Disease Innovative Drug Intelligence Integration#

Overview#

Convert natural-language questions such as "What noteworthy new drugs have recently emerged for Alzheimer's disease?" into executable multi-database query plans. Generate a decision-oriented English comprehensive report rather than a simple list of drug names.

Quick Start#

1. Identify whether the request belongs to the disease_to_drug scenario.
2. Standardize the disease entity and decompose the "innovative drug" intent.
3. Prefer the local local_tools/ code for database queries.
4. Perform entity normalization, evidence integration, innovation filtering, and tiered output.
5. Generate the final report and annotate evidence boundaries.

For detailed data structures, routing rules, scoring logic, and the report template, see disease_to_drug_playbook.md.

Local Execution Rules#

This skill no longer depends on a BioDB MCP HTTP service. For any ChEMBL, ClinicalTrials, or Search call, only use the local Python code bundled in the current skill directory:

• local_tools/chembl_api.py
• local_tools/clinicaltrials_api.py
• local_tools/search_api.py
• local_tools/run_tool.sh

Recommended commands:

bash

bash local_tools/run_tool.sh chembl_api.py search_target EGFR
bash local_tools/run_tool.sh chembl_api.py search_molecule osimertinib
bash local_tools/run_tool.sh chembl_api.py get_drug_by_id CHEMBL3545063
bash local_tools/run_tool.sh clinicaltrials_api.py get_studies --query-cond "lung cancer" --fields NCTId BriefTitle OverallStatus
bash local_tools/run_tool.sh search_api.py "latest EGFR inhibitor approval"

Execution constraints:

• Prefer importing local local_tools/ modules or executing them through bash local_tools/run_tool.sh .... Do not assume services such as http://127.0.0.1:8086 exist.
• Do not depend on bare python commands directly. Always use run_tool.sh to resolve the available interpreter. run_tool.sh prefers python3 and only falls back to python if needed.
• If Search is needed, confirm that langchain_tavily is installed and the TAVILY_API_KEY environment variable is set.
• The current database capability set only includes ChEMBL, ClinicalTrials, and Search. Ignore any other database descriptions that are not implemented in this skill.
• Search may only be executed via bash local_tools/run_tool.sh search_api.py ... or SearchAPI.run(query). Do not bypass the local tool and call an external web search directly.
• Only when the user explicitly asks for external web search and also states that the local Search tool is unavailable or insufficient may external web search be used as a final fallback. Otherwise it is forbidden.

Trigger and Detection#

Trigger this skill when the user question contains both of the following:

• A disease entity, such as diabetes, lung cancer, Alzheimer's disease, obesity, NASH, or RA.
• An innovative-drug intent, such as innovative drugs, novel-mechanism drugs, pipeline drugs, cutting-edge drugs, or noteworthy new drugs.

If the request is overly broad, such as "innovative drugs for cancer," first suggest narrowing the disease type. If the user does not want to narrow it, default to a top cancer types plus top mechanisms overview.

Workflow#

Step 0 Task Structuring#

Construct a task object such as:

json

{
  "task_type": "disease_to_drug",
  "focus": "innovative_drugs",
  "disease_raw": "diabetes",
  "time_constraint": null,
  "region_constraint": null,
  "stage_constraint": null
}

Step 1 Disease Normalization#

Output canonical_disease, subtypes, aliases, and preferred_query_terms. If the user does not specify a subtype, first analyze the overall disease, then emphasize more active R&D subtypes, for example prioritizing T2DM under diabetes.

Step 2 Innovative Drug Mapping#

Map "innovative drugs" into executable criteria:

• New mechanisms or new targets, including first-in-class tendency
• Representative recently approved drugs
• Mid-to-late stage pipeline candidates, with Phase II/III preferred
• Frontier directions such as dual or multi-target drugs and next-generation optimized molecules

Step 3 Subtask Decomposition#

Always execute these five subtasks:

• identify_targets_and_mechanisms
• retrieve_representative_drugs
• build_drug_profiles
• validate_clinical_progress
• summarize_trends

Step 4 Database Execution Order#

Default order:

1. Mechanisms and targets: ChEMBL(target/mechanism)
2. Drug candidates: ChEMBL(molecule/drug/indication)
3. Clinical progress: ClinicalTrials
4. Web fallback: Search, only through local search_api.py, and only when ChEMBL and ClinicalTrials evidence is insufficient or recent updates need verification

Notes:

• This skill no longer depends on a BioDB MCP service.
• ChEMBL, ClinicalTrials, and Search have been switched to local adapter code.
• The stable execution scope of this skill only covers these three capabilities.

