Research map / Probabilistic and scientific modeling / Molecules and proteins
MS-BART: Unified Modeling of Mass Spectra and Molecules for Structure Elucidation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MS-BART, a BART-base language model that maps MIST-predicted or RDKit-computed molecular fingerprints and SELFIES molecules into a shared vocabulary, pretrained on 4M fingerprint-molecule pairs, finetuned on experimental spectra, and aligned with a Tanimoto-based chemical-feedback rank loss for mass-spectrum structure elucidation.
Problem
De novo structure elucidation from tandem mass spectra is bottlenecked by scarce annotated experimental spectra and the heterogeneity of raw spectral signals, and generated molecules often hallucinate.
Approach
Represent spectra as MIST-predicted fingerprint tokens and molecules as SELFIES tokens in one shared vocabulary; pretrain a BART-style model with multi-task denoising and fingerprint-to-molecule translation on RDKit-computed fingerprints, then finetune on experimental data and align with a Tanimoto-ranked contrastive loss.
Key results
- On NPLIB1 Top-1, MS-BART reaches 9.66 MCES and 0.44 Tanimoto, best among methods, while DiffMS has higher Top-1 accuracy (8.34% vs 7.45%).
- On MassSpecGym Top-1, MS-BART attains 16.47 MCES and 0.23 Tanimoto; DiffMS leads accuracy (2.30% vs 1.07%).
- MS-BART inference takes about 3 s per spectrum on an RTX 4090 at beam width 100, versus about 160 s for DiffMS, roughly 53x faster.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add multi-task pretraining (ablation) | MS-Bart · NPLIB1 | Top-1 accuracy: 1.71% -> 7.45% | Helped |
| add multi-task pretraining (ablation) | MS-Bart · NPLIB1 | Top-10 accuracy: 3.05% -> 10.99% | Helped |
| use SELFIES denoising pretraining alone instead of multi-task pretraining (ablation) | MS-Bart · NPLIB1 | Top-1 accuracy: 7.45% -> 0.37% | Hurt |
| use fingerprint-to-molecule translation pretraining alone instead of multi-task pretraining (ablation) | MS-Bart · NPLIB1 | Top-1 accuracy: 7.45% -> 6.23% | Hurt |
| fine-tune the pretrained model on experimental spectra (ablation) | MS-Bart · MassSpecGym | Top-1 accuracy: 0.00% -> 1.07% | Helped |
| add chemical-feedback rank alignment after fine-tuning (ablation) | MS-Bart · MassSpecGym | Top-1 Tanimoto | Helped |
| add chemical-feedback rank alignment after fine-tuning (ablation) | MS-Bart · MassSpecGym | MCES | Hurt |
| raise fingerprint threshold from ε = 0.11 to ε = 0.15 (ablation) | MS-Bart · MassSpecGym | Top-1 accuracy: 1.07% -> 1.20% | No clear effect |
| raise fingerprint threshold from ε = 0.11 to ε = 0.20 (ablation) | MS-Bart · MassSpecGym | Top-1 accuracy: 1.07% -> 1.19% | No clear effect |
| set rank loss weight α to 1 instead of 5 (ablation) | MS-Bart · NPLIB1 | Top-1 accuracy: 7.45% -> 7.08% | No clear effect |
| set rank loss weight α to 3 instead of 5 (ablation) | MS-Bart · NPLIB1 | Top-1 accuracy: 7.45% -> 7.20% | No clear effect |
| vary decoding temperature from 0.2 to 0.8 (ablation) | MS-Bart · MassSpecGym | Top-1 accuracy: 1.07% -> 1.07% | No clear effect |
Surprising findings
- Pretraining with denoising alone degrades performance on NPLIB1 relative to no pretraining, while fingerprint-to-molecule translation gives the largest gain.
- Output quality is nearly insensitive to sampling temperature, attributed to high model confidence and a small 185-token vocabulary.
Limitations
- Performance depends heavily on the external MIST model's fingerprint prediction accuracy.
- Formula information is used only for re-ranking, not incorporated into training.
- Top-1 and Top-10 accuracy do not surpass DiffMS on most metrics.
Experimental setup
Datasets: MassSpecGym NPLIB1 4M unlabeled molecules from MassSpecGym (RDKit fingerprints)
Benchmarks: MassSpecGym NPLIB1
Metrics: Top-k accuracy Top-k maximum Tanimoto similarity Top-k minimum MCES inference time
Baselines: Spec2Mol MIST + Neuraldecipher MIST + MSNovelist MADGEN DiffMS MassSpecGym SMILES Transformer MassSpecGym SELFIES Transformer Random Generation
Builds on: BART MIST MSNovelist DiffMS MassSpecGym
Closest papers by meaning
- MassSpecGym: A benchmark for the discovery and identification of molecules (NeurIPS 2024 · similarity 0.90)
- MADGEN: Mass-Spec attends to De Novo Molecular generation (ICLR 2025 · similarity 0.89)
- GLACIER: Rethinking Mass Spectrum Prediction as an Object Detection Problem (NeurIPS 2026 · similarity 0.88)
- Unraveling Molecular Structure: A Multimodal Spectroscopic Dataset for Chemistry (NeurIPS 2024 · similarity 0.87)
- DiffMS: Diffusion Generation of Molecules Conditioned on Mass Spectra (ICML 2025 · similarity 0.87)
- PepSpecBench: A Unified Evaluation Benchmark for Peptide Tandem Mass Spectrometry Prediction (NeurIPS 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.