mdtraj-trajectory-analysis

name	mdtraj-trajectory-analysis
description	mdtraj molecular dynamics trajectory analysis (Python). Reads DCD/XTC/TRR/NetCDF/H5/PDB topologies and trajectories; computes RMSD vs time, radius of gyration, per-residue RMSF, residue-residue contact frequency maps, phi/psi torsions for Ramachandran plots (general + Gly/Pro), and 8-state DSSP secondary structure. Modules: trajectory I/O, geometry (distances/angles/dihedrals), structural analysis (RMSD/Rg/RMSF/SASA), contacts, hydrogen bonds, secondary structure (DSSP), NMR observables. For broader atom-selection grammar use mdanalysis-trajectory; for running MD simulations use OpenMM/GROMACS.
license	LGPL-2.1

mdtraj Trajectory Analysis

Overview

mdtraj is a dependency-light Python library for analyzing MD trajectories. Reads DCD/XTC/TRR/NetCDF/H5/AMBER/GROMACS/CHARMM/OpenMM into a Trajectory object backed by NumPy arrays, then exposes geometry, RMSD/Rg/RMSF/SASA, contacts, hydrogen bonds, torsions, and 8-state DSSP as pure-Python functions.

Units: mdtraj uses nm and ps internally. Multiply distances by 10 for Å, divide time by 1000 for ns. Torsions are in radians — np.degrees().

When to Use

• RMSD vs time, Rg, per-residue RMSF for stability across MD replicates
• Residue-residue contact frequency maps from a trajectory ensemble
• Backbone phi/psi for Ramachandran (general, Gly, Pro)
• 8-state DSSP per residue per frame for secondary-structure time series
• Lightweight ad-hoc analyses where MDAnalysis's full framework is overkill
• NMR observables (J-couplings, chemical shifts via SHIFTX2)
• Use mdanalysis-trajectory instead when you need MDAnalysis's selection grammar, AnalysisBase parallelism, or LAMMPS/NAMD-specific readers
• Use OpenMM/GROMACS to run the simulation — this skill is post-simulation only

Prerequisites

• mdtraj, numpy, pandas, matplotlib
• Topology (PDB/GRO/PRMTOP/PSF) + trajectory (DCD/XTC/TRR/NetCDF/H5)
• Python 3.9+, conda-forge build recommended

Check before installing — inside a pixi/conda env mdtraj is usually present:

python3 -c "import mdtraj" 2>/dev/null || conda install -c conda-forge mdtraj numpy pandas matplotlib

Quick Start

import mdtraj as md

traj = md.load("traj.xtc", top="topology.pdb")
ca = traj.topology.select("name CA")
traj.superpose(traj, frame=0, atom_indices=ca)

rmsd_ang = md.rmsd(traj, traj, frame=0, atom_indices=ca) * 10.0  # nm -> Å
print(f"Frames: {traj.n_frames}  RMSD: {rmsd_ang.min():.2f}–{rmsd_ang.max():.2f} Å")

Core API

Module 1: Trajectory I/O

Load whole or streamed. Format auto-detected from extension.

import mdtraj as md

traj = md.load("rep1.xtc", top="protein.pdb")

# Stream large trajectories — avoids OOM
for chunk in md.iterload("rep1.xtc", top="protein.pdb", chunk=500):
    rmsd_chunk = md.rmsd(chunk, chunk, frame=0)

# Save subset
ca = traj.topology.select("name CA")
traj.atom_slice(ca).save_dcd("ca_only.dcd")

Selecting atoms and slicing frames:

backbone = traj.topology.select("backbone")
chain_a  = traj.topology.select("chainid 0")

first_ns      = traj[:1000]
every_10th    = traj[::10]
last_half_bb  = traj[traj.n_frames // 2:].atom_slice(backbone)

Module 2: RMSD, Rg, RMSF

md.rmsd superposes internally; RMSF you compute manually after explicit superpose.

import mdtraj as md, numpy as np

traj = md.load("rep1.xtc", top="protein.pdb")
ca = traj.topology.select("name CA")

rmsd_ang = md.rmsd(traj, traj, frame=0, atom_indices=ca) * 10.0    # Å
rg_ang   = md.compute_rg(traj) * 10.0                              # Å
time_ns  = traj.time / 1000.0
print(f"<RMSD>={rmsd_ang.mean():.2f} Å,  <Rg>={rg_ang.mean():.2f} Å")

