Skip to content

Retention Time Deviation

The DeepLC feature (name: DeepLC) predicts peptide retention times using the DeepLC deep learning model and computes the deviation between predicted and observed retention times. Correct peptide identifications should show smaller deviations, while incorrect matches tend to have larger discrepancies.

Source name: DeepLC

Output Columns

Column Description
observed_retention_time Retention time recorded in the input data (seconds)
predicted_retention_time Retention time predicted by DeepLC
retention_time_diff Signed difference: predicted minus observed
abs_retention_time_diff Absolute value of the retention time difference
retention_time_ratio Ratio of the smaller to the larger of predicted and observed

Computation

Step 1: Preprocessing

Peptide sequences are preprocessed for DeepLC input:

  1. Flanking amino acids are stripped (e.g., K.PEPTIDE.R → PEPTIDE).
  2. Modifications are converted from mass-annotated format to UNIMOD notation using the modificationMap. For example, M[147.035] is converted to a UNIMOD-indexed modification string that DeepLC understands.

Step 2: Calibration

DeepLC benefits from calibration on high-confidence PSMs from the same LC-MS run. The calibration procedure:

  1. Sort all PSMs by the calibrationCriteria column (a search engine score). If lowerIsBetter is true, sort ascending; otherwise, sort descending.
  2. Select the top PSMs as the calibration set:
    • If calibrationSize is a float (e.g., 0.1), take the top 10% of PSMs.
    • If calibrationSize is an integer (e.g., 500), take the top 500 PSMs.
  3. Filter to target PSMs only (decoys are excluded from calibration).
  4. Calibrate the DeepLC predictor using the observed retention times of the calibration set.

Step 3: Prediction

The calibrated DeepLC model predicts retention times for all PSMs in the dataset.

Step 4: Feature Computation

Let \( t_\mathrm{obs} \) and \( t_\mathrm{pred} \) denote the observed and predicted retention times, respectively.

The signed difference \( \Delta t \), its absolute value \( |\Delta t| \), and the retention time ratio \( R_t \) are:

\[ \Delta t = t_\mathrm{pred} - t_\mathrm{obs} \]
\[ |\Delta t| = |t_\mathrm{pred} - t_\mathrm{obs}| \]
\[ R_t = \frac{\min(t_\mathrm{pred},\; t_\mathrm{obs})}{\max(t_\mathrm{pred},\; t_\mathrm{obs})} \]

\( R_t \) is bounded in \((0, 1]\) and equals 1 when the predicted and observed times are identical.

Missing Values

Any PSMs for which DeepLC cannot produce a prediction are filled with the median of each feature column.

Configuration

featureGenerator:
  - name: DeepLC
    params:
      calibrationCriteria: expect    # Column name used for selecting calibration PSMs
      lowerIsBetter: true            # Whether lower values of the criteria are better
      calibrationSize: 0.1           # Fraction (float) or count (int) of top PSMs for calibration
Parameter Default Description
calibrationCriteria (required) Name of a search engine score column to rank PSMs for calibration
lowerIsBetter false Set to true if lower values of the calibration criteria indicate better PSMs (e.g., E-values)
calibrationSize 0.15 Fraction of PSMs (float) or absolute count (int) for the calibration set