workbench 0.8.219__py3-none-any.whl → 0.8.231__py3-none-any.whl

workbench/model_scripts/chemprop/generated_model_script.py

@@ -34,7 +34,7 @@ DEFAULT_HYPERPARAMETERS = {
     "max_epochs": 400,
     "patience": 50,
     "batch_size": 32,
-    # Message Passing
+    # Message Passing (ignored when using foundation model)
     "hidden_dim": 700,
     "depth": 6,
     "dropout": 0.1,  # Lower dropout - ensemble provides regularization
@@ -45,15 +45,23 @@ DEFAULT_HYPERPARAMETERS = {
     "criterion": "mae",
     # Random seed
     "seed": 42,
+    # Foundation model support
+    # - "CheMeleon": Load CheMeleon pretrained weights (auto-downloads on first use)
+    # - Path to .pt file: Load custom pretrained Chemprop model
+    # - None: Train from scratch (default)
+    "from_foundation": None,
+    # Freeze MPNN for N epochs, then unfreeze (0 = no freezing, train all params from start)
+    # Recommended: 5-20 epochs when using foundation models to stabilize FFN before fine-tuning MPNN
+    "freeze_mpnn_epochs": 0,
 }
 
 # Template parameters (filled in by Workbench)
 TEMPLATE_PARAMS = {
     "model_type": "uq_regressor",
-    "targets": ['logd'],
-    "feature_list": ['smiles', 'mollogp', 'fr_halogen', 'nbase', 'peoe_vsa6', 'bcut2d_mrlow', 'peoe_vsa7', 'peoe_vsa9', 'vsa_estate1', 'peoe_vsa1', 'numhdonors', 'vsa_estate5', 'smr_vsa3', 'slogp_vsa1', 'vsa_estate7', 'bcut2d_mwhi', 'axp_2dv', 'axp_3dv', 'mi', 'smr_vsa9', 'vsa_estate3', 'estate_vsa9', 'bcut2d_mwlow', 'tpsa', 'vsa_estate10', 'xch_5dv', 'slogp_vsa2', 'nhohcount', 'bcut2d_logplow', 'hallkieralpha', 'c2sp2', 'bcut2d_chglo', 'smr_vsa4', 'maxabspartialcharge', 'estate_vsa6', 'qed', 'slogp_vsa6', 'vsa_estate2', 'bcut2d_logphi', 'vsa_estate8', 'xch_7dv', 'fpdensitymorgan3', 'xpc_6d', 'smr_vsa10', 'axp_0d', 'fr_nh1', 'axp_4dv', 'peoe_vsa2', 'estate_vsa8', 'peoe_vsa5', 'vsa_estate6'],
-    "id_column": "molecule_name",
-    "model_metrics_s3_path": "s3://sandbox-sageworks-artifacts/models/logd-reg-chemprop-hybrid/training",
+    "targets": ['udm_asy_res_value'],
+    "feature_list": ['smiles'],
+    "id_column": "udm_mol_bat_id",
+    "model_metrics_s3_path": "s3://ideaya-sageworks-bucket/models/logd-value-reg-chemprop-1-dt/training",
     "hyperparameters": {},
 }
 
@@ -114,26 +122,27 @@ def _create_molecule_datapoints(
 # Model Loading (for SageMaker inference)
 # =============================================================================
 def model_fn(model_dir: str) -> dict:
-    """Load ChemProp MPNN ensemble from the specified directory."""
-    from lightning import pytorch as pl
+    """Load ChemProp MPNN ensemble from the specified directory.
 
+    Optimized for serverless cold starts - uses direct PyTorch inference
+    instead of Lightning Trainer to minimize startup time.
+    """
     metadata = joblib.load(os.path.join(model_dir, "ensemble_metadata.joblib"))
+
+    # Load all ensemble models (keep on CPU for serverless compatibility)
+    # ChemProp handles device placement internally
     ensemble_models = []
     for i in range(metadata["n_ensemble"]):
         model = models.MPNN.load_from_file(os.path.join(model_dir, f"chemprop_model_{i}.pt"))
         model.eval()
         ensemble_models.append(model)
 
-    # Pre-initialize trainer once during model loading (expensive operation)
-    trainer = pl.Trainer(accelerator="auto", logger=False, enable_progress_bar=False)
-
     print(f"Loaded {len(ensemble_models)} model(s), targets={metadata['target_columns']}")
     return {
         "ensemble_models": ensemble_models,
         "n_ensemble": metadata["n_ensemble"],
         "target_columns": metadata["target_columns"],
         "median_std": metadata["median_std"],
-        "trainer": trainer,
     }
 
