Siamese Network Inference Spectrum Pairs

Siamese Network Inference on Spectrum Pairs

Summary

Apply a trained Siamese neural network to compute predicted structural similarity scores (Tanimoto or Dice) between pairs of preprocessed tandem mass spectra by computing cosine similarity between paired 200-dimensional embeddings. This skill enables rapid prediction of molecular structural resemblance directly from MS/MS data without explicit fingerprint computation.

When to use

You have a collection of preprocessed tandem mass spectra (binned into 10,000 equally-sized m/z bins, intensities square-root transformed, top 1,000 peaks retained), a trained MS2DeepScore Siamese model, and you need to predict structural similarity scores (Tanimoto or Dice) for all or a subset of unique spectrum pairs. Use this skill when batch inference on thousands of spectrum pairs is required and you want predictions to be fast and scalable.

When NOT to use

• Spectra have not been preprocessed (not binned into 10,000 m/z bins, intensities not square-root transformed, or more than 1,000 peaks retained) — preprocessing is a prerequisite.
• You need to train a new model on your own dataset rather than apply an existing trained model — use the training workflow instead.
• Input spectra are from ionization modes or m/z ranges substantially different from the model's training distribution (e.g., negative mode if model trained exclusively on positive mode), as prediction accuracy may degrade outside the model's learned chemical space.

Inputs

• Preprocessed tandem mass spectra (10,000-bin vectors with square-root-transformed intensities, m/z range 10–1000)
• Trained Siamese neural network model weights and architecture
• Spectrum pair indices (unique pairs for which similarity scores are desired)
• Ground-truth structural similarity labels (optional, for validation): Tanimoto or Dice scores computed from RDKit Daylight fingerprints

Outputs

• Predicted structural similarity scores (Tanimoto or Dice) for all spectrum pairs
• 200-dimensional spectral embeddings for each unique spectrum
• Prediction uncertainty estimates (IQR) when Monte-Carlo Dropout is applied
• Filtered similarity score matrix (optionally restricted to low-uncertainty predictions)
• Root mean squared error (RMSE) metric comparing predictions to ground-truth labels

How to apply

Load the pre-trained Siamese network model (with base network architecture: 200-dimensional embedding layer, 2×500-node hidden layers, L1/L2 regularization, batch normalization, dropout rate 0.2) and pass preprocessed spectrum pairs through the model to compute 200-dimensional embeddings for each spectrum in the pair. Compute cosine similarity between paired embeddings to yield predicted structural similarity scores. For uncertainty quantification, optionally apply Monte-Carlo Dropout by running N=10 forward passes with dropout enabled on all but the first layer, compute the interquartile range (IQR) of predictions, and filter predictions to retain only pairs with IQR below a threshold (e.g., IQR < 0.025 to remove high-uncertainty predictions). Calculate root mean squared error (RMSE) between predicted and ground-truth scores on the full or filtered subset.

Related tools

• MS2DeepScore (Provides the trained Siamese network model and inference engine for computing spectral embeddings and similarity scores) — https://github.com/matchms/ms2deepscore
• matchms (Provides spectrum data structures, preprocessing utilities, and pipeline infrastructure for loading and filtering spectra) — https://github.com/matchms/matchms
• RDKit (Used to compute ground-truth Tanimoto scores from Daylight fingerprints for model validation)
• Python (Runtime environment for executing Siamese network inference and Monte-Carlo Dropout uncertainty quantification)

Examples

from ms2deepscore.models import load_model
from ms2deepscore import MS2DeepScore
model = load_model('ms2deepscore_model.pt')
ms2ds = MS2DeepScore(model)
similarity_scores = ms2ds.pair(spectrum_1, spectrum_2)

Evaluation signals

• RMSE between predicted and ground-truth Tanimoto scores falls within expected range (~0.15 without filtering, ~0.10 with IQR < 0.025 filtering) on a held-out test set
• Predicted similarity scores lie within the valid range [0, 1] for Tanimoto/Dice scores; no NaN or infinite values
• Cosine similarity computations are numerically stable; embedding vectors have expected L2 norms and dimensionality (200)
• Monte-Carlo Dropout IQR estimates show decreasing variance with increasing model confidence; low-IQR predictions cluster away from boundary scores (0 and 1)
• Spectrum pair ordering and batch processing produce identical results regardless of execution order (reproducibility check)

Limitations

• Model accuracy (RMSE ~0.13–0.20) is limited to predicting Tanimoto scores between 0.1 and 0.9; predictions for very low or very high structural similarity may be less reliable.
• The model was trained on GNPS spectra (109,734 spectra, 15,062 unique molecules); generalization to spectra from other sources (e.g., proprietary databases, novel compound classes) is not guaranteed.
• Inference requires spectra to be preprocessed identically to training (10,000-bin vectors, square-root-transformed intensities, top 1,000 peaks); deviations in preprocessing can degrade prediction accuracy.
• Monte-Carlo Dropout uncertainty estimates (N=10 forward passes) may be unreliable for rare or out-of-distribution spectrum types; computational cost scales linearly with the number of forward passes.
• The skill does not detect or flag failure modes (e.g., spectra with missing metadata, anomalous m/z distributions, or ionization mode mismatches that could cause silently degraded predictions).

Evidence

• [other] Perform Siamese network inference on all unique spectrum pairs (n=6,485,401), computing cosine similarity between paired 200-dimensional embeddings to yield predicted Tanimoto scores.: "Perform Siamese network inference on all unique spectrum pairs (n=6,485,401), computing cosine similarity between paired 200-dimensional embeddings to yield predicted Tanimoto scores."
• [other] MS2DeepScore predicts Tanimoto scores on the 3,601-spectrum test set with root mean squared error of approximately 0.15 without uncertainty filtering and 0.1 when applying interquartile range (IQR < 0.025) thresholds to remove high-uncertainty predictions.: "MS2DeepScore predicts Tanimoto scores on the 3,601-spectrum test set with root mean squared error of approximately 0.15 without uncertainty filtering and 0.1 when applying interquartile range (IQR <"
• [methods] a base network (Fig. 1) which generates 200-dimensional embeddings for each spectrum: "a base network (Fig. 1) which generates 200-dimensional embeddings for each spectrum"
• [methods] To estimate the uncertainty of a prediction we used Monte-Carlo Dropout ensembles [17]. At inference time, dropout was applied to all but the first layer of the base network. N = 10 embeddings were: "To estimate the uncertainty of a prediction we used Monte-Carlo Dropout ensembles. At inference time, dropout was applied to all but the first layer of the base network. N = 10 embeddings were"
• [methods] Spectrum peaks were binned in 10,000 equally-sized bins ranging from 10 to 1000 m/z: "Spectrum peaks were binned in 10,000 equally-sized bins ranging from 10 to 1000 m/z"
• [methods] Peak intensities were square root transformed to avoid a too strong focus on the highest intensity peaks only: "Peak intensities were square root transformed to avoid a too strong focus on the highest intensity peaks only"
• [readme] To compute the similarities between spectra of your choice you can run the code below.: "To compute the similarities between spectra of your choice you can run the code below."
• [results] MS2DeepScore generally performs very well and can predict Tanimoto scores between 0.1 and 0.9 with a RMSE between 0.13 and 0.2: "MS2DeepScore generally performs very well and can predict Tanimoto scores between 0.1 and 0.9 with a RMSE between 0.13 and 0.2"