...

Makefile

@@ -68,15 +68,38 @@ clean_data: requirements
 data: requirements
 	$(PYTHON_INTERPRETER) scripts/process_data.py
 
-## Train model
+## Process external data for pretrain (external -> processed)
+.PHONY: data_pretrain
+data_pretrain: requirements
+	$(PYTHON_INTERPRETER) scripts/process_external.py
+
+# MPNN 支持：使用 USE_MPNN=1 启用 MPNN encoder
+# 例如：make pretrain USE_MPNN=1
+MPNN_FLAG = $(if $(USE_MPNN),--use-mpnn,)
+
+# Backbone 冻结：使用 FREEZE_BACKBONE=1 冻结 backbone，只训练 heads
+# 例如：make finetune FREEZE_BACKBONE=1
+FREEZE_FLAG = $(if $(FREEZE_BACKBONE),--freeze-backbone,)
+
+## Pretrain on external data (delivery only)
+.PHONY: pretrain
+pretrain: requirements
+	$(PYTHON_INTERPRETER) -m lnp_ml.modeling.pretrain $(MPNN_FLAG)
+
+## Train model (multi-task, from scratch)
 .PHONY: train
 train: requirements
-	$(PYTHON_INTERPRETER) -m lnp_ml.modeling.train
+	$(PYTHON_INTERPRETER) -m lnp_ml.modeling.train $(MPNN_FLAG)
+
+## Finetune from pretrained checkpoint (use FREEZE_BACKBONE=1 to freeze backbone)
+.PHONY: finetune
+finetune: requirements
+	$(PYTHON_INTERPRETER) -m lnp_ml.modeling.train --init-from-pretrain models/pretrain_delivery.pt $(FREEZE_FLAG) $(MPNN_FLAG)
 
 ## Train with hyperparameter tuning
 .PHONY: tune
 tune: requirements
-	$(PYTHON_INTERPRETER) -m lnp_ml.modeling.train --tune
+	$(PYTHON_INTERPRETER) -m lnp_ml.modeling.train --tune $(MPNN_FLAG)
 
 ## Run predictions
 .PHONY: predict

README.md

@@ -4,7 +4,106 @@
     <img src="https://img.shields.io/badge/CCDS-Project%20template-328F97?logo=cookiecutter" />
 </a>
 
-A short description of the project.
+LNP（脂质纳米颗粒）药物递送性能预测模型。
+
+## 快速开始
+
+### 1. 安装环境
+
+```bash
+pixi install
+pixi shell
+```
+
+### 2. 数据处理
+
+```bash
+# 清洗原始数据 (raw -> interim)
+make clean_data
+
+# 处理内部数据集 (interim -> processed)
+make data
+
+# 处理外部预训练数据 (external -> processed)
+make data_pretrain
+```
+
+### 3. 训练模型
+
+**方式 A：直接训练（从零开始）**
+
+```bash
+make train
+```
+
+**方式 B：预训练 + 微调（推荐）**
+
+利用外部 LiON 数据集（约 9000 条）进行预训练，再在内部数据上微调：
+
+```bash
+# Step 1: 处理外部数据
+make data_pretrain
+
+# Step 2: 在外部数据上预训练 backbone + delivery head
+make pretrain
+
+# Step 3: 加载预训练权重，在内部数据上多任务微调
+make finetune
+```
+
+**方式 C：超参数调优**
+
+```bash
+make tune
+```
+
+### 4. 测试与预测
+
+```bash
+# 在测试集上评估
+make test
+
+# 生成预测结果
+make predict
+```
+
+## 训练流程详解
+
+### 预训练 (Pretrain)
+
+在外部 LiON 数据上，仅训练 `quantified_delivery` 任务：
+
+```bash
+# 1. 先处理外部数据
+python scripts/process_external.py
+
+# 2. 预训练
+python -m lnp_ml.modeling.pretrain \
+    --train-path data/processed/train_pretrain.parquet \
+    --val-path data/processed/val_pretrain.parquet \
+    --epochs 50 \
+    --lr 1e-4
+```
+
+产出:
+- `data/processed/train_pretrain.parquet`: 处理后的训练数据
+- `data/processed/val_pretrain.parquet`: 处理后的验证数据
+- `models/pretrain_delivery.pt`: backbone + delivery head 权重
+- `models/pretrain_history.json`: 训练历史
+
+### 微调 (Finetune)
+
+加载预训练权重，在内部多任务数据上训练：
+
+```bash
+python -m lnp_ml.modeling.train \
+    --init-from-pretrain models/pretrain_delivery.pt \
+    --load-delivery-head  # 可选：是否加载 delivery head 权重
+```
+
+产出:
+- `models/model.pt`: 完整模型权重
+- `models/history.json`: 训练历史
 