 
@@ -141,13 +150,15 @@ def model_fn(model_dir: str) -> dict:
 # Inference (for SageMaker inference)
 # =============================================================================
 def predict_fn(df: pd.DataFrame, model_dict: dict) -> pd.DataFrame:
-    """Make predictions with ChemProp MPNN ensemble."""
+    """Make predictions with ChemProp MPNN ensemble.
+
+    Uses direct PyTorch inference (no Lightning Trainer) for fast serverless inference.
+    """
     model_type = TEMPLATE_PARAMS["model_type"]
     model_dir = os.environ.get("SM_MODEL_DIR", "/opt/ml/model")
 
     ensemble_models = model_dict["ensemble_models"]
     target_columns = model_dict["target_columns"]
-    trainer = model_dict["trainer"]  # Use pre-initialized trainer
 
     # Load artifacts
     label_encoder = None
@@ -202,18 +213,34 @@ def predict_fn(df: pd.DataFrame, model_dict: dict) -> pd.DataFrame:
         return df
 
     dataset = data.MoleculeDataset(datapoints)
-    dataloader = data.build_dataloader(dataset, shuffle=False)
+    dataloader = data.build_dataloader(dataset, shuffle=False, batch_size=64)
 
-    # Ensemble predictions
+    # Ensemble predictions using direct PyTorch inference (no Lightning Trainer)
     all_preds = []
     for model in ensemble_models:
+        model_preds = []
+        model.eval()
         with torch.inference_mode():
-            predictions = trainer.predict(model, dataloader)
-        preds = np.concatenate([p.numpy() for p in predictions], axis=0)
+            for batch in dataloader:
+                # TrainingBatch contains (bmg, V_d, X_d, targets, weights, lt_mask, gt_mask)
+                # For inference we only need bmg, V_d, X_d
+                bmg, V_d, X_d, *_ = batch
+                output = model(bmg, V_d, X_d)
+                model_preds.append(output.detach().cpu().numpy())
+
+        if len(model_preds) == 0:
+            print(f"Warning: No predictions generated. Dataset size: {len(datapoints)}")
+            continue
+
+        preds = np.concatenate(model_preds, axis=0)
         if preds.ndim == 3 and preds.shape[1] == 1:
             preds = preds.squeeze(axis=1)
         all_preds.append(preds)
 
+    if len(all_preds) == 0:
+        print("Error: No ensemble predictions generated")
+        return df
+
     preds = np.mean(np.stack(all_preds), axis=0)
     preds_std = np.std(np.stack(all_preds), axis=0)
     if preds.ndim == 1:
@@ -243,8 +270,11 @@ def predict_fn(df: pd.DataFrame, model_dict: dict) -> pd.DataFrame:
         df["prediction"] = df[f"{target_columns[0]}_pred"]
         df["prediction_std"] = df[f"{target_columns[0]}_pred_std"]
 
-        # Compute confidence from ensemble std
-        df = _compute_std_confidence(df, model_dict["median_std"])
+        # Compute confidence from ensemble std (or NaN if single model)
+        if model_dict["median_std"] is not None:
+            df = _compute_std_confidence(df, model_dict["median_std"])
+        else:
+            df["confidence"] = np.nan
 
     return df
 
@@ -279,54 +309,107 @@ if __name__ == "__main__":
     )
 
     # -------------------------------------------------------------------------
-    # Training-only helper function
+    # Training-only helper functions
     # -------------------------------------------------------------------------
-    def build_mpnn_model(
-        hyperparameters: dict,
-        task: str = "regression",
-        num_classes: int | None = None,
-        n_targets: int = 1,
-        n_extra_descriptors: int = 0,
-        x_d_transform: nn.ScaleTransform | None = None,
-        output_transform: nn.UnscaleTransform | None = None,
-        task_weights: np.ndarray | None = None,
-    ) -> models.MPNN:
-        """Build an MPNN model with specified hyperparameters."""
-        hidden_dim = hyperparameters["hidden_dim"]
-        depth = hyperparameters["depth"]
+    def _load_foundation_weights(from_foundation: str) -> tuple[nn.BondMessagePassing, nn.Aggregation]:
+        """Load pretrained MPNN weights from foundation model.
+
+        Args:
+            from_foundation: "CheMeleon" or path to .pt file
+
+        Returns:
+            Tuple of (message_passing, aggregation) modules
+        """
+        import urllib.request
+        from pathlib import Path
+
+        print(f"Loading foundation model: {from_foundation}")
+
+        if from_foundation.lower() == "chemeleon":
+            # Download from Zenodo if not cached
+            cache_dir = Path.home() / ".chemprop" / "foundation"
+            cache_dir.mkdir(parents=True, exist_ok=True)
+            chemeleon_path = cache_dir / "chemeleon_mp.pt"
+
+            if not chemeleon_path.exists():
+                print("  Downloading CheMeleon weights from Zenodo...")
+                urllib.request.urlretrieve(
+                    "https://zenodo.org/records/15460715/files/chemeleon_mp.pt", chemeleon_path
+                )
+                print(f"  Downloaded to {chemeleon_path}")
+
+            ckpt = torch.load(chemeleon_path, weights_only=True)
+            mp = nn.BondMessagePassing(**ckpt["hyper_parameters"])
+            mp.load_state_dict(ckpt["state_dict"])
+            print(f"  Loaded CheMeleon MPNN (hidden_dim={mp.output_dim})")
+            return mp, nn.MeanAggregation()
+
+        if not os.path.exists(from_foundation):
+            raise ValueError(f"Foundation model not found: {from_foundation}. Use 'CheMeleon' or a valid .pt path.")
+
+        ckpt = torch.load(from_foundation, weights_only=False)
+        if "hyper_parameters" in ckpt and "state_dict" in ckpt:
+            # CheMeleon-style checkpoint
+            mp = nn.BondMessagePassing(**ckpt["hyper_parameters"])
+            mp.load_state_dict(ckpt["state_dict"])
+            print(f"  Loaded custom foundation weights (hidden_dim={mp.output_dim})")
+            return mp, nn.MeanAggregation()
+
+        # Full MPNN model file
+        pretrained = models.MPNN.load_from_file(from_foundation)
+        print(f"  Loaded custom MPNN (hidden_dim={pretrained.message_passing.output_dim})")
+        return pretrained.message_passing, pretrained.agg
+
+    def _build_ffn(
+        task: str, input_dim: int, hyperparameters: dict,
+        num_classes: int | None, n_targets: int,
+        output_transform: nn.UnscaleTransform | None, task_weights: np.ndarray | None,
+    ) -> nn.Predictor:
+        """Build task-specific FFN head."""
         dropout = hyperparameters["dropout"]
         ffn_hidden_dim = hyperparameters["ffn_hidden_dim"]
         ffn_num_layers = hyperparameters["ffn_num_layers"]
 
