pypdb2pov -- PDB and mmCIF Structures as POV-Ray Scenes

pypdb2pov turns Brookhaven PDB and PDBx/mmCIF atomic structure files into POV-Ray scenes -- space-filling, ball-and-stick, or glass -- from one command or from Python.

It reads mmCIF, so structures too large for the PDB format render at all; it opens .gz, .bz2 and .xz with no decompression step; its bond search is linear rather than quadratic, so fifty thousand atoms takes under a second; it infers elements from the atom name the way the format specifies rather than guessing from the first letter; and nothing is ever dropped in silence -- skips, ambiguities and untextured atoms are all counted and reported with line numbers.

It is a Python port of pdb2pov, the 1993 C program by the same author, and keeps that program's command-line vocabulary so existing invocations still mean what they meant. Everything else here is new.

It is also a library. read_structure returns a plain list of Atom dataclasses carrying everything the readers recover, which makes it usable as a PDB/mmCIF reader in its own right and not only as a step towards a scene.

No dependencies beyond the standard library. The POV-Ray include files ship inside the package, so a pip install is the whole setup.

Author: Eric G. Suchanek, PhD -- Flux-Frontiers, Liberty TWP, OH

Sister projects

• quiltwright -- renders these scenes as multi-view quilts for Looking Glass holographic displays.
• proteusPy -- the same idea again, targeting VTK rather than a ray tracer.
• KGRAG -- the knowledge-graph federation layer the rest of the Flux-Frontiers stack is built on.

Installation

pip install pypdb2pov

From a checkout, for development:

poetry install --with dev    # pytest, ruff, ty
poetry install --with kg     # adds the pycodekg and dockg CLIs

The maintainer toolchain lives in Poetry groups, not extras, so it is never written into the published wheel -- pip install pypdb2pov pulls nothing at all. pip install pypdb2pov[dev] is deliberately not a thing.

Or don't install it at all -- the package has no dependencies, so PYTHONPATH=src python3 -m pypdb2pov ... works straight from a clone.

Python 3.12 or newer.

Quick start

pypdb2pov InputFile OutputFile [options]

pypdb2pov crambin crambin -s -h -b -d 1.5 -x 90

Filenames may be given without extensions, as they always could -- .pdb (or .cif, .ent, .atm, or a compressed variant) is appended to the input and .pov/.inc to the output. A path that exists as given is used as given, so 4hhb.cif.gz works too. - is standard input or standard output.

-h is the checkered ground, not help. It has meant that since 1993 and still does, so --help prints the usage message instead.

Scene, radii and geometry

Input, filtering and reporting

What it does

It reads mmCIF

The PDB format ran out of room years ago -- five columns for a serial number, one for a chain -- so anything above 99,999 atoms or 62 chains is distributed only as PDBx/mmCIF. Ribosomes, capsids and large complexes are mmCIF-only.

pypdb2pov 6xyz.cif.gz capsid -b -o

The reader handles the _atom_site loop with the syntax wwPDB files actually use: quoted values, semicolon-delimited text blocks, nulls, and multiple models. auth_* columns are preferred over label_* where both exist, so --chain A means the chain the literature calls A.

It opens compressed files

.gz, .bz2 and .xz are read transparently, which is how the wwPDB ships them. No decompression step, no temporary file.

It understands models

MODEL/ENDMDL are parsed as records. The default is the first model in the file, --model N picks another, and --all-models converts every one. This matters for NMR ensembles, where a naive reader stops at the first line beginning END -- which happens to be ENDMDL -- and can never reach model 7.

It infers elements properly

Elements come from columns 77-78 when they are there. When they are not -- the file predates the column -- the atom name is read the way the format specifies rather than switched on its first letter:

Three rules do the work, in order: an atom name that matches its residue name and is an element symbol is that element (ZN/ZN); inside a standard residue the element is the leading letter, full stop; otherwise the column alignment decides, because the format right-justifies a two-letter symbol into columns 13-14.

Where the answer is genuinely ambiguous -- a hand-edited FE sitting in column 14, where the format says one-letter -- the column rule wins and the conversion says so:

  1 atom name(s) read by the PDB column rule where a two-letter element was
    also possible ('FE' as F not FE);
      add an element column in 77-78, or use --legacy-elements, to override

--legacy-elements restores the first-letter guess exactly, mistakes included, for reproducing a render made with it.

It gives alternate conformations a policy

a loses a residue entirely when a depositor labels its conformers B and C with no A. first does not, and is otherwise identical on conventionally ordered files. In 1CBN that is fourteen atoms -- the side chains of Pro 22 and Leu 25 -- silently absent under a.

