Biotite

Biotite — computational molecular biology toolkit

What this is

Biotite is a comprehensive, NumPy-backed Python library for working with biological sequences and structures. Everything is built on plain ndarrays, so it is fast and interoperable, and you manipulate structures with the NumPy idioms you already know (boolean masks, fancy indexing, vectorized math). It is the toolkit of this collection — not a predictor or designer. Reach for it to prepare inputs for and analyze / score the outputs of the folding and design models.

Four pillars (one subpackage each):

Subpackage	What it does
biotite.structure	The AtomArray data model + structure I/O, filtering, superimposition, RMSD/lDDT/TM, SASA, SSE, H-bonds, geometry, the CCD (.info).
biotite.sequence	NucleotideSequence / ProteinSequence, pairwise + multiple alignment, substitution matrices, k-mer search, profiles, phylogenetics, sequence I/O.
biotite.database	Fetch + search RCSB PDB, AlphaFold DB, UniProt, NCBI Entrez, PubChem.
biotite.application / biotite.interface	Thin wrappers around external binaries (DSSP, MUSCLE, MAFFT, BLAST, Vina, ViennaRNA) and conversion bridges to RDKit / OpenMM / PyMOL.

The one mental model to internalize, read first: an AtomArray is one model of n atoms. Its coordinates are a (n, 3) float array (.coord); its per-atom metadata are parallel length-n "annotation arrays" — chain_id, res_id, res_name, atom_name, element, hetero, ins_code (plus optional b_factor, occupancy, charge, …). You select atoms by building a boolean mask over those arrays and indexing: array[array.atom_name == "CA"]. An AtomArrayStack is m models of the same atoms — coords (m, n, 3) — e.g. NMR models, a trajectory, or the N samples from a co-folding model. Units are Ångström throughout.

When to use Biotite vs. alternatives

You want…	Use
Read/edit/write PDB/mmCIF/SDF/trajectory; select atoms; compute RMSD/lDDT/TM/SASA/SSE	Biotite
Predict a complex structure from sequence + ligand	boltz, chai-lab, protenix
Design a backbone / sequence	foundry (RFdiffusion3/MPNN), genie3, boltzgen, disco
Refine/score a ligand pose in a pocket	placer
Orchestrate a multi-step pipeline on SLURM	protflow
Heavy MD analysis (many long trajectories, on-the-fly)	MDAnalysis / MDTraj (Biotite reads trajectories fine, but isn't an MD-analysis framework)
Just-need-a-quick-parse, deeply nested PDB header records	Biopython is an alternative; Biotite is faster and more array-friendly

The common role in a campaign: design (foundry/genie3/boltzgen) → predict (boltz/chai-lab/protenix) → Biotite to superimpose the prediction on the design, compute backbone RMSD / interface lDDT / TM-score, flag clashes, and filter survivors → keep the winners.

Install

pip install biotite              # pulls numpy, requests, msgpack, networkx
# or:  conda install -c conda-forge biotite

Optional extras, only if you use them:

• biotite.interface.rdkit → pip install rdkit (≥ 2024.09.1)
• biotite.interface.openmm → conda install -c conda-forge openmm
• biotite.interface.pymol → PyMOL (open-source build) on the env
• biotite.application.* → the external binary on PATH (mkdssp, muscle, mafft, clustalo, vina, RNAfold, …)

Requires Python ≥ 3.11. Details + the per-wrapper binary list: references/installation.md.

Quickstart — fetch, load, analyze, save

import biotite.database.rcsb as rcsb
import biotite.structure as struc
import biotite.structure.io as strucio

# 1) Fetch a structure from the RCSB PDB (BinaryCIF is smallest/fastest)
path = rcsb.fetch("1AKI", "bcif", target_path="/tmp")

# 2) Load it into an AtomArray (first model)
structure = strucio.load_structure(path)        # AtomArray (n atoms)

# 3) Select: protein CA atoms of chain A, via boolean masks on annotations
mask = (
    struc.filter_amino_acids(structure)
    & (structure.atom_name == "CA")
    & (structure.chain_id == "A")
)
ca = structure[mask]
print(ca.array_length(), "residues")

# 4) Analyze
sasa_per_atom = struc.sasa(structure)            # Shrake-Rupley, Ų
sse           = struc.annotate_sse(structure)    # per-residue 'a'/'b'/'c'

# 5) Save back out (format inferred from extension)
strucio.save_structure("/tmp/chainA_ca.pdb", ca)

Compare two structures (the bread-and-butter design-validation step):

# superimpose `pred` onto `design` over matched atoms, then measure deviation
fitted, transform = struc.superimpose(design, pred)   # Kabsch
idx = np.arange(design.array_length())                # design & pred already 1:1 (e.g. matched CA)
print("RMSD:", struc.rmsd(design, fitted))
print("TM:",   struc.tm_score(design, fitted, idx, idx))   # needs corresponding indices
print("lDDT:", struc.lddt(design, pred))              # superposition-free

The I/O pattern (read this once)

Two equivalent ways to get a structure in and out:

