diff --git a/src/structure_io.py b/src/structure_io.py
new file mode 100644
index 0000000..d06f545
--- /dev/null
+++ b/src/structure_io.py
@@ -0,0 +1,162 @@
+# structure_io.py
+
+"""Merge predicted waters with the protein+het atoms and write the result to PDB or CIF."""
+
+from __future__ import annotations
+
+import biotite.structure as bts
+import numpy as np
+from biotite.structure.io.pdb import PDBFile
+from biotite.structure.io.pdbx import CIFFile, set_structure as _set_structure_cif
+
+
+# Columns every AtomArray has. Any other column on the kept atoms (e.g. b_factor,
+# occupancy) must be mirrored onto the waters so the two arrays concatenate.
+_MANDATORY = (
+    "chain_id",
+    "res_id",
+    "ins_code",
+    "res_name",
+    "hetero",
+    "atom_name",
+    "element",
+)
+
+# Default water B-factor: mean B of kept atoms within this radius (A), plus an
+# offset since ordered waters move more than the atoms they touch.
+_B_FACTOR_CONTACT_RADIUS = 5.0
+_B_FACTOR_WATER_OFFSET = 10.0
+
+
+def merge_waters(
+    atoms: bts.AtomArray,
+    positions,
+    *,
+    chain_id: str | None = None,
+    b_factor: float | None = None,
+    occupancy: float = 1.0,
+) -> bts.AtomArray:
+    """Return the kept atoms with predicted waters appended as HOH oxygens.
+
+    Each row of positions becomes one O atom in its own HOH residue, in whatever
+    frame positions are given (no re-centering here).
+
+    Args:
+        atoms: Non-water atoms to keep (protein + hets), in the output frame.
+        positions: (N, 3) water coordinates; numpy array or tensor.
+        chain_id: Chain for the waters. Defaults to an unused single character.
+        b_factor: One B-factor for every water. Defaults to a per-water estimate
+            from the nearby atoms (see _default_b_factors). Written only if atoms
+            has a b_factor column.
+        occupancy: Occupancy for every water. Written only if atoms has an
+            occupancy column.
+
+    Returns:
+        The kept atoms followed by the waters, as one AtomArray.
+    """
+    coords = _to_coords(positions)
+    if chain_id is None:
+        chain_id = _pick_unused_chain_id(atoms)
+    if b_factor is None:
+        b_factor = _default_b_factors(atoms, coords)
+
+    waters = _water_array(
+        coords,
+        chain_id=chain_id,
+        b_factor=b_factor,
+        occupancy=occupancy,
+        template=atoms,
+    )
+    return atoms + waters
+
+
+def write_structure(atoms: bts.AtomArray, output_path: str) -> None:
+    """Write an AtomArray to PDB or CIF, chosen by the file extension.
+
+    A .cif extension writes mmCIF; anything else writes PDB.
+    """
+    if str(output_path).endswith(".cif"):
+        cif_file = CIFFile()
+        _set_structure_cif(cif_file, atoms)
+        cif_file.write(str(output_path))
+    else:
+        pdb_file = PDBFile()
+        pdb_file.set_structure(atoms)
+        pdb_file.write(str(output_path))
+
+
+def _water_array(
+    coords: np.ndarray,
+    *,
+    chain_id: str,
+    b_factor,
+    occupancy: float,
+    template: bts.AtomArray,
+) -> bts.AtomArray:
+    """Build oxygen-only HOH waters with the same columns as template.
+
+    b_factor may be a scalar or a per-water array.
+    """
+    n = len(coords)
+
+    waters = bts.AtomArray(n)
+    waters.coord = coords
+    waters.chain_id[:] = chain_id
+    waters.res_id[:] = np.arange(1, n + 1)
+    waters.ins_code[:] = ""
+    waters.res_name[:] = "HOH"
+    waters.atom_name[:] = "O"
+    waters.element[:] = "O"
+    waters.hetero[:] = True
+
+    # template + waters only concatenates if both have the same columns. Give the
+    # waters every extra column the kept atoms carry so they line up. A new column
+    # starts empty (0 / '' / False), which is fine except for b_factor and
+    # occupancy, which get a real value.
+    for cat in template.get_annotation_categories():
+        if cat in _MANDATORY:
+            continue
+        waters.add_annotation(cat, dtype=template.get_annotation(cat).dtype)
+        if cat == "b_factor":
+            waters.get_annotation(cat)[:] = b_factor
+        elif cat == "occupancy":
+            waters.get_annotation(cat)[:] = occupancy
+    return waters
+
+
+def _pick_unused_chain_id(atoms: bts.AtomArray) -> str:
+    """A single-character chain id not used by atoms (falls back to 'W')."""
