diff --git a/src/beignet/transforms/__init__.py b/src/beignet/transforms/__init__.py
index c2e6212e80..a1a5a7aee7 100644
--- a/src/beignet/transforms/__init__.py
+++ b/src/beignet/transforms/__init__.py
@@ -7,8 +7,6 @@
 from ._binarize import BinarizeTransform
 from ._lambda import Lambda
 from ._protein_tokenizer_transform import ProteinTokenizerTransform
-from ._structure import trim_or_pad
-from ._structure_featurizer import StructureFeaturizer
 from ._transform import Transform
 
 __all__ = [
diff --git a/src/beignet/transforms/_structure.py b/src/beignet/transforms/_structure.py
deleted file mode 100644
index d226023296..0000000000
--- a/src/beignet/transforms/_structure.py
+++ /dev/null
@@ -1,18 +0,0 @@
-import torch
-
-
-def trim_or_pad(tensor: torch.Tensor, pad_to: int, pad_idx: int = 0):
-    """Trim or pad a tensor with shape (L, ...) to a given length."""
-    L = tensor.shape[0]
-    if L >= pad_to:
-        # trim, assuming first dimension is the dim to trim
-        tensor = tensor[:pad_to]
-    elif L < pad_to:
-        padding = torch.full(
-            size=(pad_to - tensor.shape[0], *tensor.shape[1:]),
-            fill_value=pad_idx,
-            dtype=tensor.dtype,
-            device=tensor.device,
-        )
-        tensor = torch.concat((tensor, padding), dim=0)
-    return tensor
diff --git a/src/beignet/transforms/_structure_featurizer.py b/src/beignet/transforms/_structure_featurizer.py
deleted file mode 100644
index 7a670f8c5e..0000000000
--- a/src/beignet/transforms/_structure_featurizer.py
+++ /dev/null
@@ -1,104 +0,0 @@
-import typing as T
-import warnings
-from pathlib import Path
-
-import numpy as np
-import torch
-from lobster.model.openfold_utils import (
-    OFProtein,
-    atom37_to_frames,
-    get_backbone_frames,
-    make_atom14_masks,
-    make_atom14_positions,
-    make_pdb_features,
-    protein_from_pdb_string,
-)
-from lobster.transforms import trim_or_pad
-
-PathLike = T.Union[Path, str]
-
-
-class StructureFeaturizer:
-    def _openfold_features_from_pdb(
-        self, pdb_str: str, pdb_id: T.Optional[str] = None
-    ) -> OFProtein:
-        """Create rigid groups from a PDB file on disk.
-
-        The inputs to the Frame-Aligned Point Error (FAPE) loss used in AlphaFold2 are
-        tuples of translations and rotations from the reference frame. In the OpenFold
-        implementation, this is stored as `Rigid` objects. This function calls the
-        OpenFold wrapper functions which creates an `OFProtein` object,
-        and then extracts several `Rigid` objects.
-
-        Args:
-            pdb_str (str): String representing the contents of a PDB file
-
-        Returns:
-            OFProtein: _description_
-        """
-        pdb_id = "" if pdb_id is None else pdb_id
-        protein_object = protein_from_pdb_string(pdb_str)
-
-        # TODO: what is the `is_distillation` argument?
-        protein_features = make_pdb_features(
-            protein_object, description=pdb_id, is_distillation=False
-        )
-
-        return protein_features
-
-    def _process_structure_features(
-        self, features: T.Dict[str, np.ndarray], seq_len: T.Optional[int] = None
-    ):
-        """Process feature dtypes and pad to max length for a single sequence."""
-        features_requiring_padding = [
-            "aatype",
-            "between_segment_residues",
-            "residue_index",
-            "all_atom_positions",
-            "all_atom_mask",
-            # ... add additionals here.
-        ]
-
-        for k, v in features.items():
-            # Handle data types in converting from numpy to torch
-            if v.dtype == np.dtype("int32"):
-                features[k] = torch.from_numpy(v).long()  # int32 -> int64
-            elif v.dtype == np.dtype("O"):
-                features[k] = v.astype(str)[0]
-            else:
-                # the rest are all float32. TODO: does this be float64?
-                features[k] = torch.from_numpy(v)
-
-            # Trim or pad to a fixed length for all per-specific features
-            if (k in features_requiring_padding) and (seq_len is not None):
-                features[k] = trim_or_pad(features[k], seq_len)
-
-            # 'seq_length' is a tensor with shape equal to the aatype array length,
-            # and filled with the value of the original sequence length.
-            if k == "seq_length":
-                features[k] = torch.full((seq_len,), features[k][0])
-
-        # Make the mask
-        idxs = torch.arange(seq_len, dtype=torch.long)
-        mask = idxs < features["seq_length"]
-        features["mask"] = mask.long()
-
-        # Make sure input sequence string is also trimmed
-        if seq_len is not None:
-            features["sequence"] = features["sequence"][:seq_len]
-
-        features["aatype"] = features["aatype"].argmax(dim=-1)
-        return features
-
-    def __call__(self, pdb_str: str, seq_len: int, pdb_id: T.Optional[str] = None):
-        with warnings.catch_warnings():
-            warnings.simplefilter("ignore")
-            features = self._openfold_features_from_pdb(pdb_str, pdb_id)
-
-        features = self._process_structure_features(features, seq_len)
-        features = atom37_to_frames(features)
-        features = get_backbone_frames(features)
-        features = make_atom14_masks(features)
-        features = make_atom14_positions(features)
-
-        return features