-        mp = nn.BondMessagePassing(d_h=hidden_dim, depth=depth, dropout=dropout)
-        agg = nn.NormAggregation()
-        ffn_input_dim = hidden_dim + n_extra_descriptors
-
         if task == "classification" and num_classes is not None:
-            ffn = nn.MulticlassClassificationFFN(
-                n_classes=num_classes, input_dim=ffn_input_dim,
+            return nn.MulticlassClassificationFFN(
+                n_classes=num_classes, input_dim=input_dim,
                 hidden_dim=ffn_hidden_dim, n_layers=ffn_num_layers, dropout=dropout,
             )
+
+        from chemprop.nn.metrics import MAE, MSE
+        criterion_map = {"mae": MAE, "mse": MSE}
+        criterion_name = hyperparameters.get("criterion", "mae")
+        if criterion_name not in criterion_map:
+            raise ValueError(f"Unknown criterion '{criterion_name}'. Supported: {list(criterion_map.keys())}")
+
+        weights_tensor = torch.tensor(task_weights, dtype=torch.float32) if task_weights is not None else None
+        return nn.RegressionFFN(
+            input_dim=input_dim, hidden_dim=ffn_hidden_dim, n_layers=ffn_num_layers,
+            dropout=dropout, n_tasks=n_targets, output_transform=output_transform,
+            task_weights=weights_tensor, criterion=criterion_map[criterion_name](),
+        )
+
+    def build_mpnn_model(
+        hyperparameters: dict, task: str = "regression", num_classes: int | None = None,
+        n_targets: int = 1, n_extra_descriptors: int = 0,
+        x_d_transform: nn.ScaleTransform | None = None,
+        output_transform: nn.UnscaleTransform | None = None, task_weights: np.ndarray | None = None,
+    ) -> models.MPNN:
+        """Build MPNN model, optionally loading pretrained weights."""
+        from_foundation = hyperparameters.get("from_foundation")
+
+        if from_foundation:
+            mp, agg = _load_foundation_weights(from_foundation)
+            ffn_input_dim = mp.output_dim + n_extra_descriptors
         else:
-            # Map criterion name to ChemProp metric class (must have .clone() method)
-            from chemprop.nn.metrics import MAE, MSE
-
-            criterion_map = {
-                "mae": MAE,
-                "mse": MSE,
-            }
-            criterion_name = hyperparameters.get("criterion", "mae")
-            if criterion_name not in criterion_map:
-                raise ValueError(f"Unknown criterion '{criterion_name}'. Supported: {list(criterion_map.keys())}")
-            criterion = criterion_map[criterion_name]()
-
-            weights_tensor = torch.tensor(task_weights, dtype=torch.float32) if task_weights is not None else None
-            ffn = nn.RegressionFFN(
-                input_dim=ffn_input_dim, hidden_dim=ffn_hidden_dim, n_layers=ffn_num_layers,
-                dropout=dropout, n_tasks=n_targets, output_transform=output_transform, task_weights=weights_tensor,
-                criterion=criterion,
+            mp = nn.BondMessagePassing(
+                d_h=hyperparameters["hidden_dim"], depth=hyperparameters["depth"],
+                dropout=hyperparameters["dropout"],
             )
+            agg = nn.NormAggregation()
+            ffn_input_dim = hyperparameters["hidden_dim"] + n_extra_descriptors
 
