Model type
Frozen spectral/molecular encoders with learned alignment projections
MSAlign retrieves candidate molecules from tandem mass spectra by aligning pretrained molecular and spectral representations.
Conceptual summary of the documented data flow; optional inputs and configured downstream stages must be reported for a reproducible evaluation.
Frozen spectral/molecular encoders with learned alignment projections
MS/MS spectrum and a set of candidate molecular structures.
Candidate-molecule retrieval scores in a shared representation space.
Official project documentation and implementation: https://arxiv.org/abs/2605.19752
limited source coverage · Automated source review, 2026-09-16. All specifications and missing details
Deployment facts recorded for MSAlign. Fields the record does not carry are marked not recorded.
4 evaluations · 12 results. Different protocols are not a single leaderboard.
Sorted by Top-1 accuracy (retrieval) (higher is better). The best value in each column is highlighted. Decimals are rounded for display; each value links to the printed value and its source.
Applied filters: All linked evaluations
| Tested configuration | Protocol and dataset | Finding | Evidence and details |
|---|---|---|---|
| Configuration: MSAlign | Protocol: MSAlign molecular retrieval MassSpecGym formula split: no formula, R@1: MassSpecGym formula split retrieval with no formula Dataset subset: MassSpecGym formula split, no formula candidate pool (MSAlign molecular retrieval split) | 53.8% Top-1 accuracy (retrieval) percent · higher Uncertainty: Not yet extracted Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · Source checkedMethods, coverage and sourceSection 5.1 candidate retrieval: 256 mass-matched PubChem candidates per spectrum without formula; formula-conditioned methods can exploit true formula and MSAlign+Filter reduces the pool to about 100. Preserve the named source split; exact original split hashes remain unextracted. Aggregation: Not reported msalign: Primary paper PDF · Table 3, MassSpecGym formula split, row R@1, column 8 (MSAlign) |
| Configuration: MSAlign | Protocol: MSAlign molecular retrieval MassSpecGym formula split: no formula, R@20: MassSpecGym formula split retrieval with no formula Dataset subset: MassSpecGym formula split, no formula candidate pool (MSAlign molecular retrieval split) | 87.1% Top-20 accuracy (retrieval) percent · higher Uncertainty: Not yet extracted Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · Source checkedMethods, coverage and sourceSection 5.1 candidate retrieval: 256 mass-matched PubChem candidates per spectrum without formula; formula-conditioned methods can exploit true formula and MSAlign+Filter reduces the pool to about 100. Preserve the named source split; exact original split hashes remain unextracted. Aggregation: Not reported msalign: Primary paper PDF · Table 3, MassSpecGym formula split, row R@20, column 8 (MSAlign) |
| Configuration: MSAlign | Protocol: MSAlign molecular retrieval MassSpecGym formula split: no formula, R@5: MassSpecGym formula split retrieval with no formula Dataset subset: MassSpecGym formula split, no formula candidate pool (MSAlign molecular retrieval split) | 73.1% Top-5 accuracy (retrieval) percent · higher Uncertainty: Not yet extracted Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · Source checkedMethods, coverage and sourceSection 5.1 candidate retrieval: 256 mass-matched PubChem candidates per spectrum without formula; formula-conditioned methods can exploit true formula and MSAlign+Filter reduces the pool to about 100. Preserve the named source split; exact original split hashes remain unextracted. Aggregation: Not reported msalign: Primary paper PDF · Table 3, MassSpecGym formula split, row R@5, column 8 (MSAlign) |
