Mass Spectrometry Preprocessing

mass-spectrometry-preprocessing

Summary

Preparation and normalization of mass-spectrometry spectral data (m/z and intensity arrays) into a format compatible with machine learning input layers, specifically for CNN-based compound fingerprint prediction. This skill ensures spectral matrices are properly structured, intensity-normalized, and aligned for downstream model ingestion.

When to use

When you have raw mass-spectrometry data (precursor m/z, ionization mode, and fragment m/z–intensity pairs) that must be fed into a CNN model for metabolite annotation via compound fingerprint prediction. Apply this skill before training or inference with a neural network expecting fixed-dimension spectral tensors.

When NOT to use

• Input is already a pre-trained CNN model checkpoint (.h5 weights file); use model loading instead.
• Spectral data is already in a neural network–ready format (e.g., pre-binned NumPy arrays); skip to model inference.
• Analysis goal is exploratory (e.g., library matching or cosine similarity ranking) rather than CNN fingerprint prediction; consider traditional spectral similarity metrics instead.

Inputs

• Raw mass-spectrometry spectral data file (txt format with precursor m/z, ionization mode, and m/z–intensity pairs)
• List of candidate InChIKeys for validation/ranking
• Specification of target CNN input tensor shape and intensity normalization scheme

Outputs

• Preprocessed spectral matrices (fixed-dimension arrays compatible with CNN input layer)
• Normalized intensity tensors (range-scaled to model training specification)
• Matched spectrum–candidate metadata mapping for downstream scoring

How to apply

Load mass-spectrometry input data specifying precursor mass, ionization mode (positive/negative), and intensity pairs (m/z and normalized intensity values). Construct spectral matrices by binning or aligning fragment m/z values into fixed-dimension arrays, normalizing intensities (e.g., base peak = 100) to a consistent scale. Ensure the resulting tensor shape matches the CNN input layer dimensions (e.g., 1D or 2D convolution kernel expectations). Validate that all spectra in a batch have identical dimensions and that intensity ranges fall within [0, 1] or [0, 100] as the model was trained on. Save preprocessed data in a format (e.g., NumPy arrays or HDF5) that the TensorFlow/PyTorch training pipeline can directly consume.

Related tools

• TensorFlow (Framework for defining CNN input layer shape and loading preprocessed spectral tensors during training and inference) — https://www.tensorflow.org/
• PyTorch (Alternative deep learning framework for constructing and training CNN models on preprocessed spectral data)
• Open Babel (Converts chemical structures (InChIKey) to molecular fingerprints for comparison against CNN predictions) — https://openbabel.org/wiki/Python
• PyFingerprint (Generates molecular fingerprints from compounds for ground-truth labels during CNN training) — https://github.com/hcji/PyFingerprint

Examples

python3 main.py

Evaluation signals

• All preprocessed spectra have identical tensor dimensions matching the CNN input layer specification (e.g., shape [1, sequence_length] for 1D convolution).
• Intensity values are normalized to the expected range (0–1 or 0–100) with no NaN or infinite values.
• Precursor m/z and ionization mode are correctly paired with their corresponding fragment intensity arrays.
• Batch-wise preprocessing is reproducible: running the same spectrum through preprocessing yields bitwise-identical output tensors.
• CNN model successfully ingests the preprocessed data without shape mismatch errors or type casting warnings.

Limitations

• Spectral matrices must be padded or trimmed to a fixed dimension; spectra with very few fragments or many overlapping peaks may lose information or require aggressive quantization.
• Intensity normalization scheme (e.g., base peak = 100) assumes a consistent ionization and detection response; spectra from different instruments or modes may require separate normalization.
• The method does not validate chemical plausibility of the candidate InChIKey list; incorrect or out-of-distribution compounds will still yield CNN scores.
• No mechanism is provided to handle missing or corrupt m/z–intensity pairs; malformed input rows will cause parsing errors.

Evidence

• [other] Load or construct mass-spectrometry input data (m/z and intensity arrays or spectral matrices) in a format compatible with the CNN input layer.: "Load or construct mass-spectrometry input data (m/z and intensity arrays or spectral matrices) in a format compatible with the CNN input layer."
• [readme] The first row represents the precursor mass and ionization mode, followed by intensity pairs.: "The first row represents the precursor mass and ionization mode, followed by intensity pairs."
• [other] Train the CNN model using a suitable loss function (e.g., binary cross-entropy for fingerprint bits or Tanimoto-based loss) and optimizer, monitoring validation performance.: "Train the CNN model using a suitable loss function (e.g., binary cross-entropy for fingerprint bits or Tanimoto-based loss) and optimizer, monitoring validation performance."
• [readme] The second column represents the Tanimoto similarity score. Each table will be ranked in a descending order by score.: "The second column represents the Tanimoto similarity score. Each table will be ranked in a descending order by score."