Drug Pipeline Search Skill

This skill converts natural language questions into structured API queries against a pharmaceutical drug database, then presents the results in a readable format.

Workflow

1. Parse user intent — Extract key entities from the user's question
2. Build query parameters — Map entities to the query schema below
3. Execute the query — Run scripts/search.py
4. Present results — Format and display drug records to the user

Step 1: Extract Keywords

Identify the following entity types from the user's question:

Dict field format:

{"logic": "or", "data": ["value1", "value2"]}

• logic controls how multiple values are combined: "or" (any match) or "and" (all must match). Default to "or" unless the user explicitly wants all terms to apply simultaneously.
• data is the list of keyword strings to match.

Type rules:

• company, indication, phase, location → plain List[str]
• drug_name, target, drug_modality, drug_feature, route_of_administration → dict with logic and data
• Default to page_num: 0, page_size: 10 unless the user specifies otherwise
• Prefer English keywords (the database is indexed in English); translate non-English terms
• drug_modality must use exact strings from this set:

```json

[

"Steroids", "Vaccine", "Antisense RNA", "Antibody-Drug Conjugates, ADCs", "Unknown", "Protein Degrader",

"Monoclonal Antibodies", "mRNA", "Others", "Cell-based Therapies", "Imaging Agents", "Gene Therapy",

"miRNA", "Polypeptide", "Recombinant Proteins", "Small Molecule", "siRNA/RNAi", "Trispecific Antibodies",

"Polyclonal Antibodies", "Bi-specific Antibodies", "Glycoconjugates", "Radiopharmaceutical",

"Nucleic Acid-based", "Carbohydrates"

]

```

• drug_feature must use exact strings from this set:

```json

[

"505b2", "Bacterial Product", "Biologic", "Biosimilar", "Device", "Fixed-Dose Combination", "Immuno-Oncology",

"New Molecular Entity (NME)", "Non-NME", "Precision Medicine", "Reformulation", "Specialty Drug", "Viral"

]

```

• route_of_administration must use exact strings from this set:

```json

[

"Intraarterial", "Intraurethral", "Inhaled", "Intranasal", "Subcutaneous (SQ) - Unspecified", "Transdermal",

"Intraocular/Subretinal/Subconjunctival", "Subcutaneous (SQ) Injection", "Intrauterine", "Intralymphatic",

"Intradiscal", "Intra-amniotic", "Intrathecal", "Intracerebral/cerebroventricular", "Intramuscular (IM)",

"Intraarticular", "Intracochlear", "Surgical Implantation", "Hemoperfusion", "Subcutaneous (SQ) Infusion",

"Intravitreal", "Intravenous (IV)", "Oral (PO)", "Intradermal", "Percutaneous Catheter/Injection",

"Intranodal", "Intravesical", "Intracameral", "Intratympanic", "Intratumoral",

"Sublingual (SL)/Oral Transmucosal", "Intravaginal", "N/A", "Rectal", "Intracavitary",

"Intra-Cisterna Magna (ICM) Injection", "Injectable - Unspecified", "Intratracheal", "Topical",

"Instillation", "Intraintestinal", "Submucosal"

]

```

Step 2: Execute the Query

python scripts/search.py --params '<JSON string>'

Or using a parameter file:

python scripts/search.py --params-file /tmp/query.json

Add --raw to receive the unformatted JSON response.

Step 3: Interpret Results

The response contains:

• total_count — total number of matching drugs
• results — current page of drug records, each with name, phase, modality, targets, companies, indication, development progress, etc.

Step 4: Review and Fallback Search Strategies

If no results are returned, apply the fallback strategies below before giving up.

When an initial query returns zero or poor results, try these strategies in order:

Strategy 1 — Drug Name Variant Expansion

Drug names in the database may use different formats (with/without hyphens, partial codes, aliases). Expand the drug_name field to include common variants and merge deduplicated results.

