Peak center coordinate localization

peak-center-coordinate-localization

License: noncommercial — confirm your use is a permitted (noncommercial) purpose before applying; commercial use requires a separate license (see metadata.tool_license). <!-- asb-license-banner -->

Summary

Train and apply a CNN model to predict the precise rt×mz center coordinates of chromatographic peaks detected in LC-HRMS profile mode data. This skill enables automated, coordinate-level localization of peak maxima from standardized two-dimensional rt×mz areas extracted around local maxima.

When to use

You have LC-HRMS profile mode data with candidate peak regions (local maxima) exported as standardized rt×mz two-dimensional areas, and you need to automatically predict the precise peak-center location (both rt and mz coordinates) rather than relying on manual inspection or simple gradient-descent algorithms. Use this when peak-center accuracy is critical for downstream quantification, alignment across samples, or feature table construction.

When NOT to use

• Input data are already centroided (not profile mode); centroided peaks have fundamentally different structure and do not benefit from pixel-level coordinate regression.
• You only need binary peak-vs.-background classification and do not need precise center location; simpler classifiers would be more efficient.
• Reference chromatograms or ground-truth peak lists are unavailable or too small to generate sufficient training instances; the CNN requires large augmented training datasets to achieve accurate coordinate prediction.

Inputs

• LC-HRMS profile mode chromatogram data (raw or NetCDF format)
• Standardized two-dimensional rt×mz area images around local maxima
• User-defined reference list of ground-truth isolated chromatographic peaks
• Training chromatograms for reference feature extraction

Outputs

• Predicted peak-center rt coordinate (retention time)
• Predicted peak-center mz coordinate (mass-to-charge ratio)
• Peak-center coordinate prediction error metrics
• Trained CNN model weights (TensorFlow format)

How to apply

First, prepare a training dataset by extracting reference chromatographic peaks from ground-truth chromatograms using smoothing and gradient-descent peak-finding, then match these peaks to a user-defined reference list and iteratively combine matched references to generate a large number of augmented training instances (including distraction peaks and different background types). Train a CNN model with an auxiliary regression head using TensorFlow on these instances, with GPU acceleration (CUDA) to reduce training time. The model learns to output peak-center coordinates (rt and mz) alongside peak-type classification. On new LC-HRMS data, apply the trained model to standardized rt×mz areas around detected local maxima to predict their center coordinates; validate predictions by checking coordinate prediction error on held-out test sets and visually inspecting exported detection examples.

Related tools

• PeakBot (Python package that implements the complete CNN-based peak-center localization workflow, including reference feature extraction, training instance generation with GPU acceleration, CNN model training in TensorFlow, and inference on new LC-HRMS data) — https://github.com/christophuv/PeakBot
• TensorFlow (Deep learning framework used to implement and train the CNN model with regression head for peak-center coordinate prediction) — https://www.tensorflow.org/
• OpenMS / TOPPView (Optional visualization and export tool; detected peaks with predicted center coordinates can be exported as featureML files for inspection in TOPPView) — https://pubmed.ncbi.nlm.nih.gov/19425593/

Examples

python quickExample_GPU.py

Evaluation signals

• Peak-center coordinate prediction error (difference between predicted rt, mz and ground-truth center) on held-out test set is below acceptable threshold (article does not specify threshold; use domain expectations, e.g., <0.5% rt error, <5 ppm mz error).
• Predicted center coordinates lie within the bounding-box output by the model; coordinate should not fall outside suggested peak boundaries.
• Visual inspection of exported detection examples confirms that predicted centers align with the visual center of chromatographic peak regions in rt×mz space.
• Coordinate predictions are consistent across replicate samples of the same compound; peak-center rt and mz should be stable within known instrumental/biological variation.
• Downstream quantification or feature alignment using predicted center coordinates shows improved accuracy or reduced error compared to gradient-descent-only center estimates.

Limitations

• GPU with CUDA support and sufficient memory (≥4 GB for default export batch size of 2048) is required for efficient training; CPU-only execution is possible but significantly slower. Block dimension (blockdim) and grid dimension (griddim) parameters must be tuned per GPU architecture.
• Coordinate prediction accuracy depends heavily on the size and representativeness of the training dataset; small or biased reference peak sets will lead to poor generalization on new samples.
• The model trains on augmented instances (synthetic combinations of reference peaks with distraction peaks and background); if the training distribution does not match the real data distribution (e.g., novel background types, unusual peak shapes), prediction error may increase.
• No changelog is provided in the repository, making it difficult to assess version differences or known bugs in coordinate prediction across releases.

Evidence

• [readme] uses local-maxima in the LC-HRMS dataset each of which is then exported as a standarized two-dimensional area (rt x mz), which is used as the input for a machine-learning CNN model: "uses local-maxima in the LC-HRMS dataset each of which is then exported as a standarized two-dimensional area (rt x mz), which is used as the input for a machine-learning CNN model"
• [readme] reports whether the local-maxima is a chromatographic peak with left/right isomeric compounds or a signal of the background. Moreover, for chromatographic peaks it suggests a bounding-box and a peak-center: "for chromatographic peaks it suggests a bounding-box and a peak-center"
• [readme] generate a large number of training instances by iteratively combining them. Each such training instance can consist of a chromatographic peak or background signal and several other distraction peaks to generalize: "generate a large number of training instances by iteratively combining them. Each such training instance can consist of a chromatographic peak or background signal and several other distraction peaks"
• [readme] A GPU (CUDA) based approach is implemented that decreases the time required for their generation. The CNN model is implemented in the TensorFlow package: "A GPU (CUDA) based approach is implemented that decreases the time required for their generation. The CNN model is implemented in the TensorFlow package"
• [readme] It consists of several convolutional and pooling-layers and outputs a peak-type, -center, and -bounding-box.: "It consists of several convolutional and pooling-layers and outputs a peak-type, -center, and -bounding-box"
• [readme] If an exportBatchSize of 2048 requires some 4GB of GPU-memory. If you have less, try reducing this value to 1024 of 512.: "If an exportBatchSize of 2048 requires some 4GB of GPU-memory. If you have less, try reducing this value to 1024 of 512"
• [readme] the properties of the reference features are also updated to best fit the chromatographic peaks from the reference chromatograms: "the properties of the reference features are also updated to best fit the chromatographic peaks from the reference chromatograms"