 ## Project Organization
 

lnp_ml/dataset.py

@@ -346,3 +346,185 @@ def load_dataset(
         LNPDataset(val_df, config),
         LNPDataset(test_df, config),
     )
+
+
+# ============ 外部数据（仅 delivery）处理 ============
+
+# 外部数据中 Value_name 的值映射（空格 -> 下划线）
+EXTERNAL_VALUE_NAME_MAP = {
+    "log luminescence": "log_luminescence",
+    "Luminescence": "luminescence",
+    "FFL silencing": "FFL_silencing",
+    "Peptide abundance": "Peptide_abundance",
+    "hEPO": "hEPO",
+    "FVII silencing": "FVII_silencing",
+    "GFP delivery": "GFP_delivery",
+    "Discretized luminescence": "Discretized_luminescence",
+}
+
+# 外部数据中 Mix_type 的值映射
+EXTERNAL_MIX_TYPE_MAP = {
+    "Hand": "Pipetting",  # 外部数据用 "Hand" 表示 "Pipetting"
+    "Microfluidic": "Microfluidic",
+    "Pipetting": "Pipetting",
+}
+
+
+def process_external_dataframe(df: pd.DataFrame) -> pd.DataFrame:
+    """
+    处理外部 LiON 数据的 DataFrame，对齐到模型所需的列格式。
+    
+    与 process_dataframe 类似，但针对外部数据的列名差异做适配：
+    - Value_name 值中的空格需要转为下划线
+    - Mix_type 中 "Hand" 需要映射为 "Pipetting"
+    """
+    df = df.copy()
+    
+    # 1. 预处理：映射 Value_name 和 Mix_type 的值
+    if "Value_name" in df.columns:
+        df["Value_name"] = df["Value_name"].map(
+            lambda x: EXTERNAL_VALUE_NAME_MAP.get(x, x) if pd.notna(x) else x
+        )
+    
+    if "Mix_type" in df.columns:
+        df["Mix_type"] = df["Mix_type"].map(
+            lambda x: EXTERNAL_MIX_TYPE_MAP.get(x, x) if pd.notna(x) else x
+        )
+    
+    # 2. 处理 phys token 的 one-hot 列（如果不存在则从原始列生成）
+    for col, values in PHYS_ONEHOT_SPECS.items():
+        for v in values:
+            onehot_col = f"{col}_{v}"
+            if onehot_col not in df.columns:
+                if col in df.columns:
+                    df[onehot_col] = (df[col] == v).astype(float)
+                else:
+                    df[onehot_col] = 0.0
+    
+    # 3. 处理 exp token 的 one-hot 列
+    # 外部数据部分列已存在（如 Model_type_*, Delivery_target_* 等），但可能缺失某些类别
+    for col, values in EXP_ONEHOT_SPECS.items():
+        for v in values:
+            onehot_col = f"{col}_{v}"
+            if onehot_col not in df.columns:
+                if col in df.columns:
+                    df[onehot_col] = (df[col] == v).astype(float)
+                else:
+                    df[onehot_col] = 0.0
+            else:
+                # 确保是 float 类型
+                df[onehot_col] = df[onehot_col].fillna(0.0).astype(float)
+    
+    # 4. 确保 comp 列存在且为 float
+    for col in COMP_COLS:
+        if col in df.columns:
+            df[col] = pd.to_numeric(df[col], errors="coerce").fillna(0.0)
+        else:
+            df[col] = 0.0
+    
+    # 5. 确保 help 列存在
+    for col in HELP_COLS:
+        if col not in df.columns:
+            df[col] = 0.0
+        else:
+            df[col] = df[col].fillna(0.0).astype(float)
+    
+    # 6. 处理 quantified_delivery
+    if "quantified_delivery" in df.columns:
+        df["quantified_delivery"] = pd.to_numeric(df["quantified_delivery"], errors="coerce")
+    
+    return df
+
+
+class ExternalDeliveryDataset(Dataset):
+    """
+    外部 LiON 数据集，仅用于 delivery 预训练。
+    
+    返回:
+        - smiles: str
+        - tabular: Dict[str, Tensor] with keys "comp", "phys", "help", "exp"
+        - targets: Dict[str, Tensor] with key "delivery"
+        - mask: Dict[str, Tensor] with key "delivery"
+    """
+    
+    def __init__(
+        self,
+        df: pd.DataFrame,
+        config: Optional[LNPDatasetConfig] = None,
+    ):
+        self.config = config or LNPDatasetConfig()
+        self.df = process_external_dataframe(df)
+        
+        # 提取数据
+        self.smiles = self.df[SMILES_COL].tolist()
+        
+        # Tabular features
+        self.comp = self.df[self.config.comp_cols].values.astype(np.float32)
+        self.phys = self.df[self.config.phys_cols].values.astype(np.float32)
+        self.help = self.df[self.config.help_cols].values.astype(np.float32)
+        self.exp = self.df[self.config.exp_cols].values.astype(np.float32)
+        
+        # 只有 delivery 作为 target
+        self.delivery = self.df["quantified_delivery"].values.astype(np.float32) if "quantified_delivery" in self.df.columns else None
+    
+    def __len__(self) -> int:
+        return len(self.smiles)
+    
+    def __getitem__(self, idx: int) -> Dict:
+        item = {
+            "smiles": self.smiles[idx],
+            "tabular": {
+                "comp": torch.from_numpy(self.comp[idx]),
+                "phys": torch.from_numpy(self.phys[idx]),
+                "help": torch.from_numpy(self.help[idx]),
+                "exp": torch.from_numpy(self.exp[idx]),
+            },
+            "targets": {},
+            "mask": {},
+        }
+        
+        # delivery target and mask
+        if self.delivery is not None:
+            item["targets"]["delivery"] = torch.tensor(self.delivery[idx], dtype=torch.float32)
+            item["mask"]["delivery"] = torch.tensor(not np.isnan(self.delivery[idx]), dtype=torch.bool)
+        
+        return item
+
+
+def load_external_dataset(
+    path: Path,
+    train_ratio: float = 0.9,
+    seed: int = 42,
+) -> Tuple[ExternalDeliveryDataset, ExternalDeliveryDataset]:
+    """
+    加载外部 LiON 数据集并划分为 train/val。
+    
+    Args:
+        path: CSV 文件路径
+        train_ratio: 训练集比例（剩余为验证集）
+        seed: 随机种子
+        
+    Returns:
+        (train_dataset, val_dataset)
+    """
+    df = pd.read_csv(path)
+    
+    # 过滤掉 quantified_delivery 为空的行
+    if "quantified_delivery" in df.columns:
+        df = df[df["quantified_delivery"].notna()].reset_index(drop=True)
+    
+    # 随机打乱
+    df = df.sample(frac=1, random_state=seed).reset_index(drop=True)
+    
+    n = len(df)
+    n_train = int(n * train_ratio)
+    
+    train_df = df.iloc[:n_train]
+    val_df = df.iloc[n_train:]
+    
+    config = LNPDatasetConfig()
+    
+    return (
+        ExternalDeliveryDataset(train_df, config),
+        ExternalDeliveryDataset(val_df, config),
+    )

lnp_ml/featurization/smiles.py

@@ -43,7 +43,13 @@ class RDKitFeaturizer:
         return np.array(MACCSkeys.GenMACCSKeys(mol).ToList(), dtype=np.float32)
     
     def _encode_desc(self, mol: Chem.Mol) -> np.ndarray:
-        return np.array(list(Descriptors.CalcMolDescriptors(mol).values()), dtype=np.float32)
+        # 使用 float64 计算，然后 clip 到 float32 范围，避免 overflow
+        desc_values = list(Descriptors.CalcMolDescriptors(mol).values())
+        arr = np.array(desc_values, dtype=np.float64)
+        # 替换 inf/nan，clip 到 float32 范围
+        arr = np.nan_to_num(arr, nan=0.0, posinf=1e10, neginf=-1e10)
+        arr = np.clip(arr, -1e10, 1e10)
+        return arr.astype(np.float32)
 
     def _encode_one(self, smiles: str) -> Dict[str, np.ndarray]:
         mol = Chem.MolFromSmiles(smiles)

lnp_ml/modeling/models.py

@@ -121,28 +121,16 @@ class LNPModel(nn.Module):
             dropout=dropout,
         )
 
