------|-----|------|--------------|
| GEN-001 | ...Nc3ccc(N4CCN(C)CC4)c(F)n3... | 0.74 | 3.2 | Methyl-piperazine + F |
| GEN-002 | ...Nc3ccc(N4CCOCC4)c(Cl)n3... | 0.71 | 3.5 | Morpholine + Cl |
| GEN-003 | ...Nc3ccc(NC4CCNCC4)cn3... | 0.69 | 2.9 | Piperidine amine |

**Next Step**: Pass all 120 generated molecules to Phase 5 (ADMET filtering)

*Source: NVIDIA NIM via `NvidiaNIM_genmol`, `NvidiaNIM_molmim`*

Phase 6: Candidate Docking

# After ADMET filtering (Phase 5), dock all candidates
admet_passed = 67  # Compounds passing all ADMET filters

if nvidia_available:
    # Get reference score
    reference_confidence = 0.92  # From Phase 3.5
    
    # Dock all candidates
    docking_results = []
    for compound in admet_passed_compounds:
        result = tu.tools.NvidiaNIM_diffdock(
            protein=pdb_content,
            ligand=compound['sdf'],
            num_poses=5
        )
        
        best_confidence = result['poses'][0]['confidence']
        vs_reference = (best_confidence / reference_confidence - 1) * 100
        
        docking_results.append({
            'id': compound['id'],
            'smiles': compound['smiles'],
            'confidence': best_confidence,
            'vs_reference_pct': vs_reference,
            'evidence_tier': assign_tier(best_confidence, reference_confidence)
        })
    
    # Sort by confidence
    ranked = sorted(docking_results, key=lambda x: x['confidence'], reverse=True)

Report Section 6 (Final):

### 6. Final Candidate Ranking

**Scoring Method**: Docking (40%) + ADMET (30%) + Similarity (20%) + Novelty (10%)
**Reference**: Palbociclib (confidence: 0.92)

| Rank | ID | Evidence | Docking | vs Ref | ADMET | Score | Source |
|------|-----|----------|---------|--------|-------|-------|--------|
| 1 | GEN-001 | ★★★★ | 0.95 | +3.3% | 4.5 | 4.7 | GenMol |
| 2 | CPD-042 | ★★★ | 0.91 | -1.1% | 4.6 | 4.5 | ChEMBL (IC50=18nM) |
| 3 | GEN-002 | ★★★★ | 0.94 | +2.2% | 4.2 | 4.4 | GenMol |
| 4 | CPD-108 | ★★☆ | 0.89 | -3.3% | 4.4 | 4.2 | PubChem |
| 5 | GEN-015 | ★★☆ | 0.90 | -2.2% | 4.3 | 4.1 | MolMIM |

**Key Findings**:
- 2 de novo generated compounds outperform reference (★★★★)
- 3/5 top candidates from AI generation
- GEN-001 shows novel modification with improved docking score

**Methods Summary**:
| Step | Tool | Purpose |
|------|------|---------|
| Sequence retrieval | UniProt_search | Get CDK4 sequence |
| Structure prediction | NvidiaNIM_alphafold2 | Validate binding pocket |
| Docking validation | NvidiaNIM_diffdock | Confirm pose accuracy |
| Known ligands | ChEMBL_get_target_activities | Mining actives |
| De novo generation | NvidiaNIM_genmol | Novel scaffold exploration |
| ADMET filtering | ADMETAI_predict_* | Drug-likeness |
| Candidate docking | NvidiaNIM_diffdock | Final scoring |

*Full methods and parameters available in appendix*

---

Common Pitfalls and Solutions

Pitfall 1: Silent Tool Failures

Problem: ChEMBL returns empty for wrong parameter name

# WRONG - returns empty
tu.tools.ChEMBL_get_target_activities(chembl_target_id="CHEMBL203")

# CORRECT
tu.tools.ChEMBL_get_target_activities(target_chembl_id="CHEMBL203")

Solution: Always verify tool parameters before first use.

Pitfall 2: ADMET Batch Size

Problem: Sending 1000+ SMILES crashes or times out

# WRONG - too many at once
tu.tools.ADMETAI_predict_toxicity(smiles=all_1000_smiles)

# CORRECT - batch in chunks
for i in range(0, len(all_smiles), 100):
    batch = all_smiles[i:i+100]
    results = tu.tools.ADMETAI_predict_toxicity(smiles=batch)
    process_results(results)

Pitfall 3: Missing Deduplication

Problem: Same compound from multiple sources counted multiple times

# Track by canonical SMILES or InChIKey
seen_inchikeys = set()
unique_candidates = []
for compound in all_candidates:
    inchikey = get_inchikey(compound['smiles'])  # Or use canonical SMILES
    if inchikey not in seen_inchikeys:
        seen_inchikeys.add(inchikey)
        unique_candidates.append(compound)

Pitfall 4: Not Checking Existing Data

Problem: Recommending "novel" compounds that are already tested

# Check if compound already has bioactivity for target
for candidate in candidates:
    existing = tu.tools.ChEMBL_search_activities(
        molecule_chembl_id=candidate['chembl_id'],
        target_chembl_id=target_id
    )
    if existing['activities']:
        candidate['status'] = 'ALREADY_TESTED'
        candidate['known_ic50'] = existing['activities'][0]['standard_value']
    else:
        candidate['status'] = 'NOVEL'

Pitfall 5: NVIDIA NIM API Key Not Set

Problem: Tools fail silently or return errors

# WRONG - No fallback when NIM unavailable
result = tu.tools.NvidiaNIM_alphafold2(sequence=seq)
# → Error: NVIDIA_API_KEY not set

# CORRECT - Check availability first
import os
nvidia_available = bool(os.environ.get("NVIDIA_API_KEY"))

if nvidia_available:
    result = tu.tools.NvidiaNIM_alphafold2(sequence=seq)
else:
    print("NVIDIA NIM unavailable, using AlphaFold DB")
    result = tu.tools.alphafold_get_prediction(accession=uniprot_id)

Pitfall 6: GenMol Mask Syntax Errors

Problem: Invalid mask syntax causes generation failure

# WRONG - Missing mask or wrong syntax
tu.tools.NvidiaNIM_genmol(smiles="CCCc1ccc(O)cc1")  # No mask
tu.tools.NvidiaNIM_genmol(smiles="CCCc1ccc([*])cc1")  # No size range
tu.tools.NvidiaNIM_genmol(smiles="CCCc1ccc([*{1,3}])cc1")  # Wrong separator

# CORRECT - Proper mask syntax with dash separator
tu.tools.NvidiaNIM_genmol(
    smiles="CCCc1ccc([*{1-3}])cc1",  # [*{min-max}] format
    num_molecules=50,
    temperature=2.0
)

Pitfall 7: Boltz2 Polymer Format Errors

Problem: Wrong input format for protein-ligand complex

# WRONG - Direct sequence input
tu.tools.NvidiaNIM_boltz2(sequence="MKTAYIAK...", smiles="CCO")

# CORRECT - Use polymers and ligands lists
tu.tools.NvidiaNIM_boltz2(
    polymers=[{"molecule_type": "protein", "sequence": "MKTAYIAK..."}],
    ligands=[{"smiles": "CCO"}],
    sampling_steps=50,
    diffusion_samples=1
)

Pitfall 8: Over-relying on Docking Scores

Problem: Ranking by docking alone ignores ADMET and novelty

# WRONG - Rank by docking only
ranked = sorted(candidates, key=lambda x: x['docking_score'], reverse=True)

# CORRECT - Multi-factor scoring
def calculate_score(c):
    return (
        c['docking_score'] * 0.40 +      # Docking
        c['admet_score'] * 0.30 +         # ADMET
        c['similarity_score'] * 0.20 +    # Evidence
        c['novelty_score'] * 0.10         # Novelty
    )

ranked = sorted(candidates, key=calculate_score, reverse=True)

Pitfall 9: Not Reporting pLDDT for Predicted Structures

Problem: Using predicted structure without confidence assessment

# WRONG - Use structure without quality check
structure = tu.tools.NvidiaNIM_alphafold2(sequence=seq)
# Immediately use for docking...

