Retention time calibration via lowess regression

retention-time-calibration-via-lowess-regression

Summary

A statistically guided approach to correct sample-level retention time drift in LC-MS metabolomics by identifying high-confidence landmark peaks and applying LOWESS (locally weighted scatterplot smoothing) regression with boundary extension. This ensures consistent peak alignment across samples before composite feature detection.

When to use

After mass track extraction and alignment across samples, when preparing to detect elution peaks on composite mass tracks. Use this when inter-sample retention time variation exceeds acceptable alignment tolerance (e.g., >10–15 seconds in typical LC-MS), which is detected by observing that the same mass track exhibits multiple distinct elution profiles across samples rather than a single coherent peak.

When NOT to use

• Input data are already RT-corrected or come from a single-sample analysis where cross-sample alignment is not needed.
• Landmark peaks with mSelectivity <0.99 or prominence <20% of peak height are predominant; calibration will be unstable.
• Retention time variation is <5 ppm or within expected instrument precision; calibration overhead outweighs benefit.

Inputs

• list of mass tracks per sample (output from chromatograms.extract_massTracks_)
• per-sample MS1 spectra intensity and retention time data
• mSelectivity and peak prominence metrics for each candidate landmark peak
• sample list and run metadata

Outputs

• rt_cal_dict: dictionary of retention time calibration parameters per sample
• RT-corrected mass tracks per sample (updated retention times for all peaks)
• calibration quality report (LOWESS curve fit, landmark peaks used, residual statistics)

How to apply

First, identify landmark peaks on each sample's mass tracks by filtering for selectivity (mSelectivity >0.99, indicating the peak is present in only one sample or has minimal ambiguity) and prominence (>20% of peak height). These landmarks serve as internal reference points with high confidence. Next, apply chromatograms.rt_lowess_calibration with 10% boundary extension to generate a per-sample retention time calibration dictionary (rt_cal_dict). The LOWESS approach locally estimates the relationship between observed and expected retention times across the set of landmark peaks, smoothing non-linear drift while preserving genuine peak shape variation. The boundary extension ensures that peaks near the start and end of the chromatogram are also corrected. Finally, apply this rt_cal_dict to all detected peaks in that sample before aggregating intensities across samples into composite mass tracks.

Related tools

• chromatograms.rt_lowess_calibration (Applies LOWESS-based retention time correction to a single sample using identified landmark peaks and 10% boundary extension) — https://github.com/shuzhao-li/asari
• peaks.quick_detect_unique_elution_peak (Identifies potential landmark peaks with high selectivity and prominence for use as calibration anchors) — https://github.com/shuzhao-li/asari
• scipy.signal (implied) (Underlying signal processing for LOWESS smoothing and peak prominence evaluation)
• asari workflow module (Orchestrates RT calibration step within the overall LC-MS processing pipeline) — https://github.com/shuzhao-li/asari

Examples

from asari.chromatograms import rt_lowess_calibration; rt_cal_dict = rt_lowess_calibration(sample_mass_tracks, landmark_peaks_mz_rt, boundary_extension=0.1)

Evaluation signals

• Landmark peak mSelectivity ≥0.99 and prominence ≥20% of peak height; check that sufficient landmarks (typically ≥5–10) were identified per sample.
• LOWESS curve fit residuals should show no systematic bias across the retention time range; plot observed vs. fitted RT to inspect for non-linearity.
• After RT calibration, the same mass track should display a single, coherent elution profile when visualized across all samples, with peak retention times clustering within 10–20 seconds.
• Composite map peak detection should recover the same features (m/z, RT) across replicates; compare preferred_Feature_table.tsv entries before and after calibration for consistent peak counts and alignment.
• Validate calibration on known standards (if available in the dataset); their retention times should shift by <1 minute and show consistent ordering across samples.

Limitations

• Requires sufficient high-quality landmark peaks (mSelectivity >0.99) in each sample; samples with poor peak quality or atypical chromatograms may have unstable calibration.
• LOWESS regression assumes a smooth, non-parametric relationship between observed and expected RT; highly non-linear or multi-modal drift patterns may not be fully corrected.
• Calibration is performed independently per sample; systematic instrument drift across the entire batch is not corrected (batch-level RT alignment is a separate step).
• 10% boundary extension may introduce artifacts if landmarks cluster near the chromatogram edges; sparse landmark distribution can lead to extrapolation errors.

Evidence

• [methods] identify landmark peaks with mSelectivity >0.99 and prominence >20% of peak height, then apply chromatograms.rt_lowess_calibration with 10% boundary extension to obtain per-sample rt_cal_dict: "identify landmark peaks with mSelectivity >0.99 and prominence >20% of peak height, then apply chromatograms.rt_lowess_calibration with 10% boundary extension to obtain per-sample rt_cal_dict"
• [methods] See chromatograms.rt_lowess_calibration: "See chromatograms.rt_lowess_calibration"
• [methods] Build composite mass tracks by summing aligned intensity values across all samples after RT calibration.: "Build composite mass tracks by summing aligned intensity values across all samples after RT calibration"