-    def forward(
+    def _encode_and_project(
         self,
         smiles: List[str],
         tabular: Dict[str, torch.Tensor],
-    ) -> Dict[str, torch.Tensor]:
+    ) -> torch.Tensor:
         """
-        Args:
-            smiles: SMILES 字符串列表，长度为 B
-            tabular: Dict[str, Tensor]，包含:
-                - "comp": [B, 5] 配方比例
-                - "phys": [B, 12] 物理参数
-                - "help": [B, 4] Helper lipid
-                - "exp": [B, 32] 实验条件
+        内部方法：编码 SMILES 和 tabular，返回 stacked tokens。
         
         Returns:
-            Dict[str, Tensor]:
-                - "size": [B, 1]
-                - "pdi": [B, 4]
-                - "ee": [B, 3]
-                - "delivery": [B, 1]
-                - "biodist": [B, 7]
-                - "toxic": [B, 2]
+            stacked: [B, n_tokens, d_model]
         """
         # 1. Encode SMILES
         rdkit_features = self.rdkit_encoder(smiles)  # {"morgan", "maccs", "desc"}
@@ -170,23 +158,87 @@ class LNPModel(nn.Module):
         projected = self.token_projector(all_features)  # Dict[str, [B, d_model]]
 
         # 4. Stack tokens: [B, n_tokens, d_model]
-        # 按顺序 stack：Channel A (化学) + Channel B (配方/实验)
         if self.use_mpnn:
             token_order = ["mpnn", "morgan", "maccs", "desc", "comp", "phys", "help", "exp"]
         else:
             token_order = ["morgan", "maccs", "desc", "comp", "phys", "help", "exp"]
 
         stacked = torch.stack([projected[k] for k in token_order], dim=1)
+        return stacked
 
-        # 5. Cross Modal Attention
+    def forward_backbone(
+        self,
+        smiles: List[str],
+        tabular: Dict[str, torch.Tensor],
+    ) -> torch.Tensor:
+        """
+        Backbone forward：编码 -> 投影 -> 注意力 -> 融合，不经过任务头。
+        
+        用于 pretrain 阶段或需要提取特征的场景。
+        
+        Args:
+            smiles: SMILES 字符串列表，长度为 B
+            tabular: Dict[str, Tensor]
+            
+        Returns:
+            fused: [B, fusion_dim] 融合后的特征向量
+        """
+        # 编码 + 投影 + stack
+        stacked = self._encode_and_project(smiles, tabular)
+        
+        # Cross Modal Attention
         attended = self.cross_attention(stacked)
         
-        # 6. Fusion
+        # Fusion
         fused = self.fusion(attended)
         
-        # 7. Multi-Task Head
-        outputs = self.head(fused)
+        return fused
 
+    def forward_delivery(
+        self,
+        smiles: List[str],
+        tabular: Dict[str, torch.Tensor],
+    ) -> torch.Tensor:
+        """
+        仅预测 delivery（用于 pretrain）。
+        
+        Args:
+            smiles: SMILES 字符串列表，长度为 B
+            tabular: Dict[str, Tensor]
+            
+        Returns:
+            delivery: [B, 1] 预测的 delivery 值
+        """
+        fused = self.forward_backbone(smiles, tabular)
+        return self.head.delivery_head(fused)
+
+    def forward(
+        self,
+        smiles: List[str],
+        tabular: Dict[str, torch.Tensor],
+    ) -> Dict[str, torch.Tensor]:
+        """
+        完整的多任务 forward。
+        
+        Args:
+            smiles: SMILES 字符串列表，长度为 B
+            tabular: Dict[str, Tensor]，包含:
+                - "comp": [B, 5] 配方比例
+                - "phys": [B, 12] 物理参数
+                - "help": [B, 4] Helper lipid
+                - "exp": [B, 32] 实验条件
+
+        Returns:
+            Dict[str, Tensor]:
+                - "size": [B, 1]
+                - "pdi": [B, 4]
+                - "ee": [B, 3]
+                - "delivery": [B, 1]
+                - "biodist": [B, 7]
+                - "toxic": [B, 2]
+        """
+        fused = self.forward_backbone(smiles, tabular)
+        outputs = self.head(fused)
         return outputs
 
     def clear_cache(self) -> None:
@@ -195,6 +247,69 @@ class LNPModel(nn.Module):
         if self.mpnn_encoder is not None:
             self.mpnn_encoder.clear_cache()
 
+    def get_backbone_state_dict(self) -> Dict[str, torch.Tensor]:
+        """
+        获取 backbone 部分的 state_dict（不含任务头）。
+        
+        包含: token_projector, cross_attention, fusion
+        """
+        backbone_keys = []
+        for name in self.state_dict().keys():
+            if name.startswith(("token_projector.", "cross_attention.", "fusion.")):
+                backbone_keys.append(name)
+        
+        return {k: v for k, v in self.state_dict().items() if k in backbone_keys}
+
+    def get_delivery_head_state_dict(self) -> Dict[str, torch.Tensor]:
+        """获取 delivery head 的 state_dict"""
+        return {
+            k: v for k, v in self.state_dict().items()
+            if k.startswith("head.delivery_head.")
+        }
+
+    def load_pretrain_weights(
+        self,
+        pretrain_state_dict: Dict[str, torch.Tensor],
+        load_delivery_head: bool = True,
+        strict: bool = False,
+    ) -> None:
+        """
+        从预训练 checkpoint 加载 backbone 和（可选）delivery head 权重。
+        
+        Args:
+            pretrain_state_dict: 预训练模型的 state_dict
+            load_delivery_head: 是否加载 delivery head 权重
+            strict: 是否严格匹配（默认 False，允许缺失/多余的键）
+        """
+        # 筛选要加载的参数
+        keys_to_load = []
+        for name in pretrain_state_dict.keys():
+            # Backbone 部分
+            if name.startswith(("token_projector.", "cross_attention.", "fusion.")):
+                keys_to_load.append(name)
+            # Delivery head（可选）
+            elif load_delivery_head and name.startswith("head.delivery_head."):
+                keys_to_load.append(name)
+        
+        filtered_state_dict = {k: v for k, v in pretrain_state_dict.items() if k in keys_to_load}
+        
+        # 加载权重
+        missing, unexpected = [], []
+        model_state = self.state_dict()
+        for k, v in filtered_state_dict.items():
+            if k in model_state:
+                if model_state[k].shape == v.shape:
+                    model_state[k] = v
+                else:
+                    unexpected.append(f"{k} (shape mismatch: {model_state[k].shape} vs {v.shape})")
+            else:
+                unexpected.append(k)
+        
+        self.load_state_dict(model_state, strict=False)
+        
+        if strict and (missing or unexpected):
+            raise RuntimeError(f"Missing keys: {missing}, Unexpected keys: {unexpected}")
+
 