# CORRECT - Parse and report confidence
structure = tu.tools.NvidiaNIM_alphafold2(sequence=seq)
plddt_scores = parse_plddt_from_pdb(structure['pdb_content'])
mean_plddt = sum(plddt_scores) / len(plddt_scores)

# Report in output
print(f"Mean pLDDT: {mean_plddt:.1f}")
if mean_plddt < 70:
    print("WARNING: Low confidence structure, binding site predictions may be unreliable")
    
# Check binding residue confidence specifically
binding_residues = [10, 15, 50, 80]  # Known binding site residues
for res in binding_residues:
    if plddt_scores[res-1] < 70:
        print(f"WARNING: Binding residue {res} has low confidence ({plddt_scores[res-1]:.1f})")

Report Template & Evidence Grading

Report template, evidence grading system, and report section formatting examples.

Report File Template

File: [TARGET]_binder_discovery_report.md

# Small Molecule Binder Discovery: [TARGET]

**Generated**: [Date] | **Query**: [Original query] | **Status**: In Progress

---

## Executive Summary
[Researching...]

---

## 1. Target Validation
### 1.1 Target Identifiers
[Researching...]
### 1.2 Druggability Assessment
[Researching...]
### 1.3 Binding Site Analysis
[Researching...]

---

## 2. Known Ligand Landscape
### 2.1 ChEMBL Bioactivity Summary
[Researching...]
### 2.2 Approved Drugs & Clinical Compounds
[Researching...]
### 2.3 Chemical Probes
[Researching...]
### 2.4 SAR Insights
[Researching...]

---

## 3. Structural Information
### 3.1 Available Structures
[Researching...]
### 3.2 Binding Pocket Analysis
[Researching...]
### 3.3 Key Interactions
[Researching...]

---

## 4. Compound Expansion
### 4.1 Similarity Search Results
[Researching...]
### 4.2 Substructure Search Results
[Researching...]
### 4.3 Cross-Database Mining
[Researching...]

---

## 5. ADMET Filtering
### 5.1 Physicochemical Filters
[Researching...]
### 5.2 ADMET Predictions
[Researching...]
### 5.3 Structural Alerts
[Researching...]
### 5.4 Filter Summary
[Researching...]

---

## 6. Candidate Prioritization
### 6.1 Scoring Methodology
[Researching...]
### 6.2 Synthesis Feasibility
[Researching...]
### 6.3 Top 20 Candidates
[Researching...]

---

## 7. Recommendations
### 7.1 Immediate Actions
[Researching...]
### 7.2 Experimental Validation Plan
[Researching...]
### 7.3 Backup Strategies
[Researching...]

---

## 8. Data Gaps & Limitations
[Researching...]

---

## 9. Data Sources
[Will be populated as research progresses...]

---

## 10. Methods Summary

| Step | Tool | Purpose |
|------|------|---------|
| Sequence retrieval | UniProt_search | Get protein sequence |
| Structure prediction | NvidiaNIM_alphafold2 / NvidiaNIM_esmfold | 3D structure with pLDDT |
| Docking validation | NvidiaNIM_diffdock / NvidiaNIM_boltz2 | Validate binding pocket |
| Known ligands | ChEMBL_get_target_activities | Bioactivity data |
| Similarity search | ChEMBL_search_similar_molecules | Expand chemical space |
| De novo generation | NvidiaNIM_genmol / NvidiaNIM_molmim | Novel molecule design |
| ADMET filtering | ADMETAI_predict_* | Drug-likeness assessment |
| Candidate docking | NvidiaNIM_diffdock / NvidiaNIM_boltz2 | Final scoring |

---

Evidence Grading System

Docking-Enhanced Grading

When NVIDIA NIM docking is available:

• Docking > reference -> upgrade to T0
• Docking within 5% -> upgrade to T2
• Docking within 20% -> maintain current tier
• Docking >20% worse -> downgrade one tier

---

Report Section Formatting Examples

Druggability Assessment Example

### 1.2 Druggability Assessment

| Factor | Assessment | Score | Source |
|--------|------------|-------|--------|
| Target class | Receptor tyrosine kinase | High | ChEMBL |
| Tractability bucket | 1 (approved drugs) | High | Open Targets |
| Known drugs | 15 approved | High | DGIdb |
| Binding site | ATP pocket (well-characterized) | High | PDB |

**Overall Druggability**: Highly druggable

*Sources: Open Targets, DGIdb, ChEMBL*

Known Actives Example

### 2.1 Known Active Compounds (ChEMBL)

**Total Bioactivity Points**: 2,847 (IC50: 1,234 | Ki: 892 | Kd: 456 | EC50: 265)
**Compounds with IC50 < 100 nM**: 156

| Compound | ChEMBL ID | IC50 (nM) | Max Phase | SMILES (truncated) |
|----------|-----------|-----------|-----------|-------------------|
| Erlotinib | CHEMBL553 | 2 | 4 | COc1cc2ncnc(Nc3ccc... |

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

Binding Site Example

### 3.2 Binding Pocket Characterization

**Pocket Volume**: ~850 A3 (well-defined)
**Key Interaction Residues**:
- **Hinge region**: M793 (backbone H-bond donor/acceptor)
- **Gatekeeper**: T790 (small residue, allows access)
- **DFG motif**: D855 (active conformation)

**Druggability Assessment**: High (enclosed pocket, conserved interactions)

ADMET Filter Summary Example

### 5.4 Filter Summary

| Filter Stage | Input | Passed | Failed | Pass Rate |
|--------------|-------|--------|--------|-----------|
| Physicochemical (Lipinski) | 568 | 456 | 112 | 80% |
| Drug-likeness (QED > 0.3) | 456 | 398 | 58 | 87% |
| Bioavailability (> 0.3) | 398 | 312 | 86 | 78% |
| Toxicity filters | 312 | 267 | 45 | 86% |
| Structural alerts | 267 | 234 | 33 | 88% |
| **Final Candidates** | **568** | **234** | **334** | **41%** |

Top Candidates Example

### 6.3 Top 20 Candidate Compounds

| Rank | ID | SMILES | Sim. Score | ADMET | Novelty | Overall | Rationale |
|------|-----|--------|------------|-------|---------|---------|-----------|
| 1 | CPD-001 | Cc1ccc... | 0.82 | 4.5 | Novel scaffold | 4.2 | High similarity, clean ADMET |

**Scaffold Diversity**: 7 distinct scaffolds in top 20
**Commercial Availability**: 12/20 available for purchase
**Estimated Hit Rate**: 15-25% (based on similarity to actives)

Therapeutic Antibody Landscape Example

### 1.2.5 Therapeutic Antibody Landscape

| Antibody (INN) | Target | Format | Phase | PDB |
|----------------|--------|--------|-------|-----|
| Pembrolizumab | PD-1 | IgG4 | Approved | 5DK3 |

**Competitive Landscape**: 3 approved antibodies target this protein
**Strategic Implication**: Small molecule approach offers differentiation (oral dosing, CNS penetration, cost)