+        ffn = _build_ffn(task, ffn_input_dim, hyperparameters, num_classes, n_targets, output_transform, task_weights)
         return models.MPNN(message_passing=mp, agg=agg, predictor=ffn, batch_norm=True, metrics=None, X_d_transform=x_d_transform)
 
     # -------------------------------------------------------------------------
@@ -359,6 +442,14 @@ if __name__ == "__main__":
     print(f"Extra features: {extra_feature_cols if use_extra_features else 'None (SMILES only)'}")
     print(f"Hyperparameters: {hyperparameters}")
 
+    # Log foundation model configuration
+    if hyperparameters.get("from_foundation"):
+        freeze_epochs = hyperparameters.get("freeze_mpnn_epochs", 0)
+        freeze_msg = f"MPNN frozen for {freeze_epochs} epochs" if freeze_epochs > 0 else "no freezing"
+        print(f"Foundation model: {hyperparameters['from_foundation']} ({freeze_msg})")
+    else:
+        print("Foundation model: None (training from scratch)")
+
     # Load training data
     training_files = [os.path.join(args.train, f) for f in os.listdir(args.train) if f.endswith(".csv")]
     print(f"Training Files: {training_files}")
@@ -456,7 +547,7 @@ if __name__ == "__main__":
         print(f"Fold {fold_idx + 1}/{len(folds)} - Train: {len(train_idx)}, Val: {len(val_idx)}")
         print(f"{'='*50}")
 
-        # Split data
+        # Split data (val_extra_raw preserves unscaled features for OOF predictions)
         df_train, df_val = all_df.iloc[train_idx].reset_index(drop=True), all_df.iloc[val_idx].reset_index(drop=True)
         train_targets, val_targets = all_targets[train_idx], all_targets[val_idx]
         train_extra = all_extra_features[train_idx] if all_extra_features is not None else None
@@ -484,7 +575,7 @@ if __name__ == "__main__":
         train_loader = data.build_dataloader(train_dataset, batch_size=batch_size, shuffle=True, num_workers=3)
         val_loader = data.build_dataloader(val_dataset, batch_size=batch_size, shuffle=False, num_workers=3)
 
-        # Build and train model
+        # Build model
         pl.seed_everything(hyperparameters["seed"] + fold_idx)
         mpnn = build_mpnn_model(
             hyperparameters, task=task, num_classes=num_classes, n_targets=n_targets,
@@ -492,14 +583,39 @@ if __name__ == "__main__":
             output_transform=output_transform, task_weights=task_weights,
         )
 
-        trainer = pl.Trainer(
-            accelerator="auto", max_epochs=hyperparameters["max_epochs"], logger=False, enable_progress_bar=True,
-            callbacks=[
-                pl.callbacks.EarlyStopping(monitor="val_loss", patience=hyperparameters["patience"], mode="min"),
-                pl.callbacks.ModelCheckpoint(dirpath=args.model_dir, filename=f"best_{fold_idx}", monitor="val_loss", mode="min", save_top_k=1),
-            ],
-        )
-        trainer.fit(mpnn, train_loader, val_loader)
+        # Train model (with optional two-phase foundation training)
+        freeze_mpnn_epochs = hyperparameters.get("freeze_mpnn_epochs", 0)
+        use_two_phase = hyperparameters.get("from_foundation") and freeze_mpnn_epochs > 0
+
+        def _set_mpnn_frozen(frozen: bool):
+            for param in mpnn.message_passing.parameters():
+                param.requires_grad = not frozen
+            for param in mpnn.agg.parameters():
+                param.requires_grad = not frozen
+
+        def _make_trainer(max_epochs: int, save_checkpoint: bool = False):
+            callbacks = [pl.callbacks.EarlyStopping(monitor="val_loss", patience=hyperparameters["patience"], mode="min")]
+            if save_checkpoint:
+                callbacks.append(pl.callbacks.ModelCheckpoint(
+                    dirpath=args.model_dir, filename=f"best_{fold_idx}", monitor="val_loss", mode="min", save_top_k=1
+                ))
+            return pl.Trainer(accelerator="auto", max_epochs=max_epochs, logger=False, enable_progress_bar=True, callbacks=callbacks)
+
+        if use_two_phase:
+            # Phase 1: Freeze MPNN, train FFN only
+            print(f"Phase 1: Training with frozen MPNN for {freeze_mpnn_epochs} epochs...")
+            _set_mpnn_frozen(True)
+            _make_trainer(freeze_mpnn_epochs).fit(mpnn, train_loader, val_loader)
+
+            # Phase 2: Unfreeze and fine-tune all
+            print("Phase 2: Unfreezing MPNN, continuing training...")
+            _set_mpnn_frozen(False)
+            remaining_epochs = max(1, hyperparameters["max_epochs"] - freeze_mpnn_epochs)
+            trainer = _make_trainer(remaining_epochs, save_checkpoint=True)
+            trainer.fit(mpnn, train_loader, val_loader)
+        else:
+            trainer = _make_trainer(hyperparameters["max_epochs"], save_checkpoint=True)
+            trainer.fit(mpnn, train_loader, val_loader)
 