 class LNPModelWithoutMPNN(LNPModel):
     """不使用 MPNN 的简化版本"""

lnp_ml/modeling/predict.py

@@ -1,7 +1,7 @@
 """预测脚本：使用训练好的模型进行推理"""
 
 from pathlib import Path
-from typing import Dict, List
+from typing import Dict, List, Union
 
 import pandas as pd
 import torch
@@ -11,17 +11,54 @@ import typer
 
 from lnp_ml.config import MODELS_DIR, PROCESSED_DATA_DIR
 from lnp_ml.dataset import LNPDataset, collate_fn
-from lnp_ml.modeling.models import LNPModelWithoutMPNN
+from lnp_ml.modeling.models import LNPModel, LNPModelWithoutMPNN
 
 
 app = typer.Typer()
 
+# MPNN ensemble 默认路径
+DEFAULT_MPNN_ENSEMBLE_DIR = MODELS_DIR / "mpnn" / "all_amine_split_for_LiON"
 
-def load_model(model_path: Path, device: torch.device) -> LNPModelWithoutMPNN:
-    """加载训练好的模型"""
+
+def find_mpnn_ensemble_paths(base_dir: Path = DEFAULT_MPNN_ENSEMBLE_DIR) -> List[str]:
+    """自动查找 MPNN ensemble 的 model.pt 文件。"""
+    model_paths = sorted(base_dir.glob("cv_*/fold_*/model_*/model.pt"))
+    if not model_paths:
+        raise FileNotFoundError(f"No model.pt files found in {base_dir}")
+    return [str(p) for p in model_paths]
+
+
+def load_model(
+    model_path: Path,
+    device: torch.device,
+    mpnn_device: str = "cpu",
+) -> Union[LNPModel, LNPModelWithoutMPNN]:
+    """
+    加载训练好的模型。
+    
+    自动根据 checkpoint 的 config.use_mpnn 选择模型类型。
+    """
     checkpoint = torch.load(model_path, map_location=device)
     config = checkpoint["config"]
+    use_mpnn = config.get("use_mpnn", False)
     
+    if use_mpnn:
+        # 自动查找 MPNN ensemble
+        logger.info("Model was trained with MPNN, auto-detecting ensemble...")
+        ensemble_paths = find_mpnn_ensemble_paths()
+        logger.info(f"Found {len(ensemble_paths)} MPNN models")
+        
+        model = LNPModel(
+            d_model=config["d_model"],
+            num_heads=config["num_heads"],
+            n_attn_layers=config["n_attn_layers"],
+            fusion_strategy=config["fusion_strategy"],
+            head_hidden_dim=config["head_hidden_dim"],
+            dropout=config["dropout"],
+            mpnn_ensemble_paths=ensemble_paths,
+            mpnn_device=mpnn_device,
+        )
+    else:
         model = LNPModelWithoutMPNN(
             d_model=config["d_model"],
             num_heads=config["num_heads"],
@@ -30,6 +67,7 @@ def load_model(model_path: Path, device: torch.device) -> LNPModelWithoutMPNN:
             head_hidden_dim=config["head_hidden_dim"],
             dropout=config["dropout"],
         )
+    
     model.load_state_dict(checkpoint["model_state_dict"])
     model.to(device)
     model.eval()
@@ -43,7 +81,7 @@ def load_model(model_path: Path, device: torch.device) -> LNPModelWithoutMPNN:
 
 @torch.no_grad()
 def predict_batch(
-    model: LNPModelWithoutMPNN,
+    model: Union[LNPModel, LNPModelWithoutMPNN],
     loader: DataLoader,
     device: torch.device,
 ) -> Dict[str, List]:

lnp_ml/modeling/pretrain.py

@@ -0,0 +1,361 @@
+"""预训练脚本：在外部 LiON 数据上预训练 backbone + delivery head"""
+
+import json
+from pathlib import Path
+from typing import Dict, List, Optional
+
+import pandas as pd
+import torch
+import torch.nn as nn
+from torch.utils.data import DataLoader
+from loguru import logger
+from tqdm import tqdm
+import typer
+
+from lnp_ml.config import MODELS_DIR, PROCESSED_DATA_DIR
+from lnp_ml.dataset import ExternalDeliveryDataset, collate_fn
+
+# MPNN ensemble 默认路径
+DEFAULT_MPNN_ENSEMBLE_DIR = MODELS_DIR / "mpnn" / "all_amine_split_for_LiON"
+
+
+def find_mpnn_ensemble_paths(base_dir: Path = DEFAULT_MPNN_ENSEMBLE_DIR) -> List[str]:
+    """
+    自动查找 MPNN ensemble 的 model.pt 文件。
+    
+    在 base_dir 下查找所有 cv_*/fold_*/model_*/model.pt 文件。
+    """
+    model_paths = sorted(base_dir.glob("cv_*/fold_*/model_*/model.pt"))
+    if not model_paths:
+        raise FileNotFoundError(f"No model.pt files found in {base_dir}")
+    return [str(p) for p in model_paths]
+from lnp_ml.modeling.models import LNPModel, LNPModelWithoutMPNN
+
+
+app = typer.Typer()
+
+
+class EarlyStopping:
+    """早停机制"""
+
+    def __init__(self, patience: int = 10, min_delta: float = 0.0):
+        self.patience = patience
+        self.min_delta = min_delta
+        self.counter = 0
+        self.best_loss = float("inf")
+
+    def __call__(self, val_loss: float) -> bool:
+        if val_loss < self.best_loss - self.min_delta:
+            self.best_loss = val_loss
+            self.counter = 0
+            return False
+        self.counter += 1
+        return self.counter >= self.patience
+
+
+def warmup_cache(model: nn.Module, smiles_list: List[str], batch_size: int = 256) -> None:
+    """
+    预热 RDKit 特征缓存，避免训练时计算阻塞。
+    """
+    unique_smiles = list(set(smiles_list))
+    logger.info(f"Warming up RDKit cache for {len(unique_smiles)} unique SMILES...")
+    
+    for i in tqdm(range(0, len(unique_smiles), batch_size), desc="Cache warmup"):
+        batch = unique_smiles[i:i + batch_size]
+        model.rdkit_encoder(batch)
+    
+    logger.success(f"Cache warmup complete. Cached {len(model.rdkit_encoder._cache)} SMILES.")
+
+
+def train_epoch_delivery(
+    model: nn.Module,
+    loader: DataLoader,
+    optimizer: torch.optim.Optimizer,
+    device: torch.device,
+    epoch: int = 0,
+) -> Dict[str, float]:
+    """
+    单个 epoch 的预训练（仅 delivery 任务）。
+    """
+    model.train()
+    total_loss = 0.0
+    n_samples = 0
+
+    pbar = tqdm(loader, desc=f"Epoch {epoch+1} [Train]", leave=False)
+    for batch in pbar:
+        smiles = batch["smiles"]
+        tabular = {k: v.to(device) for k, v in batch["tabular"].items()}
+        targets = batch["targets"]["delivery"].to(device)
+        mask = batch["mask"]["delivery"].to(device)
+
+        optimizer.zero_grad()
+
+        # Forward: 只预测 delivery
+        pred = model.forward_delivery(smiles, tabular)  # [B, 1]
+        pred = pred.squeeze(-1)  # [B]
+
+        # 计算损失（仅对有效样本）
+        if mask.any():
+            loss = nn.functional.mse_loss(pred[mask], targets[mask])
+            loss.backward()
+            optimizer.step()
+            total_loss += loss.item() * mask.sum().item()
+            n_samples += mask.sum().item()
+            
+            pbar.set_postfix({"loss": total_loss / max(n_samples, 1)})
+
+    avg_loss = total_loss / max(n_samples, 1)
+    return {"loss": avg_loss, "n_samples": n_samples}
+
+
+@torch.no_grad()
+def validate_delivery(
+    model: nn.Module,
+    loader: DataLoader,
+    device: torch.device,
+) -> Dict[str, float]:
+    """
+    验证（仅 delivery 任务）。
+    """
+    model.eval()
+    total_loss = 0.0
+    n_samples = 0
+
+    for batch in loader:
+        smiles = batch["smiles"]
+        tabular = {k: v.to(device) for k, v in batch["tabular"].items()}
+        targets = batch["targets"]["delivery"].to(device)
+        mask = batch["mask"]["delivery"].to(device)
+
+        pred = model.forward_delivery(smiles, tabular).squeeze(-1)
+
+        if mask.any():
+            loss = nn.functional.mse_loss(pred[mask], targets[mask])
+            total_loss += loss.item() * mask.sum().item()
+            n_samples += mask.sum().item()
+
+    avg_loss = total_loss / max(n_samples, 1)
+    return {"loss": avg_loss, "n_samples": n_samples}
+
+
+def pretrain(
+    train_loader: DataLoader,
+    val_loader: DataLoader,
+    model: nn.Module,
+    device: torch.device,
+    lr: float = 1e-4,
+    weight_decay: float = 1e-5,
+    epochs: int = 50,
+    patience: int = 10,
+) -> dict:
+    """
+    预训练循环。
+
+    Returns:
+        训练历史和最佳验证损失
+    """
+    model = model.to(device)
+    optimizer = torch.optim.AdamW(model.parameters(), lr=lr, weight_decay=weight_decay)
+    scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
+        optimizer, mode="min", factor=0.5, patience=5, verbose=True
+    )
+    early_stopping = EarlyStopping(patience=patience)
+
+    history = {"train": [], "val": []}
+    best_val_loss = float("inf")
+    best_state = None
+
+    for epoch in range(epochs):
+        # Train
+        train_metrics = train_epoch_delivery(model, train_loader, optimizer, device, epoch)
+
+        # Validate
+        val_metrics = validate_delivery(model, val_loader, device)
+
+        # Log
+        logger.info(
+            f"Epoch {epoch + 1}/{epochs} | "
+            f"Train Loss: {train_metrics['loss']:.4f} | "
+            f"Val Loss: {val_metrics['loss']:.4f}"
+        )
+
+        history["train"].append(train_metrics)
+        history["val"].append(val_metrics)
+
+        # Learning rate scheduling
+        scheduler.step(val_metrics["loss"])
+
+        # Save best model
+        if val_metrics["loss"] < best_val_loss:
+            best_val_loss = val_metrics["loss"]