*Source: Thera-SAbDab via `TheraSAbDab_search_by_target`*

Structure Prediction Example

### 1.4 Structure Prediction Quality

**Method**: AlphaFold2 via NVIDIA NIM
**Mean pLDDT**: 90.94 (very high confidence)

| Confidence Level | Range | Fraction | Interpretation |
|------------------|-------|----------|----------------|
| Very High | >=90 | 74.3% | Highly reliable |
| Confident | 70-90 | 16.0% | Reliable |
| Low | 50-70 | 9.0% | Use caution |
| Very Low | <50 | 0.7% | Unreliable |

*Source: NVIDIA NIM via `NvidiaNIM_alphafold2`*

---

Citation Format

Every section MUST include source attribution:

*Source: ChEMBL via `ChEMBL_get_target_activities` (CHEMBL203)*
*Source: PDB via `get_protein_metadata_by_pdb_id` (1M17)*
*Source: ADMET-AI via `ADMETAI_predict_toxicity`*
*Source: NVIDIA NIM via `NvidiaNIM_alphafold2` (pLDDT: 90.94)*
*Source: NVIDIA NIM via `NvidiaNIM_diffdock` (confidence: 0.906)*
*Source: NVIDIA NIM via `NvidiaNIM_genmol` (100 molecules)*

---

Data Output Files

In addition to the report:

• [TARGET]_candidate_compounds.csv - Prioritized compounds with: Rank, ID, SMILES, Similarity, ADMET_Score, Overall_Score, Source
• [TARGET]_bibliography.json - Literature references (optional)

Tools Reference for Small Molecule Binder Discovery

Complete tool reference with verified parameters and fallback chains.

Phase 1: Target Validation Tools

UniProt_search

Purpose: Resolve protein name to UniProt accession

result = tu.tools.UniProt_search(
    query="EGFR human",
    organism="human",
    limit=10
)
# Returns: list of UniProt entries with accession, protein name, gene name

MyGene_query_genes

Purpose: Get Ensembl and NCBI gene IDs

result = tu.tools.MyGene_query_genes(
    q="EGFR",
    species="human",
    fields="ensembl.gene,symbol,name"
)
# Returns: gene info with ensembl.gene, symbol, name

ChEMBL_search_targets

Purpose: Get ChEMBL target ID

result = tu.tools.ChEMBL_search_targets(
    query="EGFR",
    organism="Homo sapiens",
    limit=10
)
# Returns: targets with target_chembl_id, pref_name, target_type

OpenTargets_get_target_tractability_by_ensemblID

Purpose: Assess small molecule tractability

result = tu.tools.OpenTargets_get_target_tractability_by_ensemblID(
    ensemblId="ENSG00000146648"  # Note: camelCase
)
# Returns: tractability assessments by modality (small molecule, antibody, etc.)

⚠️ Parameter: Use ensemblId (camelCase), NOT ensembl_id

DGIdb_get_gene_druggability

Purpose: Get druggability categories

result = tu.tools.DGIdb_get_gene_druggability(
    genes=["EGFR"]  # List of gene symbols
)
# Returns: druggability info, drug count, categories

ChEMBL_search_binding_sites

Purpose: Find known binding sites

result = tu.tools.ChEMBL_search_binding_sites(
    target_chembl_id="CHEMBL203"
)
# Returns: binding site names, types

InterPro_get_protein_domains

Purpose: Get domain architecture

result = tu.tools.InterPro_get_protein_domains(
    accession="P00533"  # UniProt accession
)
# Returns: domains, families, sites with positions

---

Thera-SAbDab Tools (NEW - Therapeutic Antibody Landscape)

Check therapeutic antibody competition for your target.

TheraSAbDab_search_by_target

Purpose: Find antibodies targeting specific antigen

result = tu.tools.TheraSAbDab_search_by_target(
    target="PD-1"  # Target antigen name
)
# Returns: list of therapeutic antibodies, phase, format, PDB IDs

TheraSAbDab_search_therapeutics

Purpose: Search antibodies by name or keyword

result = tu.tools.TheraSAbDab_search_therapeutics(
    query="pembrolizumab"  # Antibody name or target
)
# Returns: matching therapeutics with target, phase, format

TheraSAbDab_get_all_therapeutics

Purpose: Get summary of all therapeutic antibodies

result = tu.tools.TheraSAbDab_get_all_therapeutics()
# Returns: total count, distribution by phase, distribution by format

Why Use Thera-SAbDab:

• Competitive intelligence: See what antibodies already target your protein
• Target validation: Approved antibodies = validated target
• Structural data: Links to PDB structures of antibody-target complexes
• Strategic differentiation: Identify where small molecules offer advantages

Example Competitive Analysis:

def assess_antibody_competition(tu, target_name):
    """Assess therapeutic antibody competitive landscape."""
    
    # Search by target
    results = tu.tools.TheraSAbDab_search_by_target(target=target_name)
    
    if results.get('status') == 'success':
        antibodies = results['data'].get('therapeutics', [])
        
        # Count by phase
        phases = {}
        for ab in antibodies:
            phase = ab.get('phase', 'Unknown')
            phases[phase] = phases.get(phase, 0) + 1
        
        return {
            'total': len(antibodies),
            'by_phase': phases,
            'has_approved': 'Approved' in phases,
            'top_antibodies': antibodies[:5]
        }
    return None

---

GPCRdb Tools (for GPCR Targets)

~35% of approved drugs target GPCRs. Use GPCRdb for specialized GPCR data.

GPCRdb_get_protein

Purpose: Get GPCR protein information

result = tu.tools.GPCRdb_get_protein(
    operation="get_protein",
    protein="adrb2_human"  # Entry name format: {gene}_human
)
# Returns: GPCR family, class, species, sequence info

GPCRdb_list_proteins

Purpose: List GPCR families or proteins in a family

# List all GPCR families
families = tu.tools.GPCRdb_list_proteins(
    operation="list_proteins"
)

# List proteins in specific family
proteins = tu.tools.GPCRdb_list_proteins(
    operation="list_proteins",
    family="001"  # Class A
)

GPCRdb_get_structures

Purpose: Get GPCR structures with receptor state (active/inactive)

result = tu.tools.GPCRdb_get_structures(
    operation="get_structures",
    protein="adrb2_human",
    state="active"  # Optional: "active", "inactive", "intermediate"
)
# Returns: PDB IDs with receptor state, resolution, ligand info

GPCRdb_get_ligands

Purpose: Get known GPCR ligands

result = tu.tools.GPCRdb_get_ligands(
    operation="get_ligands",
    protein="adrb2_human"
)
# Returns: Ligands with binding affinity, type (agonist/antagonist)

GPCRdb_get_mutations

Purpose: Get experimental mutation data for GPCR

result = tu.tools.GPCRdb_get_mutations(
    operation="get_mutations",
    protein="adrb2_human"
)
# Returns: Mutations with effect on ligand binding/signaling

Why GPCRdb is Essential for GPCR Binder Discovery:

• Receptor state structures: Active vs. inactive conformations
• Ballesteros-Weinstein numbering: Standard residue numbering across GPCRs
• Curated ligand data: Agonists, antagonists, allosteric modulators
• Mutation effects: Direct effects on ligand binding

---

Phase 1.4: Structure Prediction (NVIDIA NIM)

Requires: NVIDIA_API_KEY environment variable

NvidiaNIM_alphafold2

Purpose: High-accuracy structure prediction with pLDDT confidence scores

result = tu.tools.NvidiaNIM_alphafold2(
    sequence="MRPSGTAGAALLALL...",  # Protein sequence
    algorithm="mmseqs2",             # MSA algorithm: "mmseqs2" or "jackhmmer"
    relax_prediction=False           # Optional: Run relaxation
)
# Returns: PDB structure string with pLDDT scores per residue
# pLDDT interpretation:
#   ≥90: Very high confidence
#   70-90: Confident
#   50-70: Low confidence
#   <50: Very low confidence

⚠️ Note: Async operation, may take 5-15 minutes. Max sequence length ~2000 AA.

