# Deep-Pharma Multimodal Intelligence - DataHack 3.0 — Data

来源：https://www.kaggle.com/competitions/deep-pharma-multimodal-intelligence-challenge-data-hack-3-0/data

采集时间：2026-10-19T09:59:59Z

# 📂 Dataset Description

The dataset consists of drug pairs enriched with extensive pharmacological, chemical, and clinical text data. The goal is to predict the **Severity**, **Side Effects**, and **PRR Risk** of the interaction between **Drug A** and **Drug B**.

## File Descriptions

### `train.csv`

The main training set containing drug pairs, their rich feature set, and ground truth labels.

* **Rows:** 15,284 interaction pairs.
* **Columns:** Includes Identifiers, Features (A & B), and Targets.

### `test.csv`

The test set containing **3,974** drug pairs.

* Contains all feature columns found in Train (`Drug_A_Name`, `SMILES_A`, etc.).
* **Does not contain** the Target columns.

---

## 🧬 Feature Column Details

The dataset provides dual features for every interaction: one set for **Drug A** (suffix `_A`) and one for **Drug B** (suffix `_B`).

### 1. Identifiers & Chemistry

* `Pair_ID`: Unique identifier for the interaction row.
* `Drug_A_Name`, `Drug_B_Name`: Generic international non-proprietary name (INN).
* `SMILES_A`, `SMILES_B`: Simplified Molecular Input Line Entry System. A text string representing the exact 2D chemical structure and topology of the molecule.

### 2. Pharmacology (Mechanism)

High-dimensional text fields describing how the drug works biologically.

* `Mechanism_A`, `Mechanism_B`: Description of the specific molecular targets (receptors, enzymes) the drug activates or inhibits.
* `Pharmacodynamics_A`, `Pharmacodynamics_B`: Describes the biochemical and physiologic effects of the drug on the body.

### 3. Pharmacokinetics (ADME)

Data describing how the drug moves through the body (Absorption, Distribution, Metabolism, Excretion). These fields may contain mixed text and numeric data.

* `Absorption_A`, `Absorption_B`: Details on bioavailability and absorption rates.
* `Metabolism_A`, `Metabolism_B`: General description of metabolic breakdown locations (e.g., Liver).
* `Elimination_Route_A`, `Elimination_Route_B`: How the drug exits the body (Renal, Biliary, etc.).
* `Half_Life_A`, `Half_Life_B`: The time required for the drug's concentration in the body to reduce by half.
* `Protein_Binding_A`, `Protein_Binding_B`: Percentage of the drug bound to plasma proteins (affects distribution).

### 4. Metabolic Enzymes (Crucial for DDIs)

* `CYP450_Enzymes_A`, `CYP450_Enzymes_B`: Specific Cytochrome P450 isozymes involved in the drug's metabolism (e.g., "CYP3A4, CYP2D6"). *Note: Overlap in these fields is a strong predictor of metabolic interactions.*

### 5. Clinical & Safety Profile

* `Indication_A`, `Indication_B`: The approved diseases or conditions the drug treats.
* `Warning_A`, `Warning_B`: FDA Black Box warnings or general precautions.
* `Toxicity_A`, `Toxicity_B`: Known toxic effects, LD50 values, or overdose symptoms.

---

## 🎯 Target Variables (Train Only)

### 1. Severity (Classification)

* **`Severity`**: The clinical impact of the interaction.
* Classes: `Major`, `Moderate`, `Minor`.



### 2. Specific Side Effects (Multi-Label)

* **`Target_Binary_[SideEffect]`** (50 columns):
* `1`: The side effect occurs.
* `0`: The side effect does not occur.



### 3. Risk Quantification (Regression)

* **`Target_PRR_[SideEffect]`** (50 columns):
* **Proportional Reporting Ratio (PRR)**: A float value indicating the strength of the statistical signal.
* Values `> 0` indicate a signal.
* Values `== 0` indicate no signal/no side effect.