# (a) Shortcuts — format inferred from the file extension:
import biotite.structure.io as strucio
atoms = strucio.load_structure("x.cif")      # AtomArray (or AtomArrayStack if model=None and multi-model)
strucio.save_structure("y.pdb", atoms)

# (b) Explicit File object — needed for headers, assemblies, bonds, components:
import biotite.structure.io.pdbx as pdbx
f = pdbx.CIFFile.read("x.cif")               # or BinaryCIFFile.read("x.bcif")
atoms = pdbx.get_structure(f, model=1, include_bonds=True, extra_fields=["b_factor"])
# ...edit atoms...
out = pdbx.CIFFile()
pdbx.set_structure(out, atoms)
out.write("y.cif")

• model=1 → an AtomArray (one model). Omit model → an AtomArrayStack (all models). mmCIF / BinaryCIF is the preferred modern format (PDB has hard column-width limits and can't represent large systems); pdbx reads both.
• For trajectories you must pass a topology template: XTCFile.read("traj.xtc").get_structure(template).
• Ligands / small molecules: biotite.structure.io.mol (MOLFile, SDFile).

Full format matrix (PDB / PDBx / BinaryCIF / MOL / SDF / GRO / XTC / TRR / DCD / NetCDF) + assemblies, altlocs, bonds, and sequence-file I/O (FASTA / FASTQ / GenBank / GFF): references/file-io.md.

What lives where (reference index)

• references/installation.md — pip/conda, Python version, optional-dependency matrix (rdkit/openmm/pymol), and the external binaries each biotite.application wrapper needs.
• references/structure.md — the biotite.structure workhorse: Atom / AtomArray / AtomArrayStack / BondList, annotations, indexing & the filter_* helpers, residue/chain iteration, superimpose* / rmsd / rmsf / lddt / tm_score, sasa, annotate_sse, hbond, geometry (distance/angle/dihedral), transforms, and the CCD (structure.info).
• references/sequence.md — biotite.sequence: sequence types & translation, align_optimal / align_multiple, SubstitutionMatrix, Alignment & identity, the KmerTable fast-search workflow, SequenceProfile, and phylo trees.
• references/file-io.md — every structure & sequence file format, the read → get_structure / set_structure → write contract, load_structure / save_structure, assemblies, altlocs, bonds, trajectories.
• references/database.md — rcsb / entrez / uniprot / afdb / pubchem fetch + the search / Query builders (FieldQuery, SequenceQuery, …).
• references/applications.md — the LocalApp / WebApp lifecycle and each wrapper (DSSP, MUSCLE/MAFFT/Clustal, BLAST, Vina, ViennaRNA, tantan, SRA), plus the biotite.interface bridges (RDKit, OpenMM, PyMOL).
• references/troubleshooting.md — common pitfalls (views vs. copies, missing bonds, altloc/insertion codes, PDB width limits, element inference, web rate-limits).
• examples/quickstart.py — a runnable end-to-end script.
• examples/recipes.md — copy-paste snippets by task.

Gotchas (read before you debug)

1. Indexing can return a view, not a copy. sub = array[mask] may share the underlying annotation arrays; mutate via .copy() when in doubt (NumPy semantics).
2. Many functions need bonds. sasa on hydrogens, charge/aromaticity, RDKit export, and some filters expect a BondList. Load with include_bonds=True, or build one with struc.connect_via_residue_names / connect_via_distances.
3. superimpose does not match atoms for you. Both inputs must be atom-for-atom corresponding (e.g. both filtered to CA). For differing sequences use superimpose_homologs; for distant structures superimpose_structural_homologs (TM-based). rmsd itself does no alignment.
4. PDB has hard limits (≤ 99,999 atoms, 4-char chain ids via hybrid-36, narrow columns). Use mmCIF / BinaryCIF for anything large or programmatic.
5. load_structure returns a Stack for multi-model files unless you pass model=. Code that assumes an AtomArray will break on NMR/predicted ensembles.
6. Element column matters. Some PDBs lack it; downstream mass/radius/bond logic then misbehaves. extra_fields and repair/info helpers fix this.
7. application wrappers shell out to a binary that must be installed and on PATH; interface bridges import an optional package. Neither ships with Biotite.
8. SSE codes are lowercase 'a' (alpha-helix), 'b' (beta-strand), 'c' (coil) from annotate_sse — not DSSP's 8-state. Use DsspApp for 8-state.

More failure modes + fixes: references/troubleshooting.md.

Citation

Biotite is BSD-3-Clause, developed by Patrick Kunzmann & the Biotite contributors. If you use it in published work:

Kunzmann, P. & Hamacher, K.
Biotite: a unifying open source computational biology framework in Python.
BMC Bioinformatics 19, 346 (2018). https://doi.org/10.1186/s12859-018-2367-z

Kunzmann, P. et al.
Biotite: new tools for a versatile Python bioinformatics library.
BMC Bioinformatics 24, 236 (2023). https://doi.org/10.1186/s12859-023-05345-6

Docs & tutorial: https://www.biotite-python.org