NvidiaNIM_esmfold

Purpose: Fast structure prediction (synchronous)

result = tu.tools.NvidiaNIM_esmfold(
    sequence="MRPSGTAGAALLALL..."  # Protein sequence (max 1024 AA)
)
# Returns: PDB structure string

⚠️ Note: Fast (~30 sec) but limited to 1024 residues. No MSA, slightly lower accuracy.

NvidiaNIM_msa_search

Purpose: Generate multiple sequence alignment for structure prediction

result = tu.tools.NvidiaNIM_msa_search(
    sequence="MRPSGTAGAALLALL...",
    database="uniref90"  # "uniref90", "bfd", etc.
)
# Returns: MSA in A3M format

Use: Pre-generate MSA for repeated AlphaFold2 predictions.

---

Phase 2: Known Ligand Mining Tools

ChEMBL_get_target_activities

Purpose: Get all bioactivity data for target

result = tu.tools.ChEMBL_get_target_activities(
    target_chembl_id="CHEMBL203",
    limit=500
)
# Returns: activities with molecule_chembl_id, standard_type, standard_value

⚠️ Parameter: Use target_chembl_id, NOT chembl_target_id

ChEMBL_get_molecule

Purpose: Get detailed molecule information

result = tu.tools.ChEMBL_get_molecule(
    molecule_chembl_id="CHEMBL553"
)
# Returns: molecule data including SMILES, max_phase, properties

GtoPdb_get_interactions

Purpose: Get pharmacological interactions

result = tu.tools.GtoPdb_get_interactions(
    target_id="1797"  # GtoPdb target ID
)
# Returns: ligands with pKi, pIC50, action type

OpenTargets_get_chemical_probes_by_target_ensemblID

Purpose: Find validated chemical probes

result = tu.tools.OpenTargets_get_chemical_probes_by_target_ensemblID(
    ensemblId="ENSG00000146648"
)
# Returns: probes with ratings, use recommendations

OpenTargets_get_associated_drugs_by_target_ensemblID

Purpose: Get approved/clinical drugs

result = tu.tools.OpenTargets_get_associated_drugs_by_target_ensemblID(
    ensemblId="ENSG00000146648"
)
# Returns: drugs with phase, mechanism of action

BindingDB_get_ligands_by_uniprot (NEW)

Purpose: Get ligands with measured binding affinities

result = tu.tools.BindingDB_get_ligands_by_uniprot(
    uniprot="P00533",  # UniProt accession
    affinity_cutoff=10000  # Max affinity in nM
)
# Returns: SMILES, affinity_type (Ki/IC50/Kd), affinity value, PMID

Advantages: Direct affinity measurements, literature links, compounds not in ChEMBL

BindingDB_get_ligands_by_pdb (NEW)

Purpose: Get ligands for a PDB structure

result = tu.tools.BindingDB_get_ligands_by_pdb(
    pdb_ids="1M17",  # PDB ID
    affinity_cutoff=10000,
    sequence_identity=100
)
# Returns: ligands binding to the structure's protein

BindingDB_get_targets_by_compound (NEW)

Purpose: Find off-targets for selectivity analysis

result = tu.tools.BindingDB_get_targets_by_compound(
    smiles="CC(=O)Nc1ccc(cc1)O",  # Compound SMILES
    similarity_cutoff=0.85
)
# Returns: other proteins binding similar compounds

PubChem_search_assays_by_target_gene (NEW)

Purpose: Find HTS screening data for target

result = tu.tools.PubChem_search_assays_by_target_gene(
    gene_symbol="EGFR"
)
# Returns: list of AIDs (assay IDs) for this target

PubChem_get_assay_summary (NEW)

Purpose: Get assay statistics

result = tu.tools.PubChem_get_assay_summary(aid=504526)
# Returns: active/inactive counts, target info, assay metadata

PubChem_get_assay_active_compounds (NEW)

Purpose: Get hits from HTS screen

result = tu.tools.PubChem_get_assay_active_compounds(aid=504526)
# Returns: CIDs of active compounds

PubChem_get_assay_dose_response (NEW)

Purpose: Get IC50/EC50 curves

result = tu.tools.PubChem_get_assay_dose_response(aid=1053104)
# Returns: concentration-response data

PubChem_get_compound_bioactivity (NEW)

Purpose: Get all bioactivity for a compound

result = tu.tools.PubChem_get_compound_bioactivity(cid=2244)
# Returns: all assays, targets, activity outcomes

When to Use Each Source:

---

Phase 3: Structure Analysis Tools

PDB_search_similar_structures

Purpose: Find PDB structures by sequence

result = tu.tools.PDB_search_similar_structures(
    query="P00533",  # UniProt or PDB ID
    type="sequence"  # or "structure"
)
# Returns: PDB IDs with similarity scores

get_protein_metadata_by_pdb_id

Purpose: Get structure metadata

result = tu.tools.get_protein_metadata_by_pdb_id(
    pdb_id="1M17"
)
# Returns: resolution, method, deposition date, title

get_binding_affinity_by_pdb_id

Purpose: Get co-crystallized ligand affinities

result = tu.tools.get_binding_affinity_by_pdb_id(
    pdb_id="1M17"
)
# Returns: ligand codes with Kd, Ki, IC50 values

alphafold_get_prediction

Purpose: Get AlphaFold predicted structure

result = tu.tools.alphafold_get_prediction(
    accession="P00533"  # UniProt accession
)
# Returns: structure prediction with pLDDT scores

⚠️ Parameter: Use accession, NOT uniprot

get_ligand_smiles_by_chem_comp_id

Purpose: Get ligand structure from PDB

result = tu.tools.get_ligand_smiles_by_chem_comp_id(
    chem_comp_id="AQ4"  # 3-letter ligand code from PDB
)
# Returns: SMILES, name, formula

EMDB_search_structures (NEW)

Purpose: Search EMDB for cryo-EM structures

result = tu.tools.EMDB_search_structures(
    query="EGFR membrane receptor"
)
# Returns: EMDB entries with emdb_id, resolution, title

When to use: Membrane proteins (GPCRs, ion channels), large complexes, targets where conformational states matter.

EMDB_get_structure (NEW)

Purpose: Get details for EMDB entry including associated PDB models

result = tu.tools.EMDB_get_structure(
    entry_id="EMD-12345"
)
# Returns: entry details including pdb_ids (associated atomic models)

Cryo-EM vs X-ray Decision:

---

Phase 3.5: Docking Validation (NVIDIA NIM)

Requires: NVIDIA_API_KEY environment variable

NvidiaNIM_diffdock

Purpose: Blind molecular docking (no predefined binding site needed)

result = tu.tools.NvidiaNIM_diffdock(
    protein=pdb_content,      # PDB file content as string
    ligand=ligand_sdf,        # Ligand in SDF or MOL2 format
    num_poses=10              # Number of poses to generate
)
# Returns: Docked poses with confidence scores
# Confidence interpretation:
#   >0.9: Excellent, high-confidence pose
#   0.7-0.9: Good, reliable binding mode
#   0.5-0.7: Moderate, consider alternatives
#   <0.5: Low confidence, may not bind

Use: When you have PDB structure and ligand SDF file.