Step 5 Evidence Integration and Deduplication#

Primary keys in priority order:

• Drug: ChEMBL ID > standard drug name > ClinicalTrials intervention
• Target: gene symbol or standard target name > alias

Always preserve aliases and dosage-form information to avoid incorrect merges, such as different semaglutide formulations.

Step 6 Innovation Filtering and Ranking#

Score from 0 to 5 and rank comprehensively across:

• disease_relevance
• innovation
• clinical_maturity
• evidence_strength
• representativeness

Output must be tiered into:

• Approved or clinically validated representative innovative drugs
• Mid-to-late stage pipeline candidates
• Frontier exploratory mechanism directions

Step 7 Report Generation#

Before generating the report, first read ## 10. English Report Template in references/disease_to_drug_playbook.md.

By default, the final report must strictly follow that template. Do not produce a free-form answer that merely "covers the same topics." The section order, first-level numbering, and main title skeleton must remain unchanged:

• {Disease Name} Innovative Drug Analysis Report
• 1. Problem Overview
• 2. Executive Summary
• 3. Key Disease-Related Targets and Mechanisms
• 4. Representative Innovative Drug List
• 5. Clinical Trial Progress Overview
• 6. R&D Trends and Assessment
• 7. Result Notes and Limitations

Only if the user explicitly asks for a brief version, summary version, table version, or another specific format may you deviate from the standard template. Otherwise you must use it.

Step 8 Exception Handling#

• Too many results: use representativeness plus innovation to select Top N, default 10.
• Too few results: prioritize target directions and adjacent mechanisms rather than forcing a drug list.
• Conflicting evidence: clearly state that molecular evidence exists while clinical evidence is limited.
• Missing constraints: default to time_constraint=null and region_constraint=global, then state this explicitly in the report.

Tool-Usage Constraints#

• Run the main path first, meaning local ChEMBL plus local ClinicalTrials, and only then supplement. Do not reverse the order.
• If key fields such as phase, status, or target are missing, a supplementary query must be triggered.
• Every key conclusion in the report must have at least one traceable evidence item with database name plus entity primary key.
• Do not treat "innovative drugs" as a regulatory term. It is an information-integration term and must be declared as such in the result notes.
• For the minimum required ChEMBL capability set, prefer search_target, search_molecule, get_drug_by_id, get_molecule_by_id, get_target_by_id, get_mechanism, and get_drug_indication.
• For the minimum required ClinicalTrials capability set, prefer get_studies and get_study.
• For the minimum required Search capability set, always use SearchAPI.run(query). If dependencies or API keys are missing, state in the result that supplementary retrieval could not be executed, and do not automatically switch to external web search.
• When executing local tools from the command line, always use bash local_tools/run_tool.sh <tool.py> .... Do not write python <tool.py> ... directly.
• Do not reintroduce KEGG, UniProt, STRING, Ensembl, PubChem, PDB, or any other database instructions not included in the current code layer.
• Before output, self-check whether the report sections match ## 10. English Report Template exactly. If they do not, rewrite before sending.

Quality Checklist#

• Was disease normalization completed, including aliases and subtypes?
• Does the report provide a mechanism-drug-clinical three-layer evidence chain?
• Were entity normalization and alias deduplication completed?
• Is the output tiered into approved, mid-to-late stage, and frontier directions?
• Are limitations, conflicts, and uncertainty clearly stated?

Reference File#

• disease_to_drug_playbook.md: Full SOP, routing strategy, internal JSON structures, and the English report template.