Research map / Graph learning / Graph self-supervised learning
Neural Graph Matching Improves Retrieval Augmented Generation in Molecular Machine Learning
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 MARASON, a retrieval-augmented mass spectrum simulator that uses neural graph matching between target and reference fragmentation DAGs to improve ICEBERG's spectrum prediction and top-1 retrieval accuracy.
Problem
Retrieval augmentation for molecular machine learning is poorly understood; naive concatenation of retrieved references gives little benefit, and fixed-affinity graph matching is brittle for non-identical structures. The paper targets tandem mass spectrum simulation.
Approach
Apply end-to-end learned neural graph matching (learned node affinities with a differentiable Softmax matching layer) over fragmentation DAGs of target and retrieved reference molecules, and use the aligned reference fragment intensities as input to the intensity predictor built on ICEBERG.
Key results
- On the NIST20 random split, MARASON reaches 0.278 top-1 retrieval accuracy versus 0.189 for ICEBERG and 0.202 for ICEBERG with collision energy.
- On MassSpecGym, MARASON improves top-1 retrieval accuracy from 0.3193 (FraGNNet) to 0.3403.
- Naive concatenation of the reference spectrum reduces cosine similarity (0.737 vs 0.739 no-RAG), while neural graph matching with Softmax improves it to 0.753 (shared GNN) and 0.757 (separate GNNs).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add neural graph matching-based retrieval augmentation | ICEBERG · NIST20 random split, positive adducts | Top-1 retrieval accuracy: 18.7\% to 27.8\% | Helped |
| Add neural graph matching-based retrieval augmentation | ICEBERG · NIST20 random split | Cosine similarity: 0.727$\pm$0.002 -> 0.765$\pm$0.002 | Helped |
| Add neural graph matching-based retrieval augmentation | ICEBERG · NIST20 scaffold split | Cosine similarity: 0.699 -> 0.725 | Helped |
| Add neural graph matching-based retrieval augmentation | FraGNNet · MassSpecGym | Top-1 retrieval accuracy: 31.93\% to 34.03\% | Helped |
| Add collision energy input (ablation) | ICEBERG · NIST20 random split, positive adducts | Top-1 retrieval accuracy: 0.189$\pm$0.012 -> 0.202$\pm$0.009 | Helped |
| Use NGM with Softmax matching instead of Sinkhorn (ablation) | MARASON with shared GNN · NIST20 random split, seed 1 | Cosine similarity: 0.749 -> 0.753 | Helped |
| Use separate GNN weights for matching and intensity prediction (ablation) | MARASON with NGM and Softmax · NIST20 random split, seed 1 | Cosine similarity: 0.753 -> 0.757 | Helped |
| Replace neural graph matching with Hungarian matching (ablation) | MARASON · NIST20 random split, seed 1 | Cosine similarity: 0.753 -> 0.746 | Hurt |
| Replace neural graph matching with RRWM (ablation) | MARASON · NIST20 random split, seed 1 | Cosine similarity: 0.753 -> 0.742 | Hurt |
| Concatenate the retrieved spectrum to the model input (ablation) | MARASON · NIST20 random split, seed 1 | Cosine similarity: 0.739 -> 0.737 | Hurt |
| Remove retrieval augmentation (ablation) | MARASON · NIST20 random split, seed 1 | Cosine similarity: 0.753 -> 0.739 | Hurt |
| Use retrieval augmentation when reference Tanimoto similarity is above 0.2 (ablation) | MARASON · NIST20 random split, Tanimoto similarity (0.1, 0.2] | Cosine similarity: 0.566 -> 0.550 | Hurt |
Surprising findings
- Naive RAG by concatenating the retrieved spectrum slightly hurts cosine similarity, while traditional graph matching gives only minor gains.
- Softmax matching outperformed Sinkhorn despite Sinkhorn being the usual choice for smaller graphs.
Limitations
- Gains from RAG vanish when the best reference has Tanimoto similarity below about 0.2-0.3.
- Traditional matching baselines show only minor benefit, and the approach is evaluated only on mass spectrum simulation rather than broader molecular property tasks.
- MassSpecGym experiments require pseudo-labeled collision energies from a NIST-pretrained model.
Experimental setup
Datasets: NIST20 (HCD spectra, 530,640 spectra, 25,541 structures) MassSpecGym
Benchmarks: NIST20 random split retrieval accuracy NIST20 scaffold split retrieval accuracy MassSpecGym retrieval accuracy
Metrics: cosine similarity top-k