NvidiaNIM_boltz2

Purpose: Protein-ligand complex from sequence + SMILES

result = tu.tools.NvidiaNIM_boltz2(
    polymers=[{
        "molecule_type": "protein",
        "sequence": "MRPSGTAGAALLALL..."
    }],
    ligands=[{
        "smiles": "COc1cc2ncnc(Nc3ccc(C#C)cc3)c2cc1OCCOC"
    }],
    sampling_steps=50,       # Default: 50, higher = better quality
    diffusion_samples=1      # Number of complex samples
)
# Returns: Protein-ligand complex structure
# Metrics: aggregate_score, pTM, ipTM
#   pTM: Protein structure confidence
#   ipTM: Interface prediction confidence (protein-ligand contact)

Use: When starting from SMILES only, no SDF file needed.

Docking Method Selection

---

Phase 4: Compound Expansion Tools

ChEMBL_search_similar_molecules

Purpose: Similarity search in ChEMBL

result = tu.tools.ChEMBL_search_similar_molecules(
    molecule="CC(C)Cc1ccc(cc1)C(C)C(O)=O",  # SMILES, ChEMBL ID, or name
    similarity=70  # Tanimoto threshold (0-100)
)
# Returns: similar molecules with similarity score

⚠️ Parameter: Use molecule, NOT smiles

PubChem_search_compounds_by_similarity

Purpose: Similarity search in PubChem

result = tu.tools.PubChem_search_compounds_by_similarity(
    smiles="CC(C)Cc1ccc(cc1)C(C)C(O)=O",
    threshold=0.7  # Tanimoto (0-1)
)
# Returns: CIDs with similarity scores

ChEMBL_search_substructure

Purpose: Substructure search in ChEMBL

result = tu.tools.ChEMBL_search_substructure(
    smiles="c1ccc2ncncc2c1"  # Quinazoline core
)
# Returns: molecules containing substructure

PubChem_search_compounds_by_substructure

Purpose: Substructure search in PubChem

result = tu.tools.PubChem_search_compounds_by_substructure(
    smiles="c1ccc2ncncc2c1"
)
# Returns: CIDs containing substructure

STITCH_get_chemical_protein_interactions

Purpose: Cross-database chemical-protein links

result = tu.tools.STITCH_get_chemical_protein_interactions(
    identifier="EGFR",
    species=9606  # Human
)
# Returns: chemicals with confidence scores

---

Phase 4.4: De Novo Molecule Generation (NVIDIA NIM)

Requires: NVIDIA_API_KEY environment variable

NvidiaNIM_genmol

Purpose: Scaffold hopping with masked regions

result = tu.tools.NvidiaNIM_genmol(
    smiles="COc1cc2ncnc(Nc3ccc([*{3-8}])c([*{1-3}])c3)c2cc1OCCCN1CCOCC1",
    num_molecules=100,       # Number to generate
    temperature=2.0,         # Diversity: 0.5=conservative, 2.0=diverse
    scoring="QED"            # "QED" or "logP"
)
# Returns: Generated molecules with QED/LogP scores
# Mask syntax: [*{min-max}] specifies atom count range

Mask Design Guidelines:

Temperature Selection:

NvidiaNIM_molmim

Purpose: Controlled generation from reference molecule

result = tu.tools.NvidiaNIM_molmim(
    smi="COc1cc2ncnc(Nc3ccc(Cl)cc3)c2cc1OCCN1CCOCC1",
    num_molecules=50,        # Number to generate
    algorithm="CMA-ES"       # Optimization algorithm
)
# Returns: Optimized analogs with property scores
# Generates molecules similar to reference but with optimized properties

Use: Generate close analogs of top actives with improved properties.

Generation Strategy

1. Identify seeds: Top 3-5 actives from Phase 2
         ↓
2. Design approach:
   ├─ Know specific positions to vary? → NvidiaNIM_genmol (with masks)
   └─ Want general optimization? → NvidiaNIM_molmim
         ↓
3. Generate: 50-100 molecules per seed
         ↓
4. Filter: Pass to Phase 5 (ADMET)
         ↓
5. Dock: Score survivors in Phase 6

---

Phase 5: ADMET Tools

ADMETAI_predict_physicochemical_properties

Purpose: Drug-likeness assessment

result = tu.tools.ADMETAI_predict_physicochemical_properties(
    smiles=["CC(C)Cc1ccc(cc1)C(C)C(O)=O"]  # List of SMILES
)
# Returns: MW, logP, HBD, HBA, Lipinski violations, QED, TPSA

⚠️ Parameter: smiles must be a LIST, even for single compound

ADMETAI_predict_bioavailability

Purpose: Oral absorption prediction

result = tu.tools.ADMETAI_predict_bioavailability(
    smiles=["CC(C)Cc1ccc(cc1)C(C)C(O)=O"]
)
# Returns: Bioavailability_Ma, HIA_Hou, PAMPA, Caco2, Pgp_substrate

ADMETAI_predict_toxicity

Purpose: Toxicity endpoint predictions

result = tu.tools.ADMETAI_predict_toxicity(
    smiles=["CC(C)Cc1ccc(cc1)C(C)C(O)=O"]
)
# Returns: AMES, hERG, DILI, ClinTox, LD50, Carcinogens

ADMETAI_predict_CYP_interactions

Purpose: CYP enzyme interactions

result = tu.tools.ADMETAI_predict_CYP_interactions(
    smiles=["CC(C)Cc1ccc(cc1)C(C)C(O)=O"]
)
# Returns: CYP1A2, CYP2C9, CYP2C19, CYP2D6, CYP3A4 substrate/inhibitor

ADMETAI_predict_clearance_distribution

Purpose: PK predictions

result = tu.tools.ADMETAI_predict_clearance_distribution(
    smiles=["CC(C)Cc1ccc(cc1)C(C)C(O)=O"]
)
# Returns: Clearance, Half_Life, VDss, PPB

ChEMBL_search_compound_structural_alerts

Purpose: PAINS and toxicophore detection

result = tu.tools.ChEMBL_search_compound_structural_alerts(
    smiles="CC(C)Cc1ccc(cc1)C(C)C(O)=O"
)
# Returns: structural alerts, PAINS flags

---

Phase 6: Candidate Docking (NVIDIA NIM)

Requires: NVIDIA_API_KEY environment variable

Batch Docking Strategy

After ADMET filtering, dock all candidates against the target structure.

# Batch docking workflow
candidates = admet_passed_compounds  # From Phase 5

# Get reference score first
reference_result = tu.tools.NvidiaNIM_diffdock(
    protein=pdb_content,
    ligand=reference_ligand_sdf,
    num_poses=10
)
reference_confidence = reference_result['best_pose_confidence']

# Dock all candidates
docking_results = []
for compound in candidates:
    result = tu.tools.NvidiaNIM_diffdock(
        protein=pdb_content,
        ligand=compound['sdf'],
        num_poses=5
    )
    docking_results.append({
        'id': compound['id'],
        'confidence': result['best_pose_confidence'],
        'vs_reference': (result['best_pose_confidence'] / reference_confidence - 1) * 100
    })

# Sort by confidence
ranked = sorted(docking_results, key=lambda x: x['confidence'], reverse=True)

Scoring Integration

---

Phase 6.5: Literature Evidence (NEW)

PubMed_search_articles

Purpose: Search published SAR studies

result = tu.tools.PubMed_search_articles(
    query="EGFR inhibitor SAR structure-activity",
    limit=30
)
# Returns: articles with pmid, title, abstract, publication date

EuropePMC_search_articles (for Preprints)

Purpose: Search preprints from bioRxiv, medRxiv, and other sources (latest findings, not peer-reviewed)

# Search preprints using EuropePMC (bioRxiv/medRxiv don't have search APIs)
result = tu.tools.EuropePMC_search_articles(
    query="EGFR small molecule discovery",
    source="PPR",  # PPR = Preprints only
    pageSize=15
)

# If you have a DOI, get full bioRxiv metadata:
full_metadata = tu.tools.BioRxiv_get_preprint(doi="10.1101/2023.12.01.569554")
# Returns: preprints with doi, title, posted date

⚠️ Note: Preprints NOT peer-reviewed. Use for emerging compounds/methods.