+            best_state = {k: v.cpu().clone() for k, v in model.state_dict().items()}
+            logger.info(f"  -> New best model (val_loss={best_val_loss:.4f})")
+
+        # Early stopping
+        if early_stopping(val_metrics["loss"]):
+            logger.info(f"Early stopping at epoch {epoch + 1}")
+            break
+
+    # Restore best model
+    if best_state is not None:
+        model.load_state_dict(best_state)
+
+    return {
+        "history": history,
+        "best_val_loss": best_val_loss,
+    }
+
+
+@app.command()
+def main(
+    # 数据路径（已处理的 parquet 文件）
+    train_path: Path = PROCESSED_DATA_DIR / "train_pretrain.parquet",
+    val_path: Path = PROCESSED_DATA_DIR / "val_pretrain.parquet",
+    output_dir: Path = MODELS_DIR,
+    # 模型参数
+    d_model: int = 256,
+    num_heads: int = 8,
+    n_attn_layers: int = 4,
+    fusion_strategy: str = "attention",
+    head_hidden_dim: int = 128,
+    dropout: float = 0.1,
+    # MPNN 参数（可选）
+    use_mpnn: bool = False,  # 启用 MPNN，自动从默认路径加载 ensemble
+    mpnn_checkpoint: Optional[str] = None,
+    mpnn_ensemble_paths: Optional[str] = None,  # 逗号分隔的路径列表
+    mpnn_device: str = "cpu",
+    # 训练参数
+    batch_size: int = 64,
+    lr: float = 1e-4,
+    weight_decay: float = 1e-5,
+    epochs: int = 50,
+    patience: int = 10,
+    # 设备
+    device: str = "cuda" if torch.cuda.is_available() else "cpu",
+):
+    """
+    在外部 LiON 数据上预训练 backbone + delivery head。
+
+    需要先运行 `make data_pretrain` 生成 parquet 文件。
+    
+    使用 --use-mpnn 启用 MPNN encoder（自动从 models/mpnn/all_amine_split_for_LiON 加载）。
+
+    产出:
+        - models/pretrain_delivery.pt: 包含 backbone + delivery head 权重
+        - models/pretrain_history.json: 训练历史
+    """
+    logger.info(f"Using device: {device}")
+    device_obj = torch.device(device)
+
+    # 加载已处理的 parquet 文件
+    logger.info(f"Loading train data from {train_path}")
+    train_df = pd.read_parquet(train_path)
+    train_dataset = ExternalDeliveryDataset(train_df)
+    
+    logger.info(f"Loading val data from {val_path}")
+    val_df = pd.read_parquet(val_path)
+    val_dataset = ExternalDeliveryDataset(val_df)
+    
+    logger.info(f"Train samples: {len(train_dataset)}, Val samples: {len(val_dataset)}")
+
+    train_loader = DataLoader(
+        train_dataset, batch_size=batch_size, shuffle=True, collate_fn=collate_fn
+    )
+    val_loader = DataLoader(
+        val_dataset, batch_size=batch_size, shuffle=False, collate_fn=collate_fn
+    )
+
+    # 解析 MPNN 配置
+    # 优先级：mpnn_checkpoint > mpnn_ensemble_paths > use_mpnn（自动查找）
+    ensemble_paths_list = None
+    if mpnn_ensemble_paths:
+        ensemble_paths_list = mpnn_ensemble_paths.split(",")
+    elif use_mpnn and mpnn_checkpoint is None:
+        # --use-mpnn 但没有指定具体路径，自动查找
+        logger.info(f"Auto-detecting MPNN ensemble from {DEFAULT_MPNN_ENSEMBLE_DIR}")
+        ensemble_paths_list = find_mpnn_ensemble_paths()
+        logger.info(f"Found {len(ensemble_paths_list)} MPNN models")
+    
+    enable_mpnn = mpnn_checkpoint is not None or ensemble_paths_list is not None
+
+    # 创建模型
+    logger.info(f"Creating model (use_mpnn={enable_mpnn})...")
+    if enable_mpnn:
+        model = LNPModel(
+            d_model=d_model,
+            num_heads=num_heads,
+            n_attn_layers=n_attn_layers,
+            fusion_strategy=fusion_strategy,
+            head_hidden_dim=head_hidden_dim,
+            dropout=dropout,
+            mpnn_checkpoint=mpnn_checkpoint,
+            mpnn_ensemble_paths=ensemble_paths_list,
+            mpnn_device=mpnn_device,
+        )
+    else:
+        model = LNPModelWithoutMPNN(
+            d_model=d_model,
+            num_heads=num_heads,
+            n_attn_layers=n_attn_layers,
+            fusion_strategy=fusion_strategy,
+            head_hidden_dim=head_hidden_dim,
+            dropout=dropout,
+        )
+
+    n_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
+    logger.info(f"Model parameters: {n_params:,}")
+
+    # 预热 RDKit 缓存（避免训练时阻塞）
+    all_smiles = train_df["smiles"].tolist() + val_df["smiles"].tolist()
+    warmup_cache(model, all_smiles, batch_size=256)
+
+    # 预训练
+    logger.info("Starting pretraining on external data (delivery only)...")
+    result = pretrain(
+        train_loader=train_loader,
+        val_loader=val_loader,
+        model=model,
+        device=device_obj,
+        lr=lr,
+        weight_decay=weight_decay,
+        epochs=epochs,
+        patience=patience,
+    )
+
+    # 保存预训练 checkpoint
+    output_dir.mkdir(parents=True, exist_ok=True)
+    checkpoint_path = output_dir / "pretrain_delivery.pt"
+    torch.save(
+        {
+            "model_state_dict": model.state_dict(),
+            "backbone_state_dict": model.get_backbone_state_dict(),
+            "delivery_head_state_dict": model.get_delivery_head_state_dict(),
+            "config": {
+                "d_model": d_model,
+                "num_heads": num_heads,
+                "n_attn_layers": n_attn_layers,
+                "fusion_strategy": fusion_strategy,
+                "head_hidden_dim": head_hidden_dim,
+                "dropout": dropout,
+                "use_mpnn": enable_mpnn,
+            },
+            "best_val_loss": result["best_val_loss"],
+        },
+        checkpoint_path,
+    )
+    logger.success(f"Saved pretrain checkpoint to {checkpoint_path}")
+
+    # 保存训练历史
+    history_path = output_dir / "pretrain_history.json"
+    with open(history_path, "w") as f:
+        json.dump(result["history"], f, indent=2)
+    logger.success(f"Saved pretrain history to {history_path}")
+
+    logger.success(
+        f"Pretraining complete! Best val_loss: {result['best_val_loss']:.4f}"
+    )
+
+
+if __name__ == "__main__":
+    app()
+

