Reference Database Screening

reference-database-screening

Summary

Screen predicted molecular embeddings against a reference database to identify and rank candidate molecules by similarity. This skill bridges CNN predictions and molecular identification by computing pairwise similarity scores and filtering high-confidence matches.

When to use

After a CNN model has generated predicted molecular embeddings from mass spectrometry data, use this skill to match those predictions against a curated reference database of known molecular embeddings. Specifically, apply it when you need to convert abstract embeddings into a ranked list of candidate molecules with confidence scores for metabolite identification or compound annotation tasks.

When NOT to use

• Query or reference embeddings are not from the same trained CNN model or are incompatible dimensionally
• Reference database is empty, unindexed, or embeddings are not pre-computed
• Input embeddings are already discrete molecular annotations rather than continuous vectors

Inputs

• CNN-predicted embeddings for query molecules (vector format from trained model)
• Reference database embeddings (pre-computed vectors from reference molecules)
• Similarity threshold parameter (float, e.g., cosine similarity cutoff)
• Top N candidates parameter (integer, e.g., 5)

Outputs

• Ranked candidate list with match scores and reference molecule identifiers
• Pairwise similarity score matrix (queries × reference database)
• Filtered candidate matches exceeding the similarity threshold

How to apply

Load the CNN-predicted embeddings for query molecules and reference database embeddings from the trained model output. Compute pairwise similarity scores between query embeddings and all reference database embeddings using a distance metric such as cosine similarity or Euclidean distance. Rank reference molecules by similarity score in descending order for each query compound. Apply a similarity threshold to retain only high-confidence matches and filter candidate matches accordingly. Generate a ranked candidate list with match scores and reference molecule identifiers for each query compound, optionally including the top N candidates as configured.

Related tools

• Convolutional Neural Network (CNN) (Generates predicted molecular embeddings from mass spectrometry spectra data that serve as query vectors for database screening)
• ChemEmbed (End-to-end framework that orchestrates data preprocessing, CNN prediction, and reference database screening for metabolite identification) — https://github.com/massspecdl/ChemEmbed

Examples

python main.py --config config.yaml
# where config.yaml specifies reference_database, model_path, top_n_candidates: 5, and prediction_results output file

Evaluation signals

• Similarity scores are within the valid range [0, 1] for cosine similarity or are non-negative for Euclidean distance
• Ranked candidate list is non-empty when similarity threshold is appropriate; empty results may indicate threshold is too stringent or embeddings are poorly calibrated
• Top-ranked candidate has the highest similarity score; scores decrease monotonically down the ranked list
• Number of returned candidates does not exceed the specified top_n_candidates parameter
• Retrieved reference molecules can be traced back to identifiers in the reference database pickle file or flat file

Limitations

• Screening quality depends critically on the trained CNN model's embedding space; poor model calibration yields unreliable matches
• Reference database completeness and representation bias affect coverage; absent or underrepresented compound classes will yield no or low-quality matches
• Cosine similarity or Euclidean distance may not capture all relevant molecular features; Tanimoto similarity based on SMILES can complement embedding-based scoring if available
• Computational cost scales with reference database size; very large databases (millions of compounds) may require approximate nearest-neighbor methods or database indexing
• Threshold selection is empirically driven and dataset-dependent; no universal threshold guarantees optimal precision–recall trade-off across all compound classes

Evidence

• [other] The framework is designed to match predictions from a trained CNN model against a reference database to identify potential candidate molecules for metabolite identification.: "match predictions from a trained CNN model against a reference database to identify potential candidate molecules"
• [other] Compute pairwise similarity scores between query embeddings and all reference database embeddings using a distance metric (e.g., cosine similarity or Euclidean distance).: "Compute pairwise similarity scores between query embeddings and all reference database embeddings using a distance metric (e.g., cosine similarity or Euclidean distance)"
• [other] Filter candidate matches by applying a similarity threshold to retain only high-confidence matches.: "Filter candidate matches by applying a similarity threshold to retain only high-confidence matches"
• [readme] Matches predicted embeddings with a reference database to find top candidate molecules based on cosine similarity.: "Matches predicted embeddings with a reference database to find top candidate molecules based on cosine similarity"
• [readme] top_n_candidates: Number of top candidate molecules to retrieve from the reference database.: "top_n_candidates: Number of top candidate molecules to retrieve from the reference database"