         # Load best checkpoint
         if trainer.checkpoint_callback and trainer.checkpoint_callback.best_model_path:
@@ -509,7 +625,7 @@ if __name__ == "__main__":
         mpnn.eval()
         ensemble_models.append(mpnn)
 
-        # Out-of-fold predictions (using raw features)
+        # Out-of-fold predictions (using unscaled features - model's x_d_transform handles scaling)
         val_dps_raw, _ = _create_molecule_datapoints(df_val[smiles_column].tolist(), val_targets, val_extra_raw)
         val_loader_pred = data.build_dataloader(data.MoleculeDataset(val_dps_raw), batch_size=batch_size, shuffle=False)
 
@@ -599,11 +715,17 @@ if __name__ == "__main__":
         df_val["prediction"] = df_val[f"{target_columns[0]}_pred"]
         df_val["prediction_std"] = df_val[f"{target_columns[0]}_pred_std"]
 
-        # Compute confidence from ensemble std
-        median_std = float(np.median(preds_std[:, 0]))
-        print(f"\nComputing confidence scores (median_std={median_std:.6f})...")
-        df_val = _compute_std_confidence(df_val, median_std)
-        print(f"  Confidence: mean={df_val['confidence'].mean():.3f}, min={df_val['confidence'].min():.3f}, max={df_val['confidence'].max():.3f}")
+        # Compute confidence from ensemble std (or NaN for single model)
+        if preds_std is not None:
+            median_std = float(np.median(preds_std[:, 0]))
+            print(f"\nComputing confidence scores (median_std={median_std:.6f})...")
+            df_val = _compute_std_confidence(df_val, median_std)
+            print(f"  Confidence: mean={df_val['confidence'].mean():.3f}, min={df_val['confidence'].min():.3f}, max={df_val['confidence'].max():.3f}")
+        else:
+            # Single model - no ensemble std available, confidence is undefined
+            median_std = None
+            df_val["confidence"] = np.nan
+            print("\nSingle model (n_folds=1): No ensemble std, confidence set to NaN")
 
     # -------------------------------------------------------------------------
     # Save validation predictions to S3
@@ -633,6 +755,9 @@ if __name__ == "__main__":
         "n_folds": n_folds,
         "target_columns": target_columns,
         "median_std": median_std,  # For confidence calculation during inference
+        # Foundation model provenance (for tracking/reproducibility)
+        "from_foundation": hyperparameters.get("from_foundation", None),
+        "freeze_mpnn_epochs": hyperparameters.get("freeze_mpnn_epochs", 0),
     }
     joblib.dump(ensemble_metadata, os.path.join(args.model_dir, "ensemble_metadata.joblib"))
 

workbench/model_scripts/chemprop/model_script_utils.py

@@ -249,6 +249,36 @@ def output_fn(output_df: pd.DataFrame, accept_type: str) -> tuple[str, str]:
         raise RuntimeError(f"{accept_type} accept type is not supported by this script.")
 
 
+def cap_std_outliers(std_array: np.ndarray) -> np.ndarray:
+    """Cap extreme outliers in prediction_std using IQR method.
+
+    Uses the standard IQR fence (Q3 + 1.5*IQR) to cap extreme values.
+    This prevents unreasonably large std values while preserving the
+    relative ordering and keeping meaningful high-uncertainty signals.
+
+    Args:
+        std_array: Array of standard deviations (n_samples,) or (n_samples, n_targets)
+
+    Returns:
+        Array with outliers capped at the upper fence
+    """
+    if std_array.ndim == 1:
+        std_array = std_array.reshape(-1, 1)
+        squeeze = True
+    else:
+        squeeze = False
+
+    capped = std_array.copy()
+    for col in range(capped.shape[1]):
+        col_data = capped[:, col]
+        q1, q3 = np.percentile(col_data, [25, 75])
+        iqr = q3 - q1
+        upper_bound = q3 + 1.5 * iqr
+        capped[:, col] = np.minimum(col_data, upper_bound)
+
+    return capped.squeeze() if squeeze else capped
+
+
 def compute_regression_metrics(y_true: np.ndarray, y_pred: np.ndarray) -> dict[str, float]:
     """Compute standard regression metrics.
 

workbench/model_scripts/custom_models/chem_info/mol_descriptors.py

@@ -99,7 +99,6 @@ from rdkit.ML.Descriptors import MoleculeDescriptors
 from mordred import Calculator as MordredCalculator
 from mordred import AcidBase, Aromatic, Constitutional, Chi, CarbonTypes
 
-
 logger = logging.getLogger("workbench")
 logger.setLevel(logging.DEBUG)
 

workbench/model_scripts/custom_models/chem_info/molecular_descriptors.py

@@ -15,7 +15,6 @@ import json
 from mol_standardize import standardize
 from mol_descriptors import compute_descriptors
 