lnp_ml/modeling/train.py

@@ -2,7 +2,7 @@
 
 import json
 from pathlib import Path
-from typing import Optional
+from typing import List, Optional, Union
 
 import pandas as pd
 import torch
@@ -12,7 +12,7 @@ import typer
 
 from lnp_ml.config import MODELS_DIR, PROCESSED_DATA_DIR
 from lnp_ml.dataset import LNPDataset, collate_fn
-from lnp_ml.modeling.models import LNPModelWithoutMPNN
+from lnp_ml.modeling.models import LNPModel, LNPModelWithoutMPNN
 from lnp_ml.modeling.trainer import (
     train_epoch,
     validate,
@@ -20,6 +20,21 @@ from lnp_ml.modeling.trainer import (
     LossWeights,
 )
 
+# MPNN ensemble 默认路径
+DEFAULT_MPNN_ENSEMBLE_DIR = MODELS_DIR / "mpnn" / "all_amine_split_for_LiON"
+
+
+def find_mpnn_ensemble_paths(base_dir: Path = DEFAULT_MPNN_ENSEMBLE_DIR) -> List[str]:
+    """
+    自动查找 MPNN ensemble 的 model.pt 文件。
+    
+    在 base_dir 下查找所有 cv_*/fold_*/model_*/model.pt 文件。
+    """
+    model_paths = sorted(base_dir.glob("cv_*/fold_*/model_*/model.pt"))
+    if not model_paths:
+        raise FileNotFoundError(f"No model.pt files found in {base_dir}")
+    return [str(p) for p in model_paths]
+
 
 app = typer.Typer()
 
@@ -31,8 +46,27 @@ def create_model(
     fusion_strategy: str = "attention",
     head_hidden_dim: int = 128,
     dropout: float = 0.1,
-) -> LNPModelWithoutMPNN:
-    """创建模型"""
+    # MPNN 参数（可选）
+    mpnn_checkpoint: Optional[str] = None,
+    mpnn_ensemble_paths: Optional[List[str]] = None,
+    mpnn_device: str = "cpu",
+) -> Union[LNPModel, LNPModelWithoutMPNN]:
+    """创建模型（支持可选的 MPNN encoder）"""
+    use_mpnn = mpnn_checkpoint is not None or mpnn_ensemble_paths is not None
+    
+    if use_mpnn:
+        return LNPModel(
+            d_model=d_model,
+            num_heads=num_heads,
+            n_attn_layers=n_attn_layers,
+            fusion_strategy=fusion_strategy,
+            head_hidden_dim=head_hidden_dim,
+            dropout=dropout,
+            mpnn_checkpoint=mpnn_checkpoint,
+            mpnn_ensemble_paths=mpnn_ensemble_paths,
+            mpnn_device=mpnn_device,
+        )
+    else:
         return LNPModelWithoutMPNN(
             d_model=d_model,
             num_heads=num_heads,
@@ -191,6 +225,11 @@ def main(
     fusion_strategy: str = "attention",
     head_hidden_dim: int = 128,
     dropout: float = 0.1,
+    # MPNN 参数（可选）
+    use_mpnn: bool = False,  # 启用 MPNN，自动从默认路径加载 ensemble
+    mpnn_checkpoint: Optional[str] = None,
+    mpnn_ensemble_paths: Optional[str] = None,  # 逗号分隔的路径列表
+    mpnn_device: str = "cpu",
     # 训练参数
     batch_size: int = 32,
     lr: float = 1e-4,
@@ -201,13 +240,20 @@ def main(
     tune: bool = False,
     n_trials: int = 20,
     epochs_per_trial: int = 30,
+    # 预训练权重加载
+    init_from_pretrain: Optional[Path] = None,
+    load_delivery_head: bool = True,
+    freeze_backbone: bool = False,  # 冻结 backbone，只训练 heads
     # 设备
     device: str = "cuda" if torch.cuda.is_available() else "cpu",
 ):
     """
-    训练 LNP 预测模型。
+    训练 LNP 预测模型（多任务 finetune）。
     
     使用 --tune 启用超参数调优。
+    使用 --init-from-pretrain 从预训练 checkpoint 初始化 backbone。
+    使用 --use-mpnn 启用 MPNN encoder（自动从 models/mpnn/all_amine_split_for_LiON 加载）。
+    使用 --freeze-backbone 冻结 backbone，只训练多任务 heads。
     """
     logger.info(f"Using device: {device}")
     device = torch.device(device)
@@ -258,8 +304,21 @@ def main(
         lr = best_params["lr"]
         weight_decay = best_params["weight_decay"]
     
+    # 解析 MPNN 配置
+    # 优先级：mpnn_checkpoint > mpnn_ensemble_paths > use_mpnn（自动查找）
+    ensemble_paths_list = None
+    if mpnn_ensemble_paths:
+        ensemble_paths_list = mpnn_ensemble_paths.split(",")
+    elif use_mpnn and mpnn_checkpoint is None:
+        # --use-mpnn 但没有指定具体路径，自动查找
+        logger.info(f"Auto-detecting MPNN ensemble from {DEFAULT_MPNN_ENSEMBLE_DIR}")
+        ensemble_paths_list = find_mpnn_ensemble_paths()
+        logger.info(f"Found {len(ensemble_paths_list)} MPNN models")
+    
+    enable_mpnn = mpnn_checkpoint is not None or ensemble_paths_list is not None
+    
     # 创建模型
-    logger.info("Creating model...")
+    logger.info(f"Creating model (use_mpnn={enable_mpnn})...")
     model = create_model(
         d_model=d_model,
         num_heads=num_heads,
@@ -267,11 +326,48 @@ def main(
         fusion_strategy=fusion_strategy,
         head_hidden_dim=head_hidden_dim,
         dropout=dropout,
+        mpnn_checkpoint=mpnn_checkpoint,
+        mpnn_ensemble_paths=ensemble_paths_list,
+        mpnn_device=mpnn_device,
     )
     
+    # 加载预训练权重（如果指定）
+    if init_from_pretrain is not None:
+        logger.info(f"Loading pretrain weights from {init_from_pretrain}")
+        checkpoint = torch.load(init_from_pretrain, map_location="cpu")
+        
+        # 检查配置是否兼容
+        pretrain_config = checkpoint.get("config", {})
+        if pretrain_config.get("d_model") != d_model:
+            logger.warning(
+                f"d_model mismatch: pretrain={pretrain_config.get('d_model')}, "
+                f"current={d_model}. Skipping pretrain loading."
+            )
+        else:
+            # 加载 backbone + (可选) delivery head
+            model.load_pretrain_weights(
+                pretrain_state_dict=checkpoint["model_state_dict"],
+                load_delivery_head=load_delivery_head,
+                strict=False,
+            )
+            logger.success(
+                f"Loaded pretrain weights (backbone + delivery_head={load_delivery_head})"
+            )
+    
+    # 冻结 backbone（如果指定）
+    if freeze_backbone:
+        logger.info("Freezing backbone (token_projector, cross_attention, fusion)...")
+        frozen_count = 0
+        for name, param in model.named_parameters():
+            if name.startswith(("token_projector.", "cross_attention.", "fusion.")):
+                param.requires_grad = False
+                frozen_count += 1
+        logger.info(f"Frozen {frozen_count} parameter tensors")
+    
     # 打印模型信息
-    n_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
-    logger.info(f"Model parameters: {n_params:,}")
+    n_params_total = sum(p.numel() for p in model.parameters())
+    n_params_trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)
+    logger.info(f"Model parameters: {n_params_total:,} total, {n_params_trainable:,} trainable")
     