| Configuration: MSAlign | Protocol: MSAlign molecular retrieval MassSpecGym MCES split: no formula, R@1: MassSpecGym MCES split retrieval with no formula Dataset subset: MassSpecGym MCES split, no formula candidate pool (MSAlign molecular retrieval split) | 16.2% Top-1 accuracy (retrieval) percent · higher Uncertainty: Not yet extracted Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · Source checkedMethods, coverage and sourceSection 5.1 candidate retrieval: 256 mass-matched PubChem candidates per spectrum without formula; formula-conditioned methods can exploit true formula and MSAlign+Filter reduces the pool to about 100. Preserve the named source split; exact original split hashes remain unextracted. Aggregation: Not reported msalign: Primary paper PDF · Table 3, MassSpecGym MCES split, row R@1, column 8 (MSAlign) |
| Configuration: MSAlign | Protocol: MSAlign molecular retrieval MassSpecGym MCES split: no formula, R@20: MassSpecGym MCES split retrieval with no formula Dataset subset: MassSpecGym MCES split, no formula candidate pool (MSAlign molecular retrieval split) | 59.9% Top-20 accuracy (retrieval) percent · higher Uncertainty: Not yet extracted Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · Source checkedMethods, coverage and sourceSection 5.1 candidate retrieval: 256 mass-matched PubChem candidates per spectrum without formula; formula-conditioned methods can exploit true formula and MSAlign+Filter reduces the pool to about 100. Preserve the named source split; exact original split hashes remain unextracted. Aggregation: Not reported msalign: Primary paper PDF · Table 3, MassSpecGym MCES split, row R@20, column 8 (MSAlign) |
| Configuration: MSAlign | Protocol: MSAlign molecular retrieval MassSpecGym MCES split: no formula, R@5: MassSpecGym MCES split retrieval with no formula Dataset subset: MassSpecGym MCES split, no formula candidate pool (MSAlign molecular retrieval split) | 35.6% Top-5 accuracy (retrieval) percent · higher Uncertainty: Not yet extracted Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · Source checkedMethods, coverage and sourceSection 5.1 candidate retrieval: 256 mass-matched PubChem candidates per spectrum without formula; formula-conditioned methods can exploit true formula and MSAlign+Filter reduces the pool to about 100. Preserve the named source split; exact original split hashes remain unextracted. Aggregation: Not reported msalign: Primary paper PDF · Table 3, MassSpecGym MCES split, row R@5, column 8 (MSAlign) |
| Configuration: MSAlign | Protocol: MSAlign molecular retrieval NPLIB1: no formula, R@1: NPLIB1 retrieval with no formula Dataset subset: NPLIB1, no formula candidate pool (MSAlign molecular retrieval split) | 31.8% Top-1 accuracy (retrieval) percent · higher Uncertainty: Not yet extracted Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · Source checkedMethods, coverage and sourceMSAlign on MSAlign molecular retrieval NPLIB1: no formula, R@1: NPLIB1 retrieval with no formula Section 5.1 candidate retrieval: 256 mass-matched PubChem candidates per spectrum without formula; formula-conditioned methods can exploit true formula and MSAlign+Filter reduces the pool to about 100. Preserve the named source split; exact original split hashes remain unextracted. Aggregation: Not reported msalign: Primary paper PDF · Table 3, NPLIB1, row R@1, column 8 (MSAlign) |
| Configuration: MSAlign | Protocol: MSAlign molecular retrieval NPLIB1: no formula, R@20: NPLIB1 retrieval with no formula Dataset subset: NPLIB1, no formula candidate pool (MSAlign molecular retrieval split) | 82.9% Top-20 accuracy (retrieval) percent · higher Uncertainty: Not yet extracted Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · Source checkedMethods, coverage and sourceMSAlign on MSAlign molecular retrieval NPLIB1: no formula, R@20: NPLIB1 retrieval with no formula Section 5.1 candidate retrieval: 256 mass-matched PubChem candidates per spectrum without formula; formula-conditioned methods can exploit true formula and MSAlign+Filter reduces the pool to about 100. Preserve the named source split; exact original split hashes remain unextracted. Aggregation: Not reported msalign: Primary paper PDF · Table 3, NPLIB1, row R@20, column 8 (MSAlign) |