-
 # TRAINING SECTION
 #
 # This section (__main__) is where SageMaker will execute the training job

workbench/model_scripts/custom_models/chem_info/morgan_fingerprints.py

@@ -17,7 +17,6 @@ import json
 # Local imports
 from fingerprints import compute_morgan_fingerprints
 
-
 # TRAINING SECTION
 #
 # This section (__main__) is where SageMaker will execute the training job

workbench/model_scripts/pytorch_model/generated_model_script.py

@@ -59,12 +59,12 @@ DEFAULT_HYPERPARAMETERS = {
 
 # Template parameters (filled in by Workbench)
 TEMPLATE_PARAMS = {
-    "model_type": "uq_regressor",
-    "target": "udm_asy_res_efflux_ratio",
-    "features": ['fingerprint'],
+    "model_type": "classifier",
+    "target": "class",
+    "features": ['chi2v', 'fr_sulfone', 'chi1v', 'bcut2d_logplow', 'fr_piperzine', 'kappa3', 'smr_vsa1', 'slogp_vsa5', 'fr_ketone_topliss', 'fr_sulfonamd', 'fr_imine', 'fr_benzene', 'fr_ester', 'chi2n', 'labuteasa', 'peoe_vsa2', 'smr_vsa6', 'bcut2d_chglo', 'fr_sh', 'peoe_vsa1', 'fr_allylic_oxid', 'chi4n', 'fr_ar_oh', 'fr_nh0', 'fr_term_acetylene', 'slogp_vsa7', 'slogp_vsa4', 'estate_vsa1', 'vsa_estate4', 'numbridgeheadatoms', 'numheterocycles', 'fr_ketone', 'fr_morpholine', 'fr_guanido', 'estate_vsa2', 'numheteroatoms', 'fr_nitro_arom_nonortho', 'fr_piperdine', 'nocount', 'numspiroatoms', 'fr_aniline', 'fr_thiophene', 'slogp_vsa10', 'fr_amide', 'slogp_vsa2', 'fr_epoxide', 'vsa_estate7', 'fr_ar_coo', 'fr_imidazole', 'fr_nitrile', 'fr_oxazole', 'numsaturatedrings', 'fr_pyridine', 'fr_hoccn', 'fr_ndealkylation1', 'numaliphaticheterocycles', 'fr_phenol', 'maxpartialcharge', 'vsa_estate5', 'peoe_vsa13', 'minpartialcharge', 'qed', 'fr_al_oh', 'slogp_vsa11', 'chi0n', 'fr_bicyclic', 'peoe_vsa12', 'fpdensitymorgan1', 'fr_oxime', 'molwt', 'fr_dihydropyridine', 'smr_vsa5', 'peoe_vsa5', 'fr_nitro', 'hallkieralpha', 'heavyatommolwt', 'fr_alkyl_halide', 'peoe_vsa8', 'fr_nhpyrrole', 'fr_isocyan', 'bcut2d_chghi', 'fr_lactam', 'peoe_vsa11', 'smr_vsa9', 'tpsa', 'chi4v', 'slogp_vsa1', 'phi', 'bcut2d_logphi', 'avgipc', 'estate_vsa11', 'fr_coo', 'bcut2d_mwhi', 'numunspecifiedatomstereocenters', 'vsa_estate10', 'estate_vsa8', 'numvalenceelectrons', 'fr_nh2', 'fr_lactone', 'vsa_estate1', 'estate_vsa4', 'numatomstereocenters', 'vsa_estate8', 'fr_para_hydroxylation', 'peoe_vsa3', 'fr_thiazole', 'peoe_vsa10', 'fr_ndealkylation2', 'slogp_vsa12', 'peoe_vsa9', 'maxestateindex', 'fr_quatn', 'smr_vsa7', 'minestateindex', 'numaromaticheterocycles', 'numrotatablebonds', 'fr_ar_nh', 'fr_ether', 'exactmolwt', 'fr_phenol_noorthohbond', 'slogp_vsa3', 'fr_ar_n', 'sps', 'fr_c_o_nocoo', 'bertzct', 'peoe_vsa7', 'slogp_vsa8', 'numradicalelectrons', 'molmr', 