MedRxiv_get_preprint

Purpose: Get medRxiv preprint by DOI (for search, use EuropePMC with source='PPR')

# Get preprint by DOI
result = tu.tools.MedRxiv_get_preprint(doi="10.1101/2021.04.29.21256344")

# For searching clinical preprints, use EuropePMC:
search = tu.tools.EuropePMC_search_articles(
    query="EGFR inhibitor clinical trial",
    source="PPR",
    pageSize=10
)
# Returns: preprints with doi, title, abstract, etc.

openalex_search_works

Purpose: Search with citation analysis

result = tu.tools.openalex_search_works(
    query="EGFR kinase inhibitor structure",
    limit=20
)
# Returns: works with cited_by_count, publication_year

Use: Identify high-impact papers and validate compound importance.

SemanticScholar_search_papers

Purpose: AI-ranked paper search

result = tu.tools.SemanticScholar_search_papers(
    query="EGFR small molecule binder",
    limit=20
)
# Returns: papers with relevance ranking, citations

---

Fallback Chains

Target ID Resolution

Primary: ChEMBL_search_targets
├─ Success → Use target_chembl_id
└─ Fail → GtoPdb_search_targets (for GPCR/ion channel/enzyme)
         └─ Fail → Document "Target not in databases"

GPCR-Specific Data (NEW)

If target is GPCR:
    Primary: GPCRdb_get_protein
    ├─ Success → Get GPCR family, class
    │   ├─ GPCRdb_get_structures → Active/inactive state structures
    │   ├─ GPCRdb_get_ligands → Known agonists/antagonists
    │   └─ GPCRdb_get_mutations → Mutation effects on binding
    └─ Fail (not in GPCRdb) → Use GtoPdb_search_targets

Druggability Assessment

Primary: OpenTargets_get_target_tractability_by_ensemblID
├─ Success → Use tractability data
└─ Fail → DGIdb_get_gene_druggability
         └─ Fail → Use target class as proxy

Bioactivity Data

Primary: ChEMBL_get_target_activities
├─ Success → Use ChEMBL data
└─ Fail → BindingDB_get_ligands_by_uniprot (NEW)
         ├─ Success → Use BindingDB data
         └─ Fail → GtoPdb_get_interactions
                  ├─ Success → Use GtoPdb data
                  └─ Fail → PubChem_search_assays_by_target_gene (NEW)
                           └─ Fail → Document "No bioactivity data"

Similarity Search

Primary: ChEMBL_search_similar_molecules
├─ Success → Process results
└─ Fail → PubChem_search_compounds_by_similarity
         └─ Fail → Document "Similarity search failed"

Structure Retrieval

Primary: get_protein_metadata_by_pdb_id (for each PDB)
├─ Success → Use experimental structure
└─ Fail (no PDB) → EMDB_search_structures (for membrane proteins)
         ├─ Success → Get PDB model via EMDB_get_structure
         └─ Fail → NvidiaNIM_alphafold2
                 └─ Fail (API error) → NvidiaNIM_esmfold
                         └─ Fail → alphafold_get_prediction (AlphaFold DB)
                                 └─ Fail → Document "No structural information"

Literature Search (NEW)

Primary: PubMed_search_articles (peer-reviewed)
├─ Success → Use published literature
└─ Supplement with:
         ├─ EuropePMC_search_articles (source='PPR' for preprints)
         └─ openalex_search_works (citation analysis)

Docking

Primary: NvidiaNIM_diffdock (have PDB + SDF)
├─ Success → Use docking scores
└─ Fail → NvidiaNIM_boltz2 (from sequence + SMILES)
         └─ Fail → Skip docking, use similarity-based scoring

De Novo Generation

Primary: NvidiaNIM_genmol (specific position variation)
├─ Success → Process generated molecules
└─ Fail → NvidiaNIM_molmim (general analog generation)
         └─ Fail → Use similarity search only (no generation)

ADMET Prediction

Primary: ADMETAI_predict_* (all endpoints)
├─ Success → Use predictions
└─ Fail (invalid SMILES) → Skip compound, document reason
└─ Fail (API error) → Document "ADMET unavailable"

---

Common Parameter Errors

---

Batch Processing Pattern

For efficiency, batch similar operations:

# Define calls
calls = [
    {"name": "ChEMBL_get_molecule", "arguments": {"molecule_chembl_id": id}}
    for id in chembl_ids[:50]  # Batch of 50
]

# Execute in parallel
results = tu.run_batch(calls)

# Process results
for result in results:
    if result and 'molecule_structures' in result:
        process_molecule(result)

---

Rate Limiting Awareness

NVIDIA NIM Specific Notes

API Key Check:

import os
if not os.environ.get("NVIDIA_API_KEY"):
    print("Warning: NVIDIA_API_KEY not set. NvidiaNIM tools unavailable.")
    # Fall back to non-NIM alternatives

For large expansions (>500 compounds): 1. Use batch endpoints 2. Add small delays between batches 3. Cache results for reuse 4. For docking: prioritize top 50-100 candidates only

Binder Discovery Workflow Details

Detailed procedures, code patterns, and screening protocols for each phase.

Phase 0: Tool Verification

CRITICAL: Verify tool parameters before calling unfamiliar tools.

# Check tool params to prevent silent failures
tool_info = tu.tools.get_tool_info(tool_name="ChEMBL_get_target_activities")

Known Parameter Corrections

---

Phase 1: Target Validation Details

1.1 Identifier Resolution Chain

1. UniProt_search(query=target_name, organism="human")
   -> Extract: UniProt accession, gene name, protein name

2. MyGene_query_genes(q=gene_symbol, species="human")
   -> Extract: Ensembl gene ID, NCBI gene ID

3. ChEMBL_search_targets(query=target_name, organism="Homo sapiens")
   -> Extract: ChEMBL target ID, target type

4. GtoPdb_search_targets(query=target_name)
   -> Extract: GtoPdb target ID (if GPCR/ion channel/enzyme)

Store all IDs for downstream queries:

ids = {
    'uniprot': 'P00533',
    'ensembl': 'ENSG00000146648',
    'chembl_target': 'CHEMBL203',
    'gene_symbol': 'EGFR',
    'gtopdb': '1797'  # if available
}

1.2 Druggability Assessment

Multi-Source Triangulation:

1. OpenTargets_get_target_tractability_by_ensemblID(ensemblId)
   -> Extract: Small molecule tractability score, bucket

2. DGIdb_get_gene_druggability(genes=[gene_symbol])
   -> Extract: Druggability categories, known drug count

3. OpenTargets_get_target_classes_by_ensemblID(ensemblId)
   -> Extract: Target class (kinase, GPCR, etc.)

4. GPCRdb_get_protein(protein=entry_name)  # for GPCRs
   -> Extract: GPCR family, receptor state, ligand binding data

1.2a GPCRdb Integration (for GPCR Targets)

