HMDB-like Search Engine
Mass Spectrometry Search Prototype
A lightweight testing home page for navigating the current prototype: compound search, LC-MS/adduct m/z search, neutral mass search, LC-MS/MS cosine search, compound detail pages, source hierarchy display, related spectra inspection, and MS/MS spectrum detail visualization.
Compound Search
Search compounds by name, inspect formulas and molecular weights, and open compound detail pages.
LC-MS / Adduct m/z Search
Search observed m/z values using ion mode, adduct type, tolerance, source filtering, multi-select adducts, and stacked multi-mass results.
Neutral Mass Search
Search compounds directly by neutral monoisotopic mass using tolerance, source filtering, stacked results, pagination, and sorting.
LC-MS/MS Search
Search a pasted MS/MS peak list against library spectra using fragment tolerance, spectrum kind, source filtering, greedy cosine similarity, and mirror spectrum comparison.
MS/MS Spectrum Detail Example
Open a known MS/MS spectrum detail page with metadata, compound context, HMDB link, peak graph tooltip, sortable peak table, rows selector, and pagination.
Creatine Search Demo
Open compound search with Creatine preloaded for testing pagination, sorting, and return links.
Creatine Detail
Open a known compound detail page with source hierarchy and related spectra for quick testing.
LC-MS Example
Open the LC-MS/adduct m/z search with example m/z values and default positive-mode settings.
Neutral Mass Example
Open neutral mass search with multiple example masses, ppm tolerance, source filtering support, sorting, and per-section pagination.
Current working workflow
- 1. Use the home page to choose compound search, LC-MS/adduct m/z search, neutral mass search, or LC-MS/MS cosine search.
- 2. Search compounds by name, sort/paginate results, and open compound detail pages.
- 3. Search observed m/z values with ion mode, multi-select adduct type, tolerance, and source filters.
- 4. Search neutral monoisotopic masses with tolerance, source filters, sorting, and stacked multi-mass sections.
- 5. Search MS/MS peak lists with fragment tolerance, spectrum kind, source filtering, minimum matched peaks, cosine similarity, and mirror spectrum comparison.
- 6. Inspect MS/MS spectrum metadata, compound context, HMDB external link, peak graph tooltip, and sortable paginated peak table.
Coming next
- Make related spectra HMDB ID cells open the local MS/MS spectrum detail page
- Primitive fragment peak lookup page
- Sortable LC-MS/MS search result columns after cosine scoring is validated
- Research/ML API for deeper cosine internals, all matched pairs, candidate statistics, and parameter sweeps
- Parent ion mass/adduct-aware filtering after precursor metadata design is finalized
- Frontend user guide expansion once the search workflows are stable
Model Evaluation
Machine Learning Algorithm Comparison
To determine the most robust approach for matching and aligning candidate spectra, we evaluated traditional baselines alongside three machine learning classifiers. Models were trained using query-level 5-fold cross-validation (GroupKFold) over verified positive identifications from DSTB Saliva samples.
| Algorithm | MRR | Hit@1 | Hit@5 | Hit@10 |
|---|---|---|---|---|
| Cosine Similarity (Baseline) | 0.6008 | 47.73% | 72.73% | 84.09% |
| Precursor Mass Error (Baseline) | 0.7746 | 70.45% | 84.09% | 88.64% |
| Random Forest Classifier | 0.7930 | 72.73% | 84.09% | 88.64% |
| Gradient Boosting Classifier | 0.7526 | 63.64% | 88.64% | 90.91% |
| Support Vector Machine (SVM) | 0.7701 | 65.91% | 88.64% | 95.45% |
* Key findings: Random Forest achieved the highest Mean Reciprocal Rank (MRR = 0.7930) and Hit@1 accuracy (72.73%). SVM achieved the highest Hit@10 accuracy (95.45%). All models significantly outperformed traditional cosine alignment.