'fr_tetrazole', 'numsaturatedcarbocycles', 'bcut2d_mrhi', 'kappa1', 'numamidebonds', 'fpdensitymorgan2', 'smr_vsa8', 'chi1n', 'estate_vsa6', 'fr_barbitur', 'fr_diazo', 'kappa2', 'chi0', 'bcut2d_mrlow', 'balabanj', 'peoe_vsa4', 'numhacceptors', 'fr_sulfide', 'chi3n', 'smr_vsa2', 'fr_al_oh_notert', 'fr_benzodiazepine', 'fr_phos_ester', 'fr_aldehyde', 'fr_coo2', 'estate_vsa5', 'fr_prisulfonamd', 'numaromaticcarbocycles', 'fr_unbrch_alkane', 'fr_urea', 'fr_nitroso', 'smr_vsa10', 'fr_c_s', 'smr_vsa3', 'fr_methoxy', 'maxabspartialcharge', 'slogp_vsa9', 'heavyatomcount', 'fr_azide', 'chi3v', 'smr_vsa4', 'mollogp', 'chi0v', 'fr_aryl_methyl', 'fr_nh1', 'fpdensitymorgan3', 'fr_furan', 'fr_hdrzine', 'fr_arn', 'numaromaticrings', 'vsa_estate3', 'fr_azo', 'fr_halogen', 'estate_vsa9', 'fr_hdrzone', 'numhdonors', 'fr_alkyl_carbamate', 'fr_isothiocyan', 'minabspartialcharge', 'fr_al_coo', 'ringcount', 'chi1', 'estate_vsa7', 'fr_nitro_arom', 'vsa_estate9', 'minabsestateindex', 'maxabsestateindex', 'vsa_estate6', 'estate_vsa10', 'estate_vsa3', 'fr_n_o', 'fr_amidine', 'fr_thiocyan', 'fr_phos_acid', 'fr_c_o', 'fr_imide', 'numaliphaticrings', 'peoe_vsa6', 'vsa_estate2', 'nhohcount', 'numsaturatedheterocycles', 'slogp_vsa6', 'peoe_vsa14', 'fractioncsp3', 'bcut2d_mwlow', 'numaliphaticcarbocycles', 'fr_priamide', 'nacid', 'nbase', 'naromatom', 'narombond', 'sz', 'sm', 'sv', 'sse', 'spe', 'sare', 'sp', 'si', 'mz', 'mm', 'mv', 'mse', 'mpe', 'mare', 'mp', 'mi', 'xch_3d', 'xch_4d', 'xch_5d', 'xch_6d', 'xch_7d', 'xch_3dv', 'xch_4dv', 'xch_5dv', 'xch_6dv', 'xch_7dv', 'xc_3d', 'xc_4d', 'xc_5d', 'xc_6d', 'xc_3dv', 'xc_4dv', 'xc_5dv', 'xc_6dv', 'xpc_4d', 'xpc_5d', 'xpc_6d', 'xpc_4dv', 'xpc_5dv', 'xpc_6dv', 'xp_0d', 'xp_1d', 'xp_2d', 'xp_3d', 'xp_4d', 'xp_5d', 'xp_6d', 'xp_7d', 'axp_0d', 'axp_1d', 'axp_2d', 'axp_3d', 'axp_4d', 'axp_5d', 'axp_6d', 'axp_7d', 'xp_0dv', 'xp_1dv', 'xp_2dv', 'xp_3dv', 'xp_4dv', 'xp_5dv', 'xp_6dv', 'xp_7dv', 'axp_0dv', 'axp_1dv', 'axp_2dv', 'axp_3dv', 'axp_4dv', 'axp_5dv', 'axp_6dv', 'axp_7dv', 'c1sp1', 'c2sp1', 'c1sp2', 'c2sp2', 'c3sp2', 'c1sp3', 'c2sp3', 'c3sp3', 'c4sp3', 'hybratio', 'fcsp3', 'num_stereocenters', 'num_unspecified_stereocenters', 'num_defined_stereocenters', 'num_r_centers', 'num_s_centers', 'num_stereobonds', 'num_e_bonds', 'num_z_bonds', 'stereo_complexity', 'frac_defined_stereo'],
     "id_column": "udm_mol_bat_id",
-    "compressed_features": ['fingerprint'],
-    "model_metrics_s3_path": "s3://ideaya-sageworks-bucket/models/caco2-er-reg-fp-pytorch/training",
+    "compressed_features": [],
+    "model_metrics_s3_path": "s3://ideaya-sageworks-bucket/models/caco2-er-class-pytorch-1-fr/training",
     "hyperparameters": {},
 }
 