     # 训练
     logger.info("Starting training...")
@@ -297,8 +393,10 @@ def main(
             "fusion_strategy": fusion_strategy,
             "head_hidden_dim": head_hidden_dim,
             "dropout": dropout,
+            "use_mpnn": enable_mpnn,
         },
         "best_val_loss": result["best_val_loss"],
+        "init_from_pretrain": str(init_from_pretrain) if init_from_pretrain else None,
     }, model_path)
     logger.success(f"Saved model to {model_path}")
     

models/pretrain_history.json

@@ -0,0 +1,286 @@
+{
+  "train": [
+    {
+      "loss": 0.8151603855680482,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.6867990841792819,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.645540308540888,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.5923176541020599,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.5720762926262872,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.5477570670417328,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.5280393017717573,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.5122504676513313,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.49667307051028314,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.486139440352648,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.4749755339466122,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.4636757543530298,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.4543497681877452,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.4408158337956461,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.4419790126221837,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.42850686623585116,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.41607048387867007,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.427172136486513,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.4125568530569382,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.39480836287767923,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.3885056775666858,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.3894976457588827,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.3890058272899995,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.3741690826284791,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.3534914434345719,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.3349389765134386,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.32965143874976194,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.32094062546116675,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.32526135008251184,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.31289531808423826,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.3088379208288558,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.2994744991261045,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.2981521815160671,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.29143649446979303,
+      "n_samples": 8721
+    },
+    {
+      "loss": 0.29075756723379653,
+      "n_samples": 8721
+    }
+  ],
+  "val": [
+    {
+      "loss": 0.7625711447683281,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.7092331695236781,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.7014068689723995,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.6595172673863646,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.6312279044905191,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.6349272860831151,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.6587623598744133,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.6093261837651732,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.6125607111474924,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.6005943137518024,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.6876292386783289,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.5940848466228036,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.5820883587079644,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.6302792748938035,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.5849901610914275,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.5830434826428553,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.5643168952858116,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.5592790719340829,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.600335100686833,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.5646457721097674,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.6288956836004376,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.5771863222183704,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.5738056593250687,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.5636531712085593,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.5465074849879163,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.5701294839843508,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.5570075802420438,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.5711473401701241,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.5576858864741552,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.5624132422716871,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.5655298555506272,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.5568078993151677,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.567752199383958,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.5683093779442603,
+      "n_samples": 969
+    },
+    {
+      "loss": 0.5741443767974497,
+      "n_samples": 969
+    }
+  ]
+}

scripts/process_external.py

@@ -0,0 +1,98 @@
+"""外部数据预处理脚本：external -> processed"""
+
+from pathlib import Path
+
+import pandas as pd
+import typer
+from loguru import logger
+from sklearn.model_selection import train_test_split
+
+from lnp_ml.config import EXTERNAL_DATA_DIR, PROCESSED_DATA_DIR
+from lnp_ml.dataset import process_external_dataframe, LNPDatasetConfig, get_phys_cols, get_exp_cols, COMP_COLS, HELP_COLS
+
+
+app = typer.Typer()
+
+
+@app.command()
+def main(
+    input_path: Path = EXTERNAL_DATA_DIR / "all_data_LiON.csv",
+    output_dir: Path = PROCESSED_DATA_DIR,
+    train_ratio: float = 0.9,
+    seed: int = 42,
+):
+    """
+    处理外部 LiON 数据，生成预训练用的 parquet 文件。
+    
+    输出:
+        - processed/train_pretrain.parquet
+        - processed/val_pretrain.parquet
+        - processed/feature_columns_pretrain.txt
+    """
+    logger.info(f"Loading external data from {input_path}")
+    df = pd.read_csv(input_path)
+    logger.info(f"Loaded {len(df)} samples")
+    
+    # 过滤掉 quantified_delivery 为空的行
+    if "quantified_delivery" in df.columns:
+        before_len = len(df)
+        df = df[df["quantified_delivery"].notna()].reset_index(drop=True)
+        logger.info(f"Filtered NaN delivery: {before_len} -> {len(df)} samples")
+    
+    # 处理数据（列对齐、one-hot 生成）
+    logger.info("Processing dataframe (column alignment, one-hot encoding)...")
+    df = process_external_dataframe(df)
+    
+    # 获取所需列
+    config = LNPDatasetConfig()
+    feature_cols = (
+        ["smiles"]
+        + config.comp_cols
+        + config.phys_cols
+        + config.help_cols
+        + config.exp_cols
+        + ["quantified_delivery"]
+    )
+    
+    # 只保留需要的列
+    available_cols = [c for c in feature_cols if c in df.columns]
+    missing_cols = [c for c in feature_cols if c not in df.columns]
+    if missing_cols:
+        logger.warning(f"Missing columns (will be filled with 0): {missing_cols}")
+        for col in missing_cols:
+            df[col] = 0.0
+    
+    df = df[feature_cols]
+    
+    # 划分 train/val
+    logger.info(f"Splitting data: train_ratio={train_ratio}, seed={seed}")
+    train_df, val_df = train_test_split(
+        df, train_size=train_ratio, random_state=seed, shuffle=True
+    )
+    train_df = train_df.reset_index(drop=True)
+    val_df = val_df.reset_index(drop=True)
+    
+    logger.info(f"Train samples: {len(train_df)}, Val samples: {len(val_df)}")
+    
+    # 保存
+    output_dir.mkdir(parents=True, exist_ok=True)
+    
+    train_path = output_dir / "train_pretrain.parquet"
+    val_path = output_dir / "val_pretrain.parquet"
+    
+    train_df.to_parquet(train_path, index=False)
+    val_df.to_parquet(val_path, index=False)
+    
+    logger.success(f"Saved train data to {train_path}")
+    logger.success(f"Saved val data to {val_path}")
+    
+    # 保存特征列配置
+    cols_path = output_dir / "feature_columns_pretrain.txt"
+    with open(cols_path, "w") as f:
+        f.write("\n".join(feature_cols))
+    logger.success(f"Saved feature columns to {cols_path}")
+
+
+if __name__ == "__main__":
+    app()
+