| Configuration: MSAlign | Protocol: MSAlign molecular retrieval NPLIB1: no formula, R@5: NPLIB1 retrieval with no formula Dataset subset: NPLIB1, no formula candidate pool (MSAlign molecular retrieval split) | 60.4% Top-5 accuracy (retrieval) percent · higher Uncertainty: Not yet extracted Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · Source checkedMethods, coverage and sourceMSAlign on MSAlign molecular retrieval NPLIB1: no formula, R@5: NPLIB1 retrieval with no formula Section 5.1 candidate retrieval: 256 mass-matched PubChem candidates per spectrum without formula; formula-conditioned methods can exploit true formula and MSAlign+Filter reduces the pool to about 100. Preserve the named source split; exact original split hashes remain unextracted. Aggregation: Not reported msalign: Primary paper PDF · Table 3, NPLIB1, row R@5, column 8 (MSAlign) |
| Configuration: MSAlign | Protocol: MSAlign molecular retrieval Spectraverse: no formula, R@1: Spectraverse retrieval with no formula Dataset subset: Spectraverse, no formula candidate pool (MSAlign molecular retrieval split) | 32.3% Top-1 accuracy (retrieval) percent · higher Uncertainty: Not yet extracted Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · Source checkedMethods, coverage and sourceSection 5.1 candidate retrieval: 256 mass-matched PubChem candidates per spectrum without formula; formula-conditioned methods can exploit true formula and MSAlign+Filter reduces the pool to about 100. Preserve the named source split; exact original split hashes remain unextracted. Aggregation: Not reported msalign: Primary paper PDF · Table 3, Spectraverse, row R@1, column 8 (MSAlign) |
| Configuration: MSAlign | Protocol: MSAlign molecular retrieval Spectraverse: no formula, R@20: Spectraverse retrieval with no formula Dataset subset: Spectraverse, no formula candidate pool (MSAlign molecular retrieval split) | 79.6% Top-20 accuracy (retrieval) percent · higher Uncertainty: Not yet extracted Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · Source checkedMethods, coverage and sourceSection 5.1 candidate retrieval: 256 mass-matched PubChem candidates per spectrum without formula; formula-conditioned methods can exploit true formula and MSAlign+Filter reduces the pool to about 100. Preserve the named source split; exact original split hashes remain unextracted. Aggregation: Not reported msalign: Primary paper PDF · Table 3, Spectraverse, row R@20, column 8 (MSAlign) |
| Configuration: MSAlign | Protocol: MSAlign molecular retrieval Spectraverse: no formula, R@5: Spectraverse retrieval with no formula Dataset subset: Spectraverse, no formula candidate pool (MSAlign molecular retrieval split) | 59.1% Top-5 accuracy (retrieval) percent · higher Uncertainty: Not yet extracted Coverage: Not reported scored / Not reported eligible | Author-reported evaluation · Source checkedMethods, coverage and sourceSection 5.1 candidate retrieval: 256 mass-matched PubChem candidates per spectrum without formula; formula-conditioned methods can exploit true formula and MSAlign+Filter reduces the pool to about 100. Preserve the named source split; exact original split hashes remain unextracted. Aggregation: Not reported msalign: Primary paper PDF · Table 3, Spectraverse, row R@5, column 8 (MSAlign) |
Source checking is not independent reproduction. Release 2026-11-21-c8bea25e5abb.
Related profile: MSAlign. This page retains the exact record and its evaluation context.
Evaluated retrieval configuration from MSAlign Table 3; model-specific inputs and training are retained in the protocol.
Frozen DreaMS and ChemBERTa encoders connected by lightweight MLP projections trained with a candidate-based contrastive objective. The documented inputs are MS/MS spectrum and a set of candidate molecular structures. The output consists of candidate-molecule retrieval scores in a shared representation space.
MSAlign arXiv:2605.19752v1, submitted 19 May 2026. The applicable input limits require configuration-specific checking.
Inspected pinned official documentation, relevant implementation files and named primary-paper sections. Claims are limited to those artifacts. Remaining field extraction and identity conflicts are explicit; no new performance claims, model runs or human review are implied.