@@ -152,24 +152,30 @@ def predict_fn(df: pd.DataFrame, model_dict: dict) -> pd.DataFrame:
         print("Decompressing features for prediction...")
         matched_df, features = decompress_features(matched_df, features, compressed_features)
 
-    # Track missing features
-    missing_mask = matched_df[features].isna().any(axis=1)
-    if missing_mask.any():
-        print(f"Warning: {missing_mask.sum()} rows have missing features")
+    # Impute missing values (categorical with mode, continuous handled by scaler)
+    missing_counts = matched_df[features].isna().sum()
+    if missing_counts.any():
+        missing_features = missing_counts[missing_counts > 0]
+        print(f"Imputing missing values: {missing_features.to_dict()}")
+
+        # Load categorical imputation values if available
+        impute_path = os.path.join(model_dir, "categorical_impute.json")
+        if os.path.exists(impute_path):
+            with open(impute_path) as f:
+                cat_impute_values = json.load(f)
+            for col in categorical_cols:
+                if col in cat_impute_values and matched_df[col].isna().any():
+                    matched_df[col] = matched_df[col].fillna(cat_impute_values[col])
+        # Continuous features are imputed by FeatureScaler.transform() using column means
 
     # Initialize output columns
     df["prediction"] = np.nan
     if model_type in ["regressor", "uq_regressor"]:
         df["prediction_std"] = np.nan
 
-    complete_df = matched_df[~missing_mask].copy()
-    if len(complete_df) == 0:
-        print("Warning: No complete rows to predict on")
-        return df
-
-    # Prepare data for inference (with standardization)
+    # Prepare data for inference (with standardization and continuous imputation)
     x_cont, x_cat, _, _, _ = prepare_data(
-        complete_df, continuous_cols, categorical_cols, category_mappings=category_mappings, scaler=scaler
+        matched_df, continuous_cols, categorical_cols, category_mappings=category_mappings, scaler=scaler
     )
 
     # Collect ensemble predictions
@@ -191,28 +197,20 @@ def predict_fn(df: pd.DataFrame, model_dict: dict) -> pd.DataFrame:
         class_preds = np.argmax(avg_probs, axis=1)
         predictions = label_encoder.inverse_transform(class_preds)
 
-        all_proba = pd.Series([None] * len(df), index=df.index, dtype=object)
-        all_proba.loc[~missing_mask] = [p.tolist() for p in avg_probs]
-        df["pred_proba"] = all_proba
+        df["pred_proba"] = [p.tolist() for p in avg_probs]
         df = expand_proba_column(df, label_encoder.classes_)
     else:
         # Regression
         predictions = preds.flatten()
-        df.loc[~missing_mask, "prediction_std"] = preds_std.flatten()
+        df["prediction_std"] = preds_std.flatten()
 
         # Add UQ intervals if available
         if uq_models and uq_metadata:
-            X_complete = complete_df[features]
-            df_complete = df.loc[~missing_mask].copy()
-            df_complete["prediction"] = predictions  # Set prediction before compute_confidence
-            df_complete = predict_intervals(df_complete, X_complete, uq_models, uq_metadata)
-            df_complete = compute_confidence(df_complete, uq_metadata["median_interval_width"], "q_10", "q_90")
-            # Copy UQ columns back to main dataframe
-            for col in df_complete.columns:
-                if col.startswith("q_") or col == "confidence":
-                    df.loc[~missing_mask, col] = df_complete[col].values
-
-    df.loc[~missing_mask, "prediction"] = predictions
+            df["prediction"] = predictions  # Set prediction before compute_confidence
+            df = predict_intervals(df, matched_df[features], uq_models, uq_metadata)
+            df = compute_confidence(df, uq_metadata["median_interval_width"], "q_10", "q_90")
+
+    df["prediction"] = predictions
     return df
 
 
@@ -275,11 +273,11 @@ if __name__ == "__main__":
     all_df = pd.concat([pd.read_csv(f, engine="python") for f in training_files])
     check_dataframe(all_df, "training_df")
 
-    # Drop rows with missing features
+    # Drop rows with missing target (required for training)
     initial_count = len(all_df)
-    all_df = all_df.dropna(subset=features)
+    all_df = all_df.dropna(subset=[target])
     if len(all_df) < initial_count:
-        print(f"Dropped {initial_count - len(all_df)} rows with missing features")
+        print(f"Dropped {initial_count - len(all_df)} rows with missing target")
 
     print(f"Target: {target}")
     print(f"Features: {features}")
@@ -301,6 +299,23 @@ if __name__ == "__main__":
     print(f"Categorical: {categorical_cols}")
     print(f"Continuous: {len(continuous_cols)} columns")
 
+    # Report and handle missing values in features
+    # Compute categorical imputation values (mode) for use at inference time
+    cat_impute_values = {}
+    for col in categorical_cols:
+        mode_val = all_df[col].mode().iloc[0] if not all_df[col].mode().empty else all_df[col].cat.categories[0]
+        cat_impute_values[col] = str(mode_val)  # Convert to string for JSON serialization
+
+    missing_counts = all_df[features].isna().sum()
+    if missing_counts.any():
+        missing_features = missing_counts[missing_counts > 0]
+        print(f"Missing values in features (will be imputed): {missing_features.to_dict()}")
+        # Impute categorical features with mode (most frequent value)
+        for col in categorical_cols:
+            if all_df[col].isna().any():
+                all_df[col] = all_df[col].fillna(cat_impute_values[col])
+        # Continuous features are imputed by FeatureScaler.transform() using column means
+
     # -------------------------------------------------------------------------
     # Classification setup
     # -------------------------------------------------------------------------
@@ -506,6 +521,9 @@ if __name__ == "__main__":
     with open(os.path.join(args.model_dir, "feature_metadata.json"), "w") as f:
         json.dump({"continuous_cols": continuous_cols, "categorical_cols": categorical_cols}, f)
 
+    with open(os.path.join(args.model_dir, "categorical_impute.json"), "w") as f:
+        json.dump(cat_impute_values, f)
+
     with open(os.path.join(args.model_dir, "hyperparameters.json"), "w") as f:
         json.dump(hyperparameters, f, indent=2)