+    used = set(atoms.chain_id.tolist()) if atoms.array_length() else set()
+    for c in "WXYZUVTSRQ0123456789":
+        if c not in used:
+            return c
+    return "W"
+
+
+def _default_b_factors(atoms: bts.AtomArray, coords: np.ndarray) -> np.ndarray:
+    """Per-water default B-factor from the local environment.
+
+    Each water takes the mean B-factor of the kept atoms near it, plus an offset.
+    A water with no atom in range uses the overall mean. Returns zeros when atoms
+    has no b_factor.
+    """
+
+    if len(coords) == 0:
+        return np.zeros(0, dtype=np.float32)
+    if "b_factor" not in atoms.get_annotation_categories() or atoms.array_length() == 0:
+        return np.zeros(len(coords), dtype=np.float32)
+
+    b = atoms.b_factor
+    cell_list = bts.CellList(atoms, cell_size=_B_FACTOR_CONTACT_RADIUS)
+    within = cell_list.get_atoms(coords, radius=_B_FACTOR_CONTACT_RADIUS, as_mask=True)
+    counts = within.sum(axis=1)
+    local_mean = (within * b[None, :]).sum(axis=1) / np.maximum(counts, 1)
+    base = np.where(counts > 0, local_mean, b.mean())
+    return (base + _B_FACTOR_WATER_OFFSET).astype(np.float32)
+
+
+def _to_coords(positions) -> np.ndarray:
+    """Tensor or array positions to an (N, 3) float32 array."""
+    if hasattr(positions, "detach"):  # torch.Tensor
+        positions = positions.detach().cpu().numpy()
+    return np.asarray(positions, dtype=np.float32).reshape(-1, 3)
diff --git a/tests/test_structure_io.py b/tests/test_structure_io.py
new file mode 100644
index 0000000..183039b
--- /dev/null
+++ b/tests/test_structure_io.py
@@ -0,0 +1,156 @@
+"""Unit tests for src/structure_io.py -- merging predicted waters and writing PDB/CIF."""
+
+import biotite.structure as bts
+import numpy as np
+import pytest
+import torch
+from biotite.structure.io.pdb import PDBFile
+from biotite.structure.io.pdbx import CIFFile, get_structure as get_structure_cif
+
+from src.structure_io import merge_waters, write_structure
+
+
+def _make_protein(n=4, chain="A", b_factor=20.0, extra_occupancy=False):
+    """A minimal protein AtomArray shaped like parse_asu_with_biotite's output
+    (carries a b_factor field), optionally with an occupancy field too."""
+    atoms = bts.AtomArray(n)
+    atoms.coord = np.arange(n * 3, dtype=np.float32).reshape(n, 3)
+    atoms.chain_id[:] = chain
+    atoms.res_id[:] = np.arange(1, n + 1)
+    atoms.ins_code[:] = ""
+    atoms.res_name[:] = "ALA"
+    atoms.atom_name[:] = "CA"
+    atoms.element[:] = "C"
+    atoms.hetero[:] = False
+    atoms.add_annotation("b_factor", dtype=float)
+    atoms.b_factor[:] = b_factor
+    if extra_occupancy:
+        atoms.add_annotation("occupancy", dtype=float)
+        atoms.occupancy[:] = 1.0
+    return atoms
+
+
+def _read_back(path):
+    if str(path).endswith(".cif"):
+        return get_structure_cif(CIFFile.read(str(path)), model=1)
+    return PDBFile.read(str(path)).get_structure(model=1)
+
+
+@pytest.mark.unit
+class TestMergeWaters:
+    def test_appends_waters_as_hoh_oxygens(self):
+        prot = _make_protein(4)
+        pos = np.array([[10.0, 0, 0], [11.0, 0, 0]], dtype=np.float32)
+
+        merged = merge_waters(prot, pos)
+
+        assert merged.array_length() == 4 + 2
+        w = merged[merged.res_name == "HOH"]
+        assert w.array_length() == 2
+        assert (w.element == "O").all()
+        assert (w.atom_name == "O").all()
+        assert w.hetero.all()
+        # positions written verbatim -- no re-centering in the IO layer
+        assert np.allclose(w.coord, pos)
+        # kept atoms are untouched and come first
+        assert (merged[:4].res_name == "ALA").all()
+
+    def test_water_chain_avoids_collision(self):
+        prot = _make_protein(chain="A")
+        merged = merge_waters(prot, np.zeros((1, 3), dtype=np.float32))
+        assert merged[merged.res_name == "HOH"].chain_id[0] == "W"
+
+        # when W is taken, the picker moves on
+        prot_w = _make_protein(chain="W")
+        merged_w = merge_waters(prot_w, np.zeros((1, 3), dtype=np.float32))
+        assert merged_w[merged_w.res_name == "HOH"].chain_id[0] != "W"
+
+    def test_water_res_ids_are_sequential(self):
+        prot = _make_protein()
+        merged = merge_waters(prot, np.zeros((3, 3), dtype=np.float32))
+        assert merged[merged.res_name == "HOH"].res_id.tolist() == [1, 2, 3]
+
+    def test_b_factor_defaults_to_local_mean_plus_offset(self):
+        # Two atoms within 5 A of the water (B = 10, 30) and one far away (B = 80);
+        # the water should take mean(10, 30) + 10.0.
