付録: 参考文献・リソース集

┌─────────────────┐
│   AI Model      │ ← LLM、Agent、推論モデル
│ (GPT, Claude,   │   (GitHub Copilotなど)
│  Gemini, etc.)  │
└─────────┬───────┘
          │ AI-Tool Interaction Protocol
          │
┌─────────▼───────┐
│  ToolUniverse   │ ← プロトコル実装レイヤー
│   Ecosystem     │
└─────────┬───────┘
          │
┌─────────▼───────┐
│  1,000+ Tools   │ ← MLモデル、API、データベース等
│ (Heterogeneous  │   (バックエンドの違いを吸収)
│   Backends)     │
└─────────────────┘

{
    "name": "UniProt_get_entry_by_accession",
    "description": "Get complete JSON entry for a UniProtKB accession",
    "parameters": {
        "type": "object",
        "properties": {
            "accession": {
                "type": "string",
                "description": "UniProtKB accession (e.g., P05067)",
                "required": true
            }
        }
    },
    "return_schema": {
        "type": "object",
        "description": "Complete UniProtKB entry with protein information"
    }
}

{
    "name": "Tool_identifier",
    "arguments": {
        "parameter1": "value1",
        "parameter2": "value2"
    }
}

# タンパク質情報の取得（REST API: UniProt）
{"name": "UniProt_get_entry_by_accession", "arguments": {"accession": "P05067"}}

# 薬物安全性の分析（データベース: FAERS）
{"name": "FAERS_count_reactions_by_drug_event", "arguments": {"medicinalproduct": "aspirin"}}

# MLモデルによるドッキング予測（機械学習: Boltz-2）
{"name": "boltz2_docking", "arguments": {"protein_id": "1ABC", "ligand_smiles": "CCO"}}

from tooluniverse import ToolUniverse

tu = ToolUniverse()
tu.load_tools()

# 統一インターフェイスで直接実行
result = tu.run({
    "name": "OpenTargets_get_associated_targets_by_disease_efoId",
    "arguments": {"efoId": "EFO_0000537"}  # hypertension
})

# MCPサーバーを起動
tooluniverse-smcp --port 8000

# Claude、GitHub Copilot等のAIアシスタントがMCP経由で接続
# → ToolUniverse上のすべてのツールが利用可能に

# AIモデルがツールを発見する例
query = "predict protein binding affinity"

# プロトコルが関連ツールを返す
tools_found = [
    "boltz2_docking",
    "ADMETAI_predict_properties",
    "ChEMBL_search_similar_molecules"
]

# ツール実行
result = tu.run({
    "name": "boltz2_docking",
    "arguments": {
        "protein_id": "P05067",
        "ligand_smiles": "CC(=O)OC1=CC=CC=C1C(=O)O"
    }
})

# 一貫した構造のレスポンス
{
    "binding_affinity": -8.2,
    "binding_probability": 0.85,
    "confidence_score": 0.92,
    "metadata": {
        "model_version": "boltz-2",
        "execution_time": "2.3s"
    }
}

# 不正な入力の例
{"name": "UniProt_get_entry_by_accession", "arguments": {"accession": "INVALID_ID"}}

# 構造化されたエラーレスポンス
{
    "status": "error",
    "error_type": "ValidationError",
    "message": "Invalid UniProt accession format",
    "details": {
        "parameter": "accession",
        "expected_format": "P12345 or Q9Y261",
        "received": "INVALID_ID"
    }
}

# 合成ワークフローの例
workflow = {
    "name": "drug_discovery_pipeline",
    "arguments": {
        "disease": "hypercholesterolemia",
        "steps": [
            "target_identification",
            "compound_screening",
            "ADMET_prediction",
            "patent_analysis"
        ]
    }
}

# 専門家への相談リクエスト
{
    "name": "consult_human_expert",
    "arguments": {
        "question": "Which HMG-CoA reductase inhibitor shows best safety profile?",
        "context": {"compounds": ["lovastatin", "pravastatin", "simvastatin"]},
        "expertise_required": "pharmacology"
    }
}

# 外部MCPサーバーのツールを登録
{
    "name": "register_mcp_tools",
    "arguments": {
        "server_url": "mcp://expert-system.company.com:8080",
        "tool_categories": ["proprietary_ml_models", "private_databases"]
    }
}