~35% of all approved drugs target GPCRs. For GPCR targets, use specialized data:

def check_if_gpcr_and_enrich(tu, target_name, uniprot_id):
    """Check if target is GPCR and get specialized data."""

    entry_name = f"{target_name.lower()}_human"

    gpcr_info = tu.tools.GPCRdb_get_protein(
        operation="get_protein",
        protein=entry_name
    )

    if gpcr_info.get('status') == 'success':
        structures = tu.tools.GPCRdb_get_structures(
            operation="get_structures",
            protein=entry_name
        )
        ligands = tu.tools.GPCRdb_get_ligands(
            operation="get_ligands",
            protein=entry_name
        )
        mutations = tu.tools.GPCRdb_get_mutations(
            operation="get_mutations",
            protein=entry_name
        )

        return {
            'is_gpcr': True,
            'gpcr_family': gpcr_info['data'].get('family'),
            'gpcr_class': gpcr_info['data'].get('receptor_class'),
            'structures': structures['data'].get('structures', []),
            'ligands': ligands['data'].get('ligands', []),
            'mutation_data': mutations['data'].get('mutations', [])
        }

    return {'is_gpcr': False}

1.2.5 Therapeutic Antibody Landscape

Check Thera-SAbDab for therapeutic antibodies against target:

def check_therapeutic_antibodies(tu, target_name):
    results = tu.tools.TheraSAbDab_search_by_target(target=target_name)

    if results.get('status') == 'success':
        antibodies = results['data'].get('therapeutics', [])
        by_phase = {'Approved': [], 'Phase 3': [], 'Phase 2': [], 'Phase 1': [], 'Preclinical': []}
        for ab in antibodies:
            phase = ab.get('phase', 'Unknown')
            for key in by_phase.keys():
                if key.lower() in phase.lower():
                    by_phase[key].append(ab)
                    break

        return {
            'total_antibodies': len(antibodies),
            'by_phase': by_phase,
            'antibodies': antibodies[:10],
            'competitive_alert': len(by_phase.get('Approved', [])) > 0
        }
    return None

1.3 Binding Site Analysis

1. ChEMBL_search_binding_sites(target_chembl_id)
   -> Extract: Binding site names, types

2. get_binding_affinity_by_pdb_id(pdb_id)  # For each PDB with ligand
   -> Extract: Kd, Ki, IC50 values for co-crystallized ligands

3. InterPro_get_protein_domains(uniprot_accession)
   -> Extract: Domain architecture, active sites

1.4 Structure Prediction (NVIDIA NIM)

Requires: NVIDIA_API_KEY environment variable

Option A: AlphaFold2 (High accuracy, async)

NvidiaNIM_alphafold2(
    sequence=kinase_domain_sequence,
    algorithm="mmseqs2",
    relax_prediction=False
)
-> Returns: PDB structure with pLDDT confidence scores
-> Use when: Accuracy is critical, time is available (~5-15 min)

Option B: ESMFold (Fast, synchronous)

NvidiaNIM_esmfold(sequence=kinase_domain_sequence)
-> Returns: PDB structure (max 1024 AA)
-> Use when: Quick assessment needed (~30 sec)

---

Phase 2: Known Ligand Mining Details

2.1 ChEMBL Bioactivity Data

1. ChEMBL_get_target_activities(target_chembl_id, limit=500)
   -> Filter: standard_type in ["IC50", "Ki", "Kd", "EC50"]
   -> Filter: standard_value < 10000 nM
   -> Extract: ChEMBL molecule IDs, SMILES, potency values

2. ChEMBL_get_molecule(molecule_chembl_id)  # For top actives
   -> Extract: Full molecular data, max_phase, oral flag

2.5 BindingDB Affinity Data

def get_bindingdb_ligands(tu, uniprot_id, affinity_cutoff=10000):
    """Get ligands from BindingDB with measured affinities."""

    result = tu.tools.BindingDB_get_ligands_by_uniprot(
        uniprot=uniprot_id,
        affinity_cutoff=affinity_cutoff
    )

    if result:
        ligands = []
        for entry in result:
            ligands.append({
                'smiles': entry.get('smile'),
                'affinity_type': entry.get('affinity_type'),
                'affinity_nM': entry.get('affinity'),
                'pmid': entry.get('pmid'),
                'monomer_id': entry.get('monomerid')
            })
        ligands.sort(key=lambda x: float(x['affinity_nM']) if x['affinity_nM'] else 1e6)
        return ligands[:50]
    return []

def find_compound_polypharmacology(tu, smiles, similarity_cutoff=0.85):
    """Find off-target interactions for selectivity analysis."""
    return tu.tools.BindingDB_get_targets_by_compound(
        smiles=smiles,
        similarity_cutoff=similarity_cutoff
    )

2.6 PubChem BioAssay Screening Data

def get_pubchem_assays_for_target(tu, gene_symbol):
    """Get bioassays and active compounds from PubChem."""

    assays = tu.tools.PubChem_search_assays_by_target_gene(
        gene_symbol=gene_symbol
    )

    results = {'assays': [], 'total_active_compounds': 0}

    if assays.get('data', {}).get('aids'):
        for aid in assays['data']['aids'][:10]:
            summary = tu.tools.PubChem_get_assay_summary(aid=aid)
            actives = tu.tools.PubChem_get_assay_active_compounds(aid=aid)
            active_cids = actives.get('data', {}).get('cids', [])

            results['assays'].append({
                'aid': aid,
                'summary': summary.get('data', {}),
                'active_count': len(active_cids)
            })
            results['total_active_compounds'] += len(active_cids)
    return results

When to Use Each Source:

---

Phase 3: Structure Analysis Details

3.1 PDB Structure Retrieval

1. PDB_search_similar_structures(query=uniprot_accession, type="sequence")
   -> Extract: PDB IDs with ligands

2. get_protein_metadata_by_pdb_id(pdb_id)
   -> Extract: Resolution, method, ligand codes

3. alphafold_get_prediction(accession=uniprot_accession)
   -> Extract: Predicted structure (if no experimental)

3.1b EMDB Cryo-EM Structures

Prioritize for: Membrane proteins (GPCRs, ion channels), large complexes.

def get_cryoem_structures(tu, target_name, uniprot_accession):
    """Get cryo-EM structures for membrane targets."""
    emdb_results = tu.tools.EMDB_search_structures(
        query=f"{target_name} membrane receptor"
    )

    structures = []
    for entry in emdb_results[:5]:
        details = tu.tools.EMDB_get_structure(entry_id=entry['emdb_id'])
        pdb_models = details.get('pdb_ids', [])
        structures.append({
            'emdb_id': entry['emdb_id'],
            'resolution': entry.get('resolution', 'N/A'),
            'title': entry.get('title', 'N/A'),
            'conformational_state': details.get('state', 'Unknown'),
            'pdb_models': pdb_models
        })
    return structures

When to use cryo-EM over X-ray:

---

Phase 3.5: Docking Validation (NVIDIA NIM)

Requires: NVIDIA_API_KEY environment variable

Reference Compound Docking

Option A: DiffDock (Blind docking, PDB + SDF input)

NvidiaNIM_diffdock(
    protein=pdb_content,
    ligand=reference_sdf,
    num_poses=10
)
-> Returns: Docking poses with confidence scores
-> Use: When you have PDB structure and ligand SDF file

Option B: Boltz2 (From sequence + SMILES)

NvidiaNIM_boltz2(
    polymers=[{"molecule_type": "protein", "sequence": kinase_sequence}],
    ligands=[{"smiles": "COc1cc2ncnc(Nc3ccc(C#C)cc3)c2cc1OCCOC"}],
    sampling_steps=50,
    diffusion_samples=1
)
-> Returns: Protein-ligand complex structure
-> Use: When starting from SMILES, no SDF needed