Stable record: discovery-model-msalignExplanatory profile: limited source coverage · Automated source review, 2026-09-16. Review applies to the cited claims; unresolved fields are listed below. Numerical results retain their own review status.
| Property | Description and evidence |
|---|---|
| Model type | Frozen spectral/molecular encoders with learned alignment projectionsSources (2)https://arxiv.org/abs/2605.19752: page.html; msalign: Primary paper PDF · MSAlign paper Section 3 Architecture and Training, Section 4 on splitting, and Section 5.1 Experimental Setting |
| Architecture | Frozen DreaMS and ChemBERTa encoders connected by lightweight MLP projections trained with a candidate-based contrastive objective.Sources (2)https://arxiv.org/abs/2605.19752: page.html; msalign: Primary paper PDF · MSAlign paper Section 3 Architecture and Training, Section 4 on splitting, and Section 5.1 Experimental Setting |
| Inputs | MS/MS spectrum and a set of candidate molecular structures.Sources (2)https://arxiv.org/abs/2605.19752: page.html; msalign: Primary paper PDF · MSAlign paper Section 3 Architecture and Training, Section 4 on splitting, and Section 5.1 Experimental Setting |
| Outputs | Candidate-molecule retrieval scores in a shared representation space.Sources (2)https://arxiv.org/abs/2605.19752: page.html; msalign: Primary paper PDF · MSAlign paper Section 3 Architecture and Training, Section 4 on splitting, and Section 5.1 Experimental Setting |
| Parameters | Approximately 4M trainable projection parameters; frozen DreaMS and ChemBERTa backbones are reported as 96M and 92M respectively.Sources (2)https://arxiv.org/abs/2605.19752: page.html; msalign: Primary paper PDF · MSAlign paper Section 3 Architecture and Training, Section 4 on splitting, and Section 5.1 Experimental Setting |
| Known versions | MSAlign arXiv:2605.19752v1, submitted 19 May 2026.Sources (2)https://arxiv.org/abs/2605.19752: page.html; msalign: Primary paper PDF · MSAlign paper Section 3 Architecture and Training, Section 4 on splitting, and Section 5.1 Experimental Setting |
| Training data | Projection layers are fitted separately on the NPLIB1, MassSpecGym or Spectraverse training splits. The paper distinguishes spectrum/molecule pair counts from unique molecules and controls candidate retrieval using mass matching.Sources (2)https://arxiv.org/abs/2605.19752: page.html; msalign: Primary paper PDF · MSAlign paper Section 3 Architecture and Training, Section 4 on splitting, and Section 5.1 Experimental Setting |
| Training cutoff | NPLIB1, MassSpecGym and Spectraverse are separately split training resources. The inspected paper does not define one latest measurement date covering all three. · Not reported in inspected sourcesSources (2)https://arxiv.org/abs/2605.19752: page.html; msalign: Primary paper PDF · MSAlign paper Section 3 Architecture and Training, Section 4 on splitting, and Section 5.1 Experimental Setting |
| Context limits | The pipeline inherits spectrum and molecule preprocessing from its frozen DreaMS and ChemBERTa encoders. The inspected MSAlign architecture section does not state a single joint input limit. · Not reported in inspected sourcesSources (2)https://arxiv.org/abs/2605.19752: page.html; msalign: Primary paper PDF · MSAlign paper Section 3 Architecture and Training, Section 4 on splitting, and Section 5.1 Experimental Setting |
| Weights licence | The inspected preprint does not state distribution terms for the learned projection checkpoints. Frozen encoder licences remain separate from projection-weight rights. · Not reported in inspected sourcesSources (2)https://arxiv.org/abs/2605.19752: page.html; msalign: Primary paper PDF · MSAlign paper Section 3 Architecture and Training, Section 4 on splitting, and Section 5.1 Experimental Setting; page.html: inspected official source |
| Access | Official project documentation and implementation: https://arxiv.org/abs/2605.19752Sources (2)https://arxiv.org/abs/2605.19752: page.html; msalign: Primary paper PDF · MSAlign paper Section 3 Architecture and Training, Section 4 on splitting, and Section 5.1 Experimental Setting |
| Code licence | The inspected preprint describes the algorithm but does not supply a separate code licence; no code-distribution permission is inferred from the paper licence. · Not reported in inspected sourcesSourceshttps://arxiv.org/abs/2605.19752: page.html · page.html: inspected official source |
Source checking verifies the cited claim or transcription. It does not establish independent reproduction.
Trace each statement to its source and review. A context-only reference supports the record generally; it does not verify an individual field. Source checking does not reproduce an experiment.
One row per statement and cited source. Multiple citations are not independent evaluations. Shared locators are labelled explicitly.