{
  "drug_name": {"logic": "or", "data": ["SHR-A1904", "SHR A1904", "A1904", "SHR1904"]},
  "page_num": 0,
  "page_size": 50
}

Common variant patterns to try:

• Remove or replace hyphens: SHR-A1904 → SHR A1904, SHRA1904
• Strip prefix/suffix: 9MW-2821 → MW-2821, 9MW2821
• Known alias: include trade names or INN alongside internal codes

Strategy 2 — Company-First with Application-Layer Filtering

When drug name matching is unreliable, use the company as the anchor. Fetch a broad set of the company's drugs, then filter by modality/indication/target in post-processing.

{
  "company": ["Roche", "Roche Inc"],
  "page_num": 0,
  "page_size": 500
}

After retrieving results, apply local filters:

• modality == "Monoclonal Antibodies"
• indication contains "breast cancer"
• drug_name matches known code pattern

Use this strategy when the drug code is ambiguous or the API match rate is low.

Strategy 3 — Broad Target/Modality Search with Post-Filtering

When neither name nor company is reliable, search by biological target and modality, then narrow results client-side.

{
  "target": {"logic": "or", "data": ["CLDN18.2", "Nectin-4", "HER2"]},
  "drug_modality": {"logic": "or", "data": ["Monoclonal Antibodies"]},
  "page_num": 0,
  "page_size": 200
}

After retrieval, filter by company name or drug code pattern using substring matching (e.g. code starts with SHR, 9MW, A166).

> Note: If the API supports regex, patterns like (SHR|9MW|A166) can be passed directly in drug_name.data to broaden matching in a single call.

Decision Tree

Initial query returns results?
├── Yes → present results
└── No  → Strategy 1: expand drug_name variants
          └── Still no results → Strategy 2: company anchor + local filter
                                 └── Still no results → Strategy 3: target/modality broad search
Any step hits HTTP 429?
└── Pause entire chain 30s → resume from current strategy
    (sleep ≥5s between every request to avoid triggering 429)

Conversion Examples

User: "Find PD-1 antibodies in Phase 3"

{
  "target": {"logic": "or", "data": ["PD-1"]},
  "drug_modality": {"logic": "or", "data": ["Monoclonal Antibodies"]},
  "phase": ["III"],
  "page_num": 0,
  "page_size": 30
}

User: "Roche bispecific antibodies for lung cancer"

{
  "company": ["Roche"],
  "drug_modality": {"logic": "or", "data": ["Bi-specific Antibodies"]},
  "indication": ["lung cancer"],
  "page_num": 0,
  "page_size": 30
}

User: "Oral small molecule KRAS G12C inhibitors"

{
  "target": {"logic": "or", "data": ["KRAS"]},
  "drug_modality": {"logic": "or", "data": ["Small Molecule"]},
  "route_of_administration": {"logic": "or", "data": ["Oral (PO)"]},
  "page_num": 0,
  "page_size": 30
}

User: "Drugs targeting both PD-1 and VEGF"

{
  "target": {"logic": "and", "data": ["PD-1", "VEGF"]},
  "page_num": 0,
  "page_size": 30
}

User: "Look up pembrolizumab"

{
  "drug_name": {"logic": "or", "data": ["pembrolizumab"]},
  "page_num": 0,
  "page_size": 30
}

Dependencies

• Python 3.8+
• requests library (pip install requests)
• Environment variable NOAH_API_TOKEN — API authentication token (required)
• Register for a free account at noah.bio to obtain your API key.

Security & Packaging Notes

• This skill only calls NoahAI official HTTPS endpoints under https://www.noah.bio/api/ and does not contact third-party services.
• It requires exactly one environment variable: NOAH_API_TOKEN. Store it in the environment or a local .env file, and never place it inline in commands, chats, or packaged files.
• The token is scoped to read medical public details only and cannot access private user records.
• The skill does not intentionally persist request parameters locally. Any server-side retention is determined by the NoahAI API service and its operational logging policies.
• It does not request persistent or system-level privileges and does not modify system configuration.
• The skill is source-file based (Python scripts only) and does not require runtime installs, package downloads, or external bootstrap steps.