Docking Score Interpretation

---

Phase 4: Compound Expansion Details

4.1 Similarity Search

1. ChEMBL_search_similar_molecules(molecule=top_active_smiles, similarity=70)
   -> Extract: Similar compounds not yet tested on target

2. PubChem_search_compounds_by_similarity(smiles, threshold=0.7)
   -> Extract: PubChem CIDs with similar structures

Strategy:

• Use 3-5 diverse actives as seeds
• Similarity threshold: 70-85% (balance novelty vs. activity)
• Prioritize compounds NOT in ChEMBL bioactivity for target

4.2 Substructure Search

1. ChEMBL_search_substructure(smiles=core_scaffold)
2. PubChem_search_compounds_by_substructure(smiles=core_scaffold)

4.3 Cross-Database Mining

1. STITCH_get_chemical_protein_interactions(identifier=target_gene)
2. DGIdb_get_drug_gene_interactions(genes=[gene_symbol])

4.4 De Novo Molecule Generation (NVIDIA NIM)

Option A: GenMol (Scaffold Hopping with Masked Regions)

NvidiaNIM_genmol(
    smiles="COc1cc2ncnc(Nc3ccc([*{3-8}])c([*{1-3}])c3)c2cc1OCCCN1CCOCC1",
    num_molecules=100,
    temperature=2.0,
    scoring="QED"
)

Mask Design Strategy:

Temperature Selection:

Option B: MolMIM (Controlled Generation from Reference)

NvidiaNIM_molmim(
    smi="COc1cc2ncnc(Nc3ccc(Cl)cc3)c2cc1OCCN1CCOCC1",
    num_molecules=50,
    algorithm="CMA-ES"
)

Generation Workflow: 1. Identify top 3-5 actives from Phase 2 2. Design masked SMILES for GenMol OR use as reference for MolMIM 3. Generate 50-100 molecules per seed 4. Pass generated molecules to Phase 5 (ADMET filtering) 5. Dock survivors in Phase 6 for final ranking

---

Phase 5: ADMET Filtering Details

5.1 Physicochemical Properties

ADMETAI_predict_physicochemical_properties(smiles=[compound_list])
-> Filter: Lipinski violations <= 1
-> Filter: QED > 0.3
-> Filter: MW 200-600

5.2 ADMET Endpoints

1. ADMETAI_predict_bioavailability(smiles=[compound_list])
   -> Filter: Oral bioavailability > 0.3

2. ADMETAI_predict_toxicity(smiles=[compound_list])
   -> Filter: AMES < 0.5, hERG < 0.5, DILI < 0.5

3. ADMETAI_predict_CYP_interactions(smiles=[compound_list])
   -> Flag: CYP3A4 inhibitors (drug interaction risk)

5.3 Structural Alerts

ChEMBL_search_compound_structural_alerts(smiles=compound_smiles)
-> Flag: PAINS, reactive groups, toxicophores

---

Phase 6: Candidate Docking & Prioritization Details

Batch Docking Workflow

candidates = admet_passed_compounds

# Get reference score first
reference_result = tu.tools.NvidiaNIM_diffdock(
    protein=pdb_content,
    ligand=reference_ligand_sdf,
    num_poses=10
)
reference_confidence = reference_result['best_pose_confidence']

# Dock all candidates
docking_results = []
for compound in candidates:
    result = tu.tools.NvidiaNIM_diffdock(
        protein=pdb_content,
        ligand=compound['sdf'],
        num_poses=5
    )
    docking_results.append({
        'id': compound['id'],
        'confidence': result['best_pose_confidence'],
        'vs_reference': (result['best_pose_confidence'] / reference_confidence - 1) * 100
    })

ranked = sorted(docking_results, key=lambda x: x['confidence'], reverse=True)

Scoring Framework

Synthesis Feasibility

SA Score Interpretation:

• 1-3: Easy synthesis
• 3-5: Moderate complexity
• 5-10: Challenging synthesis

---

Phase 6.5: Literature Evidence

Literature Search for Validation

def search_binder_literature(tu, target_name, compound_scaffolds):
    """Search literature for compound and target evidence."""

    # PubMed: Published SAR studies
    sar_papers = tu.tools.PubMed_search_articles(
        query=f"{target_name} inhibitor SAR structure-activity",
        limit=30
    )

    # EuropePMC: Preprints (bioRxiv/medRxiv)
    preprints = tu.tools.EuropePMC_search_articles(
        query=f"{target_name} small molecule discovery",
        source="PPR",
        pageSize=15
    )

    # Citation analysis
    key_papers = sar_papers[:10]
    for paper in key_papers:
        citation = tu.tools.openalex_search_works(
            query=paper['title'],
            limit=1
        )
        paper['citations'] = citation[0].get('cited_by_count', 0) if citation else 0

    return {
        'published_sar': sar_papers,
        'preprints': preprints,
        'high_impact_papers': sorted(key_papers, key=lambda x: x.get('citations', 0), reverse=True)
    }

---

Fallback Chains

Target ID Resolution

Primary: ChEMBL_search_targets
-> Fail -> GtoPdb_search_targets (for GPCR/ion channel/enzyme)
         -> Fail -> Document "Target not in databases"

Druggability Assessment

Primary: OpenTargets_get_target_tractability_by_ensemblID
-> Fail -> DGIdb_get_gene_druggability
         -> Fail -> Use target class as proxy

Bioactivity Data

Primary: ChEMBL_get_target_activities
-> Fail -> BindingDB_get_ligands_by_uniprot
         -> Fail -> GtoPdb_get_interactions
                  -> Fail -> PubChem_search_assays_by_target_gene
                           -> Fail -> Document "No bioactivity data"

Similarity Search

Primary: ChEMBL_search_similar_molecules
-> Fail -> PubChem_search_compounds_by_similarity
         -> Fail -> Document "Similarity search failed"

Structure Retrieval

Primary: get_protein_metadata_by_pdb_id
-> Fail (no PDB) -> EMDB_search_structures (for membrane proteins)
         -> Fail -> NvidiaNIM_alphafold2
                 -> Fail -> NvidiaNIM_esmfold
                         -> Fail -> alphafold_get_prediction
                                 -> Fail -> Document "No structural information"

Docking

Primary: NvidiaNIM_diffdock (have PDB + SDF)
-> Fail -> NvidiaNIM_boltz2 (from sequence + SMILES)
         -> Fail -> Skip docking, use similarity-based scoring

De Novo Generation

Primary: NvidiaNIM_genmol (specific position variation)
-> Fail -> NvidiaNIM_molmim (general analog generation)
         -> Fail -> Use similarity search only (no generation)

Literature Search

Primary: PubMed_search_articles (peer-reviewed)
-> Supplement: EuropePMC_search_articles (source='PPR' for preprints)
-> Supplement: openalex_search_works (citation analysis)

---

Batch Processing Pattern

# Define calls
calls = [
    {"name": "ChEMBL_get_molecule", "arguments": {"molecule_chembl_id": id}}
    for id in chembl_ids[:50]
]

# Execute in parallel
results = tu.run_batch(calls)

# Process results
for result in results:
    if result and 'molecule_structures' in result:
        process_molecule(result)

---

Rate Limiting Awareness

NVIDIA NIM Runtimes

API Key Check:

import os
if not os.environ.get("NVIDIA_API_KEY"):
    print("Warning: NVIDIA_API_KEY not set. NvidiaNIM tools unavailable.")