3 evidence rows matching the loaded filters
| Property and statement | Original source and location | Review and provenance |
|---|---|---|
Facets: method types1 values
| msalign: Primary paper PDF MSAlign arXiv 2605.19752v2 abstract and Section 3 "Foundation model alignment" (p.4): aligns "two frozen foundation models (DreaMS for mass spectra and MolDeBERTa for molecules)" through trained MLP adapters Shared locator for this statement’s cited sources; not a separate locator for each citation. Version: 2605.19752v1 | conflicting claim ["source-hash-verification","automated-source-review"] · 2026-10-12 Audit detailsCorrect the method type from none to foundation_model (use-case baseline review 2026-10-12, table 1). Checked as the review required. Score fusion (Section 3, "Late score fusion") combines four alignment scorers and three mass scores with weights fitted on validation data; two of the scorers use both foundation encoders. Evidence from docs/reviews/use-cases/baselines-2026-10-12.md. Independent review 2026-10-12: approved. MSAlign v1 Section 5.1 "Baselines" and Table 1: FFN and DeepSets are fingerprint predictors from MassSpecGym; JESTR, Emb-Cos and SAIL are trained with InfoNCE losses; MSAlign aligns two frozen foundation models. See docs/reviews/2026-10-12-method-type-corrections.md. Conflicting source claim: Field: Claim: metadata-correction-tags-20261012-3d4413567b5191f22f5a Source artifact SHA-256: Hash scope: SHA-256 of retrieved original artifact bytes Format: original_artifact Extraction artifact SHA-256: |
Facets: method types1 values
| MSAlign arXiv v2 Table 3 MSAlign arXiv 2605.19752v2 abstract and Section 3 "Foundation model alignment" (p.4): aligns "two frozen foundation models (DreaMS for mass spectra and MolDeBERTa for molecules)" through trained MLP adapters Shared locator for this statement’s cited sources; not a separate locator for each citation. Version: 2605.19752v2, 25 September 2026 | conflicting claim ["source-hash-verification","automated-source-review"] · 2026-10-12 Audit detailsCorrect the method type from none to foundation_model (use-case baseline review 2026-10-12, table 1). Checked as the review required. Score fusion (Section 3, "Late score fusion") combines four alignment scorers and three mass scores with weights fitted on validation data; two of the scorers use both foundation encoders. Evidence from docs/reviews/use-cases/baselines-2026-10-12.md. Independent review 2026-10-12: approved. MSAlign v1 Section 5.1 "Baselines" and Table 1: FFN and DeepSets are fingerprint predictors from MassSpecGym; JESTR, Emb-Cos and SAIL are trained with InfoNCE losses; MSAlign aligns two frozen foundation models. See docs/reviews/2026-10-12-method-type-corrections.md. Conflicting source claim: Field: Claim: metadata-correction-tags-20261012-3d4413567b5191f22f5a Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record Extraction artifact SHA-256: |
| Relationship: family discovery-model-msalign Individual claims | msalign: Primary paper PDF Table 3, column MSAlign; Section 5.1 Version: 2605.19752v1 | source checked ["automated-source-review"] · 2026-09-23 Audit detailsSource-backed evaluated identity only; no independent reproduction. Field: Claim: msalign-2026-table3-method-msalign-discovery-model-msalign-identity-claim Source artifact SHA-256: Hash scope: SHA-256 of retrieved original artifact bytes Format: original_artifact |
Release 2026-11-21-c8bea25e5abb · Record review: source checked
Technical metadata for this record is tracked under stable ID msalign-2026-table3-method-msalign.
Source locator: Table 3, column MSAlign; Section 5.1
MSAlign. Rewire biological model benchmark database, release 2026-11-21-c8bea25e5abb. https://benchmarks.rewirebio.io/database/configuration/msalign-2026-table3-method-msalign/
Release 2026-11-21-c8bea25e5abb. Rankings and values change between releases, so keep the release id in the citation. Cite the original publication as well for any value you reuse: each result links its source.
The compilation and annotations made by Rewire in this release are offered under CC BY 4.0; values and text taken from cited sources stay under the terms of those sources and are not covered by this licence. Data licence