+        prot = bts.AtomArray(3)
+        prot.coord = np.array([[0, 0, 0], [2, 0, 0], [100, 0, 0]], dtype=np.float32)
+        prot.chain_id[:] = "A"
+        prot.res_id[:] = [1, 2, 3]
+        prot.ins_code[:] = ""
+        prot.res_name[:] = "ALA"
+        prot.atom_name[:] = "CA"
+        prot.element[:] = "C"
+        prot.hetero[:] = False
+        prot.add_annotation("b_factor", dtype=float)
+        prot.b_factor[:] = [10.0, 30.0, 80.0]
+
+        merged = merge_waters(prot, np.array([[1.0, 0, 0]], dtype=np.float32))
+        assert np.allclose(merged[merged.res_name == "HOH"].b_factor, 20.0 + 10.0)
+
+    def test_b_factor_falls_back_to_global_mean_when_isolated(self):
+        # A water with no atom within 5 A uses the overall mean + offset.
+        prot = _make_protein(b_factor=42.0)
+        merged = merge_waters(prot, np.array([[1000.0, 0, 0]], dtype=np.float32))
+        assert np.allclose(merged[merged.res_name == "HOH"].b_factor, 42.0 + 10.0)
+
+    def test_explicit_b_factor_overrides_default(self):
+        prot = _make_protein(b_factor=42.0)
+        merged = merge_waters(prot, np.zeros((1, 3), dtype=np.float32), b_factor=5.0)
+        assert np.allclose(merged[merged.res_name == "HOH"].b_factor, 5.0)
+
+    def test_accepts_torch_tensor_positions(self):
+        prot = _make_protein()
+        pos = torch.tensor([[1.0, 2.0, 3.0]], dtype=torch.float32)
+        merged = merge_waters(prot, pos)
+        assert np.allclose(merged[merged.res_name == "HOH"].coord, pos.numpy())
+
+    def test_empty_positions_returns_atoms_unchanged(self):
+        prot = _make_protein(4)
+        merged = merge_waters(prot, np.zeros((0, 3), dtype=np.float32))
+        assert merged.array_length() == 4
+        assert (merged.res_name == "HOH").sum() == 0
+
+    def test_matches_extra_annotations_for_concat(self):
+        # atoms carrying occupancy (an extra field) must still merge cleanly,
+        # with waters mirroring the field.
+        prot = _make_protein(extra_occupancy=True)
+        merged = merge_waters(prot, np.zeros((2, 3), dtype=np.float32), occupancy=0.5)
+        w = merged[merged.res_name == "HOH"]
+        assert "occupancy" in merged.get_annotation_categories()
+        assert np.allclose(w.occupancy, 0.5)
+
+
+@pytest.mark.unit
+class TestWriteStructure:
+    @pytest.mark.parametrize("ext", [".pdb", ".cif"])
+    def test_round_trip_preserves_waters(self, tmp_path, ext):
+        prot = _make_protein(5)
+        pos = np.array([[20.0, 1, 2], [21.0, 3, 4]], dtype=np.float32)
+        merged = merge_waters(prot, pos)
+
+        out = tmp_path / f"pred{ext}"
+        write_structure(merged, str(out))
+        assert out.exists()
+
+        back = _read_back(out)
+        assert back.array_length() == 7
+        w = back[back.res_name == "HOH"]
+        assert w.array_length() == 2
+        assert np.allclose(np.sort(w.coord[:, 0]), [20.0, 21.0])
+
+    def test_extension_selects_format(self, tmp_path):
+        prot = _make_protein(2)
+        merged = merge_waters(prot, np.zeros((1, 3), dtype=np.float32))
+
+        pdb_path = tmp_path / "s.pdb"
+        cif_path = tmp_path / "s.cif"
+        write_structure(merged, str(pdb_path))
+        write_structure(merged, str(cif_path))
+
+        assert (
+            pdb_path.read_text()
+            .lstrip()
+            .startswith(("HEADER", "ATOM", "HETATM", "CRYST", "MODEL"))
+        )
+        assert cif_path.read_text().lstrip().startswith(("data_", "#"))