Research map / Probabilistic and scientific modeling / Molecules and proteins

MassSpecGym: A benchmark for the discovery and identification of molecules

NeurIPS 2024 · Probabilistic and scientific modeling › Molecules and proteins · approach: Structure-based 3D molecule generation

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.

Introduces MassSpecGym, a public benchmark of 231k curated tandem mass spectra paired with molecules, with standardized splits and baselines for de novo generation, molecule retrieval, and spectrum simulation.

Problem

Molecular identification from tandem mass spectrometry (MS/MS) spectra is hard and lacks standardized, ML-ready benchmarks with leakage-aware splits and evaluation metrics.

Approach

Curate a large, cleaned MS/MS dataset from public libraries and in-house data, split it with MCES-based clustering to prevent structural leakage, and define three challenges (de novo generation, molecule retrieval, spectrum simulation) with baselines and a modular codebase.

Key results

• MIST is the strongest molecule retrieval baseline among those compared, outperforming Fingerprint FFN and DeepSets on the MCES @ 1 metric (exact values in main text, not given in appendix).
• The final cleaned dataset contains 231,104 tandem mass spectra over 31,602 unique SMILES after multi-stage filtering from 414,049 spectra with metadata.

Surprising findings

• Data cleaning removed 135,190 spectra with less than 50% explained intensity by formula decomposition, though these accounted for only 926 unique structures.

Limitations

• Only positive ESI mode LC-MS/MS spectra with [M+H]+ or [M+Na]+ adducts are included, limiting real-world coverage.
• Heterogeneous, imbalanced public libraries may contain mislabeled spectra despite filtering.
• Limited to MS/MS spectra; excludes EI and NMR spectra and isotope/chimeric signal handling.

Experimental setup

Datasets: MoNA MassBank GNPS NIST20 MassIVE GNPS (GeMS-A10) PubChem

Benchmarks: MassSpecGym

Metrics: MCES @ 1 Tanimoto similarity JSS top-k accuracy

Baselines: Random chemical generation SMILES Transformer SELFIES Transformer Fingerprint FFN DeepSets DeepSets + Fourier features MIST FFN Fingerprint GNN FraGNNet

Builds on: MIST NEIMS FraGNNet DeepSets matchms

Closest papers by meaning

• MS-BART: Unified Modeling of Mass Spectra and Molecules for Structure Elucidation (NeurIPS 2025 · similarity 0.90)
• PepSpecBench: A Unified Evaluation Benchmark for Peptide Tandem Mass Spectrometry Prediction (NeurIPS 2026 · similarity 0.89)
• 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 0.89)
• DiffMS: Diffusion Generation of Molecules Conditioned on Mass Spectra (ICML 2025 · similarity 0.87)
• Unraveling Molecular Structure: A Multimodal Spectroscopic Dataset for Chemistry (NeurIPS 2024 · similarity 0.85)