The choice is made per residue, which matters for microheterogeneity: in that same entry residue 22 is modelled as serine at 0.20 occupancy and proline at 0.60, sharing one sequence position. Choosing atom by atom would take proline's ring and serine's hydroxyl from the same place and draw a residue that does not exist. --altloc occupancy picks the proline, whole.

The bond search is linear, not quadratic

Comparing every pair of atoms is fine for crambin's 327 and hopeless above about ten thousand. Atoms go into a uniform grid of cells one cutoff wide, so each looks only at the 27 cells around it. 50,000 atoms takes well under a second.

The grid returns exactly what the pairwise scan returns, in the same order -- the traversal order the hydrogen rule depends on is preserved deliberately, and tests/test_bonds.py diffs the two over random structures.

--bonds covalent adds what a single cutoff cannot express -- a 1.1 Å C-H and a 2.05 Å disulphide in the same structure, without also bonding every carbon to its neighbour's neighbour.

Nothing is dropped silently, and problems have line numbers

Scanning atom file <4hhb.pdb>... got <4779> atoms.
  216 atom(s) in alternate conformations skipped (--keep-altlocs to keep them)
  2 unparseable coordinate record(s) skipped
      line 1841: ATOM   1839  CB  SER A 234      ????    ????   ????
  4 atom(s) have no dedicated texture and render as Atom_X (GD, RU)

--strict turns the skip into a failure. --info reports the same census without writing anything.

No fixed limits

No line buffer, no pre-counted atom array, no pre-count pass. Files are read once, and a long record is a long record rather than two short ones.

Library use

from pypdb2pov import convert

convert("1crn.pdb", "crambin", ball_stick=True, bond_threshold=1.9)

or, with the pieces exposed:

from pypdb2pov import (
    ParseOptions, SceneOptions, read_structure,
    find_bonds, prepare_structure, write_scene,
)

structure, stats = read_structure("4hhb.cif.gz", ParseOptions(chains="A"))
print(structure.title, len(structure), structure.element_counts())

options = SceneOptions(ball_stick=True, object_only=True, name="hemoglobin_a")
prepare_structure(structure, options)                    # rotate, centre, flip
bonds = find_bonds(structure, options.bond_threshold)
write_scene(structure, options, "hemoglobin_a.inc", bonds)

Structure is a plain list of Atom dataclasses carrying everything the readers recover -- name, element, residue, chain, insertion code, altLoc, occupancy, B-factor, formal charge, model.

Rendering

Scenes reference the bundled include files by name, so POV-Ray needs the directory on its library path:

povray +Icrambin.pov +W800 +H600 +A0.3 -D +L"$(pypdb2pov --include-dir)"

The includes are shipped inside the package, which is what makes an install self-contained.

Development

poetry install --with dev
make hooks         # install the pre-commit hooks
make test          # the suite -- the fast inner loop
make lint          # every hook over the whole tree: ruff, ty, pytest
make check         # test, then convert the bundled structures

ruff, ty and pytest run on every commit through .pre-commit-config.yaml and again in CI, so make carries only the setup step and the conversion smoke test nothing else does.

102 tests, no network, no fixtures outside tests/data. The numeric expectations are pinned to crambin, which ships with the suite: the camera distance quiltwright documents at 40.075, the enclosing sphere at 18.759, and the bond count. Both constants fall out of the deliberately truncated degree-to-radian values the C used, and both are reproduced here.

Differences from the C, in full

Everything here is deliberate. Nothing else differs.

The -t .atm path is the one place a default differs in output: the format has no element column, and the port infers elements there the same way it does for PDB rather than falling back to the 1993 guess. Pass --legacy-elements for the old behaviour.

Citation

If you use pypdb2pov in research or a project, please cite it:

APA

Suchanek, E. G. (2026). pypdb2pov: PDB and mmCIF Structures as POV-Ray Scenes (Version 0.1.1) [Software]. Flux-Frontiers. https://github.com/Flux-Frontiers/pypdb2pov

BibTeX

@software{suchanek_pypdb2pov,
  author    = {Suchanek, Eric G.},
  title     = {{pypdb2pov}: PDB and mmCIF Structures as POV-Ray Scenes},
  version   = {0.1.1},
  year      = {2026},
  publisher = {Flux-Frontiers},
  url       = {https://github.com/Flux-Frontiers/pypdb2pov},
}

License

BSD 3-Clause License. Copyright (c) 1993-2026 Eric G. Suchanek, Ph.D.

Support

• Issues -- GitHub Issues
• Sister projects: quiltwright, KGRAG
• Built on the Python standard library, and on POV-Ray

Thirty-three years of molecular graphics, still one command -- egs