# Per-CA RMSF — average-structure reference
ca_traj = traj.atom_slice(ca)
ca_traj.superpose(ca_traj, frame=0)
diff     = ca_traj.xyz - ca_traj.xyz.mean(axis=0)
rmsf_ang = np.sqrt((diff ** 2).sum(axis=2).mean(axis=0)) * 10.0
res_ids  = [a.residue.resSeq for a in ca_traj.topology.atoms]
print(f"Max RMSF: residue {res_ids[np.argmax(rmsf_ang)]} = {rmsf_ang.max():.2f} Å")

Module 3: Contacts and Distance Maps

Threshold distances to get contact frequency.

import mdtraj as md, numpy as np

traj = md.load("rep1.xtc", top="protein.pdb")
distances_nm, pairs = md.compute_contacts(traj, contacts="all", scheme="closest-heavy")
# distances_nm: (n_frames, n_pairs);  pairs: (n_pairs, 2) of residue indices

contact_freq = (distances_nm < 0.5).mean(axis=0)                  # 5 Å cutoff
n_res    = traj.n_residues
freq_map = np.zeros((n_res, n_res))
for (i, j), f in zip(pairs, contact_freq):
    freq_map[i, j] = freq_map[j, i] = f
print(f"Persistent contacts (>0.8): {(contact_freq > 0.8).sum()}")

Module 4: Torsions (Ramachandran)

phi/psi returned in radians; intersect on residue since first residue has no phi and last has no psi.

import mdtraj as md, numpy as np

traj = md.load("rep1.xtc", top="protein.pdb")
phi_ix, phi_rad = md.compute_phi(traj)
psi_ix, psi_rad = md.compute_psi(traj)

def res_of(indices): return np.array([traj.topology.atom(ix[1]).residue.index for ix in indices])
phi_res, psi_res = res_of(phi_ix), res_of(psi_ix)
common  = np.intersect1d(phi_res, psi_res)
phi_deg = np.degrees(phi_rad[:, np.isin(phi_res, common)])
psi_deg = np.degrees(psi_rad[:, np.isin(psi_res, common)])

Filter to Gly / Pro residues:

res_names = [traj.topology.residue(r).name for r in common]
gly_cols  = [i for i, n in enumerate(res_names) if n == "GLY"]
pro_cols  = [i for i, n in enumerate(res_names) if n == "PRO"]
phi_gly, psi_gly = phi_deg[:, gly_cols].ravel(), psi_deg[:, gly_cols].ravel()
phi_pro, psi_pro = phi_deg[:, pro_cols].ravel(), psi_deg[:, pro_cols].ravel()

Module 5: DSSP Secondary Structure (8-state)

md.compute_dssp(traj, simplified=False) returns (n_frames, n_residues) of one-character codes:

Code	Meaning
H	alpha-helix
B	beta-bridge
E	beta-sheet
G	3_10-helix
I	pi-helix
T	turn
S	bend
(space) or C	coil

import mdtraj as md

traj = md.load("rep1.xtc", top="protein.pdb")
dssp = md.compute_dssp(traj, simplified=False)
dssp[dssp == " "] = "C"
print(f"DSSP grid: {dssp.shape}  unique codes: {set(dssp.flatten())}")

Module 6: Hydrogen Bonds and SASA

import mdtraj as md
import pandas as pd

traj = md.load("rep1.xtc", top="protein.pdb")

# Baker–Hubbard: returns (n_hbonds, 3) of [donor, H, acceptor] atom indices
hbonds = md.baker_hubbard(traj, freq=0.5, periodic=False)
label  = lambda a: f"{traj.topology.atom(a).residue}-{traj.topology.atom(a).name}"
hbonds_df = pd.DataFrame({"donor": [label(d) for d, _, _ in hbonds],
                         "acceptor": [label(a) for _, _, a in hbonds]})

# SASA in nm^2 per residue per frame
sasa_nm2 = md.shrake_rupley(traj, mode="residue")
total_ang2 = sasa_nm2.sum(axis=1) * 100.0                # nm^2 -> Å^2
print(f"Persistent H-bonds: {len(hbonds_df)},  <SASA>: {total_ang2.mean():.0f} Ų")

Key Concepts

Units (nm, ps, radians)

traj.xyz is in nm, traj.time in ps, torsions in radians. Convert before plotting:

import numpy as np
xyz_ang  = traj.xyz * 10.0
time_ns  = traj.time / 1000.0
phi_deg  = np.degrees(phi_rad)

Pair indices vs resSeq

md.compute_contacts(..., contacts="all") returns 0-based residue indices (contiguous across chains), not PDB resSeq. Reshape into (n_res, n_res) before heatmaps. Map back via traj.topology.residue(i).resSeq only when labeling.

