Research map / Probabilistic and scientific modeling / Molecules and proteins
MADGEN: Mass-Spec attends to De Novo Molecular 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.
Proposes MADGEN, a two-stage de novo molecule generator that retrieves a Murcko scaffold from an MS/MS spectrum via contrastive ranking and then generates the full molecule from that scaffold using a spectrum-conditioned graph-transformer Markov bridge with classifier-free guidance.
Problem
De novo annotation of MS/MS spectra into molecular structures is very hard because of huge chemical diversity and limited reference databases; direct generation over the full molecular space is intractable.
Approach
Decompose de novo generation into scaffold retrieval (contrastive spectrum-scaffold alignment) and scaffold-conditioned generation, where spectral peaks are tokenized, self-attended, and injected via cross-attention plus classifier-free guidance into a discrete Markov bridge over edges.
Key results
- With the oracle scaffold retriever, MADGEN reaches 49.0% top-1 accuracy on NIST23 and 18.7% on CANOPUS, and 10.5% top-1 on MassSpecGym.
- With the predictive scaffold retriever, scaffold prediction accuracy ranges from 13.2% to 40.3%, and top-1 accuracy is 4.6% (NIST23), 2.10% (CANOPUS), 1.31% (MassSpecGym).
- Tokenization with self-attention plus cross-attention and node-level CFG improves top-1 accuracy from 4.3% (binning + MLP concatenation) to 49.0% on NIST23.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the oracle correct-scaffold retriever instead of the predictive retriever | MADGEN · NIST23 | Top-1 accuracy: 4.6% -> 49.0% | Helped |
| Use the oracle correct-scaffold retriever instead of the predictive retriever | MADGEN · CANOPUS | Top-1 accuracy: 2.10% -> 18.7% | Helped |
| Use the oracle correct-scaffold retriever instead of the predictive retriever | MADGEN · MassSpecGym | Top-1 accuracy: 1.31% -> 10.5% | Helped |
| Replace tokenization + self-attention and cross-attention with binning + MLP and concatenation (ablation) | MADGEN oracle scaffold generation · NIST | Accuracy: 49.0% -> 4.30% | Hurt |
| Remove self-attention from spectral tokenization (ablation) | MADGEN oracle scaffold generation · NIST | Accuracy: 49.0% -> 13.0% | Hurt |
| Remove CFG while retaining tokenization, self-attention, and cross-attention (ablation) | MADGEN oracle scaffold generation · NIST | Accuracy: 49.0% -> 42.5% | Hurt |
| Apply CFG to edges only rather than nodes only (ablation) | MADGEN oracle scaffold generation · NIST | Accuracy: 49.0% -> 42.0% | Hurt |
| Apply CFG to both nodes and edges rather than nodes only (ablation) | MADGEN oracle scaffold generation · NIST | Accuracy: 49.0% -> 45.9% | Hurt |
Surprising findings
- Predictive scaffold retrieval is the bottleneck: the oracle retriever raises MassSpecGym top-1 accuracy from 1.31% to 10.5%.
Limitations
- Predictive scaffold retrieval accuracy is low (13.2%-40.3%), so end-to-end performance is limited.
- Performance degrades as the number of free atoms to connect to the scaffold increases (CANOPUS worst).
- Two-stage pipeline may compound errors; not end-to-end.
Experimental setup
Datasets: NIST23 CANOPUS MassSpecGym
Benchmarks: NIST23 CANOPUS MassSpecGym
Metrics: top-1 accuracy top-10 accuracy Tanimoto similarity MCES scaffold prediction accuracy
Baselines: Spec2Mol MSNovelist MassGenie SMILES Transformer SELFIES Transformer Random generation
Builds on: CLIP JESTR RetroBridge classifier-free guidance Spec2Mol
Closest papers by meaning
- MS-BART: Unified Modeling of Mass Spectra and Molecules for Structure Elucidation (NeurIPS 2025 · similarity 0.89)
- MassSpecGym: A benchmark for the discovery and identification of molecules (NeurIPS 2024 · similarity 0.89)
- DiffMS: Diffusion Generation of Molecules Conditioned on Mass Spectra (ICML 2025 · similarity 0.85)
- GLACIER: Rethinking Mass Spectrum Prediction as an Object Detection Problem (NeurIPS 2026 · similarity 0.84)
- From Artificially Real to Real: Leveraging Pseudo Data from Large Language Models for Low-Resource Molecule Discovery (AAAI 2024 · similarity 0.84)
- PepSpecBench: A Unified Evaluation Benchmark for Peptide Tandem Mass Spectrometry Prediction (NeurIPS 2026 · similarity 0.84)
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.