Drug Discovery

Agents

drug_discovery_expert

Parent orchestrator for drug discovery workflows. Handles target identification and scoring via Open Targets, compound search and bioactivity lookup via ChEMBL and PubChem, and delegates specialized analysis to three child agents.

Capabilities:

• Target search and identification from gene symbols or Ensembl IDs
• Composite druggability scoring (5-component weighted model)
• Compound search and bioactivity retrieval (IC50, Ki, EC50)
• Drug indication mapping from Open Targets
• Target-disease evidence retrieval
• Delegates molecular analysis to cheminformatics_expert
• Delegates clinical translation to clinical_dev_expert
• Delegates pharmacogenomics to pharmacogenomics_expert

cheminformatics_expert

Child agent for molecular analysis and property prediction. Invoked by drug_discovery_expert for compound characterization.

Capabilities:

• Molecular descriptor calculation (MW, LogP, TPSA, HBD, HBA, rotatable bonds)
• Lipinski Rule of Five compliance checking
• Fingerprint similarity computation (Morgan/ECFP)
• ADMET property prediction (absorption, distribution, metabolism, excretion, toxicity)
• 3D conformation generation (ETKDG + MMFF94)
• CAS-to-SMILES conversion via PubChem
• Structure similarity search
• Binding site identification from PDB content
• Side-by-side molecule comparison

clinical_dev_expert

Child agent for clinical translation and drug combination analysis. Invoked by drug_discovery_expert for downstream development assessment.

Capabilities:

• Target-disease evidence retrieval from Open Targets
• Drug combination synergy scoring (Bliss Independence, Loewe Additivity, HSA)
• Full dose-response combination matrix analysis
• Drug safety profiling and adverse event assessment
• Target tractability assessment (small molecule, antibody, PROTAC)
• Clinical trial phase data lookup
• Side-by-side drug candidate comparison

pharmacogenomics_expert

Child agent for drug-gene-variant interaction analysis. Invoked by drug_discovery_expert for pharmacogenomic profiling.

Capabilities:

• Protein mutation effect prediction using ESM2 (protein language model)
• Protein embedding extraction for wild-type vs mutant comparison
• Drug-variant interaction lookup from Open Targets
• Pharmacogenomic evidence retrieval from ChEMBL
• Variant impact scoring with combined clinical and drug context
• Expression-drug sensitivity correlation analysis
• Mutation frequency and co-occurrence pattern analysis

Example Workflows

Target Identification & Scoring

User: Score these targets for druggability: BRAF, EGFR, KRAS, TP53

[drug_discovery_expert]
- Queries Open Targets for each gene target
- Computes composite druggability scores (genetic association,
  known drugs, expression specificity, pathogenicity, literature)
- Ranks targets by weighted score
- Reports top candidates with confidence classification

Compound Profiling

User: Profile imatinib — get bioactivity data and ADMET properties

[drug_discovery_expert → cheminformatics_expert]
- Searches ChEMBL for imatinib, retrieves bioactivity (IC50, Ki)
- Delegates to cheminformatics_expert for molecular descriptors
- Computes Lipinski compliance and ADMET prediction
- Screens for PAINS and Brenk structural alerts
- Returns compound profile with druggability assessment

Drug Combination Synergy

User: Score synergy for my dose-response combination matrix

[drug_discovery_expert → clinical_dev_expert]
- Delegates to clinical_dev_expert with dose-response data
- Computes Bliss Independence, Loewe Additivity, and HSA scores
- Classifies each combination as synergistic, additive, or antagonistic
- Stores scored matrix with annotations in workspace

Pharmacogenomic Screening

User: Predict mutation effects for TP53 R248W and R273H variants

[drug_discovery_expert → pharmacogenomics_expert]
- Delegates to pharmacogenomics_expert with protein sequence
- Runs ESM2 fill-mask scoring for each mutation position
- Retrieves drug-variant interaction evidence from Open Targets
- Computes variant impact scores with clinical context
- Reports combined mutation effect and pharmacogenomic evidence

Dependencies

lobster-drug-discovery requires the following core libraries:

These are installed automatically with the package.

Optional Dependencies

Extended capabilities require optional extras:

# Core only (API-based tools)
pip install lobster-drug-discovery

# With RDKit molecular analysis
pip install "lobster-drug-discovery[chemistry]"

# With protein language models (ESM2)
pip install "lobster-drug-discovery[plm]"

# Full installation
pip install "lobster-drug-discovery[full]"

Tools that require optional dependencies degrade gracefully when those dependencies are not installed.

Services

lobster-drug-discovery includes domain-specific services bundled with the package:

Services are installed automatically with the agent package.

External APIs

This package queries public bioinformatics APIs. No authentication is required.

Sub-Agent Architecture

drug_discovery_expert (supervisor-accessible)
├── cheminformatics_expert (sub-agent, not directly accessible)
├── clinical_dev_expert (sub-agent, not directly accessible)
└── pharmacogenomics_expert (sub-agent, not directly accessible)

The drug_discovery_expert handles target identification, scoring, and compound lookup directly using shared tools (ChEMBL, Open Targets, PubChem). When the user requests molecular property analysis or ADMET prediction, it delegates to cheminformatics_expert. For clinical development assessment, synergy scoring, or safety profiling, it delegates to clinical_dev_expert. For drug-gene-variant interactions or mutation effect prediction, it delegates to pharmacogenomics_expert.

Configuration

# .lobster_workspace/config.toml
enabled = ["drug_discovery_expert"]

Acknowledgments

The drug discovery agent was inspired by the excellent work of the CellType CLI team (github.com/celltype/cli). Their approach to making computational biology accessible through intelligent CLI tooling helped shape the design philosophy behind this package. Thank you for pushing the field forward.