8-state vs simplified DSSP

simplified=False → 8 states. simplified=True → 3 states (H/E/C), which loses 3_10 (G), pi-helix (I), beta-bridge (B), bend (S), turn (T).

Common Workflows

Workflow 1: Multi-Replicate Stability + Flexibility + Contacts + Ramachandran + DSSP

Goal: structural-analysis report for three MD replicates — RMSD/Rg over time, per-residue RMSF, residue contact frequency map (rep1), Ramachandran general + Gly + Pro (rep1), DSSP 8-state time series (rep1). Plots use matplotlib defaults; the consumer picks palette downstream.

import mdtraj as md, numpy as np, matplotlib.pyplot as plt
from pathlib import Path

replicas = {"rep1": "rep1.xtc", "rep2": "rep2.xtc", "rep3": "rep3.xtc"}
topology = "protein.pdb"
outdir = Path("figures"); outdir.mkdir(exist_ok=True)

trajs = {name: md.load(p, top=topology) for name, p in replicas.items()}
ca    = next(iter(trajs.values())).topology.select("name CA")

# 1. RMSD vs time + Rg (Å) for all replicates
fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(12, 4))
for name, traj in trajs.items():
    traj.superpose(traj, frame=0, atom_indices=ca)
    ax1.plot(traj.time / 1000.0, md.rmsd(traj, traj, frame=0, atom_indices=ca) * 10.0, lw=1, label=name)
    ax2.plot(traj.time / 1000.0, md.compute_rg(traj) * 10.0, lw=1, label=name)
ax1.set(xlabel="Time (ns)", ylabel="RMSD (Å)", title="Backbone RMSD")
ax2.set(xlabel="Time (ns)", ylabel="Rg (Å)",   title="Radius of gyration")
for ax in (ax1, ax2): ax.legend(frameon=False)
fig.tight_layout(); fig.savefig(outdir / ); plt.close(fig)


fig, ax = plt.subplots(figsize=(, ))
 name, traj  trajs.items():
    ct = traj.atom_slice(ca); ct.superpose(ct, frame=)
    rmsf = np.sqrt(((ct.xyz - ct.xyz.mean(axis=)) ** ).(axis=).mean(axis=)) * 
    ax.plot([a.residue.resSeq  a  ct.topology.atoms], rmsf, lw=, label=name)
ax.(xlabel=, ylabel=, title=)
ax.legend(frameon=); fig.tight_layout()
fig.savefig(outdir / ); plt.close(fig)


rep1 = trajs[]
dist_nm, pairs = md.compute_contacts(rep1, contacts=, scheme=)
freq = (dist_nm < ).mean(axis=)
freq_map = np.zeros((rep1.n_residues, rep1.n_residues))
 (i, j), f  (pairs, freq):
    freq_map[i, j] = freq_map[j, i] = f
np.fill_diagonal(freq_map, )
fig, ax = plt.subplots(figsize=(, ))
im = ax.imshow(freq_map, vmin=, vmax=, origin=)
ax.(xlabel=, ylabel=, title=)
fig.colorbar(im, ax=ax, label=)
fig.tight_layout(); fig.savefig(outdir / ); plt.close(fig)


phi_ix, phi_rad = md.compute_phi(rep1)
psi_ix, psi_rad = md.compute_psi(rep1)
 ():  np.array([t.topology.atom(i[]).residue.index  i  ix])
phi_res, psi_res = res_of(phi_ix, rep1), res_of(psi_ix, rep1)
common = np.intersect1d(phi_res, psi_res)
phi_deg = np.degrees(phi_rad[:, np.isin(phi_res, common)]).ravel()
psi_deg = np.degrees(psi_rad[:, np.isin(psi_res, common)]).ravel()
H, _, _ = np.histogram2d(phi_deg, psi_deg, bins=, =[[-, ], [-, ]])
fig, ax = plt.subplots(figsize=(, ))
ax.imshow(H.T, origin=, extent=[-, , -, ], aspect=)
ax.(xlabel=, ylabel=, title=)
fig.tight_layout(); fig.savefig(outdir / ); plt.close(fig)


res_names = [rep1.topology.residue(r).name  r  common]
 label, cols, fname  [(, [i  i, n  (res_names)  n == ], ),
                           (, [i  i, n  (res_names)  n == ], )]:
      cols: 
    phi = np.degrees(phi_rad[:, cols]).ravel()
    psi = np.degrees(psi_rad[:, cols]).ravel()
    H, _, _ = np.histogram2d(phi, psi, bins=, =[[-, ], [-, ]])
    fig, ax = plt.subplots(figsize=(, ))
    ax.imshow(H.T, origin=, extent=[-, , -, ], aspect=)
    ax.(xlabel=, ylabel=, title=)
    fig.tight_layout(); fig.savefig(outdir / fname); plt.close(fig)


dssp = md.compute_dssp(rep1, simplified=); dssp[dssp == ] = 
codes_order = [, , , , , , , ]
code_to_int = {c: i  i, c  (codes_order)}
dssp_int = np.vectorize( c: code_to_int.get(c, ))(dssp)
fig, ax = plt.subplots(figsize=(, ))
ax.imshow(dssp_int.T, aspect=, origin=, interpolation=,
          extent=[rep1.time[] / , rep1.time[-] / , , rep1.n_residues])
ax.(xlabel=, ylabel=, title=)
fig.tight_layout(); fig.savefig(outdir / ); plt.close(fig)


np.savez(outdir / , dssp=dssp, dssp_int=dssp_int, codes_order=codes_order)
()

Workflow 2: H-Bond Persistence + SASA Across Replicates

Goal: persistent hydrogen bonds (≥50% of frames) and total SASA time series per replicate.

import mdtraj as md, numpy as np, pandas as pd

replicas = {"rep1": "rep1.xtc", "rep2": "rep2.xtc", "rep3": "rep3.xtc"}
top = "protein.pdb"
rows = []
for name, p in replicas.items():
    t = md.load(p, top=top)
    hbonds = md.baker_hubbard(t, freq=0.5, periodic=False)
    sasa_ang2 = (md.shrake_rupley(t, mode="residue").sum(axis=1) * 100.0)
    rows.append({"replica": name, "n_persistent_hbonds": len(hbonds),
                 "sasa_mean_ang2": sasa_ang2.mean(), "sasa_std_ang2": sasa_ang2.std()})
print(pd.DataFrame(rows).to_string(index=False))

Key Parameters

Parameter	Module	Default	Range / Options	Effect
chunk (md.iterload)	I/O	100	100-5000	Frames per stream chunk; trade speed vs RAM
atom_indices (md.rmsd)	Stability	None	array of ints	Restrict superposition + RMSD (e.g., CA only)
frame (md.rmsd)	Stability	0	0-n_frames-1	Reference frame
contacts (md.compute_contacts)	Contacts	"all"	"all", [(i,j),...]	All residue pairs or custom list
scheme (md.compute_contacts)	Contacts	"closest-heavy"	"ca", "closest", "closest-heavy", "sidechain", "sidechain-heavy"	Atom subset for inter-residue distance
cutoff (contact threshold, user-set)	Contacts	0.5 nm	0.4-0.8 nm	In-contact distance (5 Å for CA, 4 Å for closest-heavy)
simplified (md.compute_dssp)	DSSP	True	True/False	False → 8-state codes
freq (md.baker_hubbard)	H-bonds	0.1	0.0-1.0	Min frame fraction to report

Best Practices

1. Convert units before plotting — nm → Å (×10), ps → ns (÷1000), rad → deg (np.degrees). #1 cause of "my RMSD is 10× too small".
2. Superpose before RMSF — md.rmsd superposes internally; RMSF doesn't. Without explicit traj.superpose(...), RMSF picks up rigid-body motion.
3. md.iterload for big trajectories — 1 µs at 1 ps/frame is millions of frames. Streaming avoids OOM.
4. simplified=False for fine-grained DSSP — 3-state collapses 3_10 / pi-helix / bend transitions.
5. Pair indices, not resSeq — md.compute_contacts returns 0-based contiguous indices. Map to resSeq only for labeling.
6. Strip waters early — traj.atom_slice(traj.topology.select("protein")) cuts memory/time 5–20× on solvated systems.

Common Recipes

Recipe: Convert Trajectory Format

import mdtraj as md

traj = md.load("input.nc", top="topology.parm7")
traj.save_xtc("output.xtc")
traj.save_pdb("output.pdb")
print(f"Converted {traj.n_frames} frames -> output.xtc")

Recipe: Per-Residue Mean ± SD RMSF Across Replicates

Ensemble flexibility input for flexibility-aware docking or homology refinement.

import mdtraj as md, numpy as np, pandas as pd

paths = {"rep1": "rep1.xtc", "rep2": "rep2.xtc", "rep3": "rep3.xtc"}
top = "protein.pdb"
rmsf_by_rep = {}
for name, p in paths.items():
    t  = md.load(p, top=top)
    ct = t.atom_slice(t.topology.select("name CA")); ct.superpose(ct, frame=0)
    rmsf_by_rep[name] = np.sqrt(((ct.xyz - ct.xyz.mean(axis=0)) ** 2).sum(axis=2).mean(axis=0)) * 10.0

df = pd.DataFrame(rmsf_by_rep)
df["mean"], df["std"] = df.mean(axis=1), df.std(axis=1)
df.to_csv("rmsf_replicates.csv", index_label="residue_index")

Recipe: Save Frames Above RMSD Threshold

Curate "open" conformations for clustering or downstream analysis.

import mdtraj as md

traj = md.load("rep1.xtc", top="protein.pdb")
ca   = traj.topology.select("name CA")
traj.superpose(traj, frame=0, atom_indices=ca)
rmsd_ang = md.rmsd(traj, traj, frame=0, atom_indices=ca) * 10.0

open_state = traj[rmsd_ang > 4.0]
open_state.save_dcd("open_state.dcd")
print(f"Saved {open_state.n_frames}/{traj.n_frames} frames")

Recipe: Cross-Replicate Helix Fraction

import mdtraj as md, numpy as np

paths = ["rep1.xtc", "rep2.xtc", "rep3.xtc"]
helix_fracs = [(md.compute_dssp(md.load(p, top="protein.pdb"), simplified=False) == "H").mean() for p in paths]
print(f"α-helix per replicate: {[f'{h:.2f}' for h in helix_fracs]}")
print(f"Mean ± std: {np.mean(helix_fracs):.2f} ± {np.std(helix_fracs):.2f}")

Troubleshooting

Problem	Cause	Solution
RuntimeError: No topology	XTC/DCD/TRR carry no topology	Pass top="protein.pdb" (or .prmtop) to md.load
RMSD 10× too small	mdtraj returns nm	Multiply by 10.0
RMSF spikes at termini	Intrinsic terminal flexibility, not a bug	Slice termini off or report explicitly
DSSP returns only H/E/C	simplified=True default	Pass simplified=False
Empty/' ' DSSP cells	Space = coil, not missing	dssp[dssp == " "] = "C"
MemoryError on big trajectory	Whole trajectory in RAM	Switch to md.iterload(..., chunk=500)
phi/psi arrays differ in residue count	First residue has no phi; last has no psi	Intersect residue index sets before pairing
Ramachandran off-center	Angles still in radians	Wrap with np.degrees; set range=[[-180, 180], [-180, 180]]
md.compute_contacts returns NaN	Some pairs lack heavy atoms for closest-heavy (e.g., glycine sidechain)	Use scheme="closest" or "ca"; drop NaN columns
Contact map asymmetric	Only filled upper triangle	Mirror: freq_map[j, i] = freq_map[i, j]
traj.time all zeros	Some writers leave time blank	Reconstruct: traj.time = np.arange(traj.n_frames) * dt_ps

Related Skills

• mdanalysis-trajectory — richer atom-selection grammar and AnalysisBase framework
• autodock-vina-docking / smina-molecular-docking — feed representative MD frames as docking targets
• chembl-database-bioactivity — overlay ligand bioactivity on MD-derived flexibility maps

References

• mdtraj documentation — official docs, API reference, tutorials
• McGibbon et al. (2015) — "MDTraj: A Modern Open Library for the Analysis of Molecular Dynamics Trajectories", Biophys J
• DSSP algorithm — Kabsch & Sander (1983)
• Baker–Hubbard H-bond criterion — Baker & Hubbard (1984)
• Shrake–Rupley SASA — Shrake & Rupley (1973)
• mdtraj GitHub — source, issues, examples