tdfpy 4.0.1

A Python package for extracting data from Bruker timsTOF data files (.tdf and .tdf_bin). Includes a Numba-accelerated centroiding algorithm for efficient extraction of ion mobility data.

Overview

tdfpy provides a high-level Python API for reading Bruker timsTOF .d folders. It handles DDA, DIA, and PRM acquisition modes and exposes familiar Python objects — no need to think about raw PASEF frames or SQLite queries.

• DDA — iterate MS1 frames and precursors (MS2 spectra)
• DIA — iterate MS1 frames and DIA isolation windows
• PRM — iterate targets and their transitions
• Composable peak pipeline — read_spectrum → optional region exclusion / smoothing / noise-filter chain → centroider. Returns (N, 3) [m/z, intensity, 1/K0] arrays
• Two centroiders — MergePeaksCentroider (default, Numba-JIT'd greedy merge in float m/z) and WatershedCentroider (intensity-ordered region growing in integer TOF-index space)
• Composable noise filters — chain MadThreshold, VerticalNoiseFilter, HorizontalHaloFilter, and others via noise=[…]. String shorthand (noise="mad") preserved for terseness
• Lazy spectral access — frame metadata is loaded upfront; raw peak data is only read when you call .peaks, .raw_peaks(), or .centroid()

Installation

pip install tdfpy

Requires Python 3.12+. Pure Python — analysis.tdf_bin is decoded directly, so there is no Bruker native library and no platform restriction (Linux, macOS and Windows, x86-64 and ARM). On Python 3.12/3.13 the zstandard package is installed automatically; Python 3.14+ uses the standard library's zstd module.

Quick Start

from tdfpy import DDA, DIA, PRM

# DDA acquisition
with DDA("sample.d") as dda:
    for frame in dda.ms1:
        peaks = frame.centroid()  # shape (N, 3): [m/z, intensity, 1/K0]

    for precursor in dda.precursors:
        print(precursor.largest_peak_mz, precursor.charge)
        peaks = precursor.peaks  # MS2 centroided by tdfpy (mobility collapse + merge)

# DIA acquisition
with DIA("sample.d") as dia:
    for frame in dia.ms1:
        peaks = frame.centroid()

    for window in dia.windows:
        print(window.isolation_mz, window.isolation_width)
        peaks = window.centroid()

# PRM acquisition
with PRM("sample.d") as prm:
    for target in prm.targets:
        print(target.monoisotopic_mz, target.charge)

    for transition in prm.transitions:
        print(transition.isolation_mz, transition.collision_energy)
        peaks = transition.peaks  # list of raw per-scan (N, 2) arrays

Frame-batched window processing and acquisition validation are described in the analysis guide. An optional MCP server gives AI agents tools for inspecting acquisitions and extracting spectra. Use tdfpy validate sample.d --full to check all binary frames without modifying an acquisition.

Lookups and Queries

Frames, precursors, and windows can be accessed by ID or queried by m/z and retention time:

with DDA("sample.d") as dda:
    frame = dda.ms1[1]           # by frame ID
    precursor = dda.precursors[1]  # by precursor ID

    # query by m/z and RT window
    hits = dda.precursors.query(
        mz=1292.63,
        mz_tolerance=20.0,   # ppm
        rt=2400.0,           # seconds
        rt_tolerance=30.0,
    )

Peak extraction

Every frame-element method (Frame.raw_peaks(), Frame.centroid(), and the matching methods on DiaWindow and PrmTransition) accepts the same composable arguments:

from tdfpy import (
    ChargeStateRegion, Smooth, MadThreshold, VerticalNoiseFilter,
    HorizontalHaloFilter, MergePeaksCentroider, WatershedCentroider,
)

peaks = frame.centroid(
    # Region exclusion: drop the singly-charged contamination band
    exclude=ChargeStateRegion(),

    # Intensity smoothing: box-sum to amplify ion-mobility streaks pre-filter
    smooth=Smooth(scan_half_width=5, mz_idx_half_width=2),

    # Noise filter pipeline (applied in order, pre-centroid)
    noise=[
        VerticalNoiseFilter(min_streak_scans=5, num_iterations=2),
        HorizontalHaloFilter(),  # clear the left/right m/z halo
        MadThreshold(k=3),
    ],

    # Centroider — swap algorithms without changing the surrounding code
    centroid=MergePeaksCentroider(mz_tolerance=8, mz_tolerance_type="ppm", min_peaks=3),
    # or:  WatershedCentroider(attach_scan_half_width=10, attach_mz_idx_half_width=3)
)

Terser shorthand still works for common cases:

peaks = frame.centroid()                       # all defaults
peaks = frame.centroid(noise="mad")            # MAD-based intensity floor
peaks = frame.centroid(noise=500.0)            # absolute threshold

Watershed centroider

The watershed centroider works in integer (scan, TOF-index) space — avoiding the float-m/z binning that the greedy merger does. Useful when peaks are closely spaced or when seed selection needs to be robust to noisy spikes:

peaks = frame.centroid(
    centroid=WatershedCentroider(
        attach_scan_half_width=10, attach_mz_idx_half_width=3,
        min_seed_intensity=50.0,
        # Position-preserving intensity smoothing before seed selection (on by default)
        smooth_scan_half_width=5, smooth_mz_idx_half_width=3,
    ),
)

Custom pipelines

For ordering or transformations beyond what the convenience methods cover, call the pipeline ops directly:

from tdfpy import (
    read_spectrum, exclude_region, smooth, apply_noise, convert,
    ChargeStateRegion, VerticalNoiseFilter, HorizontalHaloFilter,
)

s = read_spectrum(td, frame_id)
s = exclude_region(s, ChargeStateRegion(), td=td, frame_id=frame_id)
s = smooth(s, scan_half_width=5, mz_idx_half_width=2)   # box-sum, amplify IM streaks
s = apply_noise(s, (VerticalNoiseFilter(), HorizontalHaloFilter()), td=td, frame_id=frame_id)
peaks = convert(s, td, frame_id)

Noise filtering vs min_peaks

Intensity-based noise filters (MadThreshold, PercentileThreshold, etc.) can be too aggressive: they cannot distinguish low-abundance real signal from electronic noise. Raising min_peaks on the centroider is often a more reliable filter, since electronic noise typically manifests as singletons while real peaks appear across multiple scans. The structural VerticalNoiseFilter extends this idea to the IM axis.

Documentation

Full documentation at tacular-omics.github.io/tdfpy