Neural Network Inference And Embedding Extraction

neural-network-inference-and-embedding-extraction

Summary

Extract fixed-dimensional vector embeddings and compute pairwise structural similarity predictions from tandem mass spectra using a trained Siamese neural network. This skill produces both continuous similarity scores (Tanimoto predictions) and intermediate spectrum representations suitable for dimensionality reduction and clustering.

When to use

When you have paired or unpaired MS/MS spectra and need to compute structural similarity scores without explicit molecular fingerprint computation, or when you want to generate low-dimensional embeddings for spectral visualization, clustering, or retrieval tasks. Use this when you have a trained Siamese model and want to move from raw spectra to either similarity matrices or embedding arrays.

When NOT to use

• Spectra have not been cleaned and preprocessed (no peak filtering, no binning, no intensity normalization applied yet).
• You need interpretable molecular fingerprints or explicit structural annotations; MS2DeepScore predicts similarity without generating fingerprints.
• Spectra are from ionization modes or compound classes not represented in the training data (model trained on >500,000 spectra from GNPS, Mona, MassBank, MSnLib in positive/negative modes); cross-ionization predictions available in MS2DeepScore 2.0+ but not in the original 2021 model.

Inputs

• Cleaned tandem mass spectra (MS/MS) in matchms Spectrum objects or formats supporting msp, mzml, mgf, mzxml, json, usi
• Trained Siamese neural network model (PyTorch .pt file)
• Spectrum pairs (tuples of two Spectrum objects) or individual Spectrum objects

Outputs

• Tanimoto similarity score predictions (continuous values 0–1) for spectrum pairs
• Embedding array: 2D numpy array where each row is a fixed-dimensional vector representation of a spectrum
• Uncertainty estimates (IQR values) for each prediction when using Monte-Carlo Dropout

How to apply

Load a trained MS2DeepScore Siamese neural network model and pass cleaned, binned spectrum pairs (or individual spectra for embedding extraction) through the base network. For similarity scoring: compute cosine similarity between embeddings of spectrum pairs to produce Tanimoto score predictions (range 0–1). For embedding extraction: call the embedding layer directly to obtain fixed-dimensional vectors for each spectrum, optionally applying Monte-Carlo Dropout at inference time to estimate prediction uncertainty via interquartile range (IQR) filtering. Spectrum preprocessing must include peak intensity normalization (square root transform), removal of low-intensity peaks (<0.1% of max), and binning into 10,000 equally-sized bins across the 10–1000 m/z range. Filter predictions by IQR threshold if uncertainty quantification is required for higher accuracy.

Related tools

• MS2DeepScore (Siamese neural network model for predicting structural similarity scores and generating spectrum embeddings from MS/MS spectra) — https://github.com/matchms/ms2deepscore
• matchms (Spectrum data structure and pipeline management; handles spectrum loading, cleaning, and filtering) — https://github.com/matchms/matchms
• PyTorch (Deep learning framework for loading trained model and executing forward passes during inference)
• scikit-learn (t-SNE dimensionality reduction for visualizing embeddings in 2D chemical space)

Examples

from ms2deepscore.models import load_model
from ms2deepscore import MS2DeepScore
model = load_model('ms2deepscore_model.pt')
ms2ds = MS2DeepScore(model)
embeddings = ms2ds.get_embedding_array(cleaned_spectra)
similarity_scores = ms2ds.pair_predict(spectrum_pairs)

Evaluation signals

• Similarity scores lie in valid range [0, 1] representing predicted Tanimoto overlap; check for NaN or out-of-range values.
• Embedding array dimensions match expected size: number of spectra × model embedding dimension (e.g., if model outputs 256-dim vectors, array shape is N × 256).
• Root mean squared error (RMSE) between predicted Tanimoto scores and ground-truth Tanimoto scores (computed on molecular fingerprints) is ~0.15 without uncertainty filtering, improving to ~0.10 with IQR-based filtering; check reproducibility against published Figure 4.
• Precision-recall curves for high structural similarity pairs (Tanimoto > 0.6) show MS2DeepScore outperforming modified Cosine and Spec2Vec baselines across full threshold range (0–1.0).
• Embeddings from the same compound cluster together in dimensionality-reduced visualization (UMAP/t-SNE) with visually distinct separation from unrelated compounds.

Limitations

• Spectrum metadata (parent mass, elemental formula, adduct information) are not used in the Siamese network training, so predictions rely only on fragment peaks.
• Model was trained on spectra binned to 10,000 m/z bins (10–1000 range); spectra outside this range or with different binning strategies may degrade performance.
• Uncertainty quantification via Monte-Carlo Dropout sampling and IQR filtering adds computational overhead; trade-off between accuracy gain and inference time must be evaluated per use case.
• Cross-ionization mode predictions (negative to positive or vice versa) are not supported in the original 2021 MS2DeepScore model; available only in MS2DeepScore 2.0+.
• Training on >500,000 spectra with strong bias toward common lipids, metabolites, and natural products; performance on rare compound classes or non-standard ionization conditions is not characterized.

Evidence

• [intro] MS2DeepScore, a deep learning approach that is trained to predict structural similarities (Tanimoto or Dice scores based on molecular fingerprints) directly from pairs of MS/MS spectra: "predict structural similarities (Tanimoto or Dice scores based on molecular fingerprints) directly from pairs of MS/MS spectra without first"
• [intro] achieve a root mean squared error for predicted Tanimoto scores of about 0.15 when run without uncertainty restrictions, and down to 0.1 with stronger restrictions on model uncertainty: "achieve a root mean squared error for predicted Tanimoto scores of about 0.15 when run without uncertainty restrictions, and down to 0.1 with stronger restrictions on model uncertainty"
• [intro] can create mass spectral embeddings that can be used for additional spectral clustering: "can create mass spectral embeddings that can be used for additional spectral clustering"
• [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"
• [methods] 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: "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"
• [results] filtered out scores, according to increasingly stringent interquartile range (IQR) thresholds: "filtered out scores, according to increasingly stringent interquartile range (IQR) thresholds"
• [full_text] MS2DeepScore outperforms both modified Cosine and Spec2Vec across the full precision-recall range for identifying structurally related compounds: "MS2DeepScore outperforms both modified Cosine and Spec2Vec across the full precision-recall range for identifying structurally related compounds"
• [readme] To calculate chemical similarity scores, MS2DeepScore first calculates an embedding (vector) representing each spectrum. This intermediate product can also be used to visualize spectra in 'chemical space' by using a dimensionality reduction technique, like UMAP.: "To calculate chemical similarity scores, MS2DeepScore first calculates an embedding (vector) representing each spectrum. This intermediate product can also be used to visualize spectra in 'chemical"
• [readme] The model works for spectra in both positive and negative ionization modes and even predictions across ionization modes can be made by this model.: "The model works for spectra in both positive and negative ionization modes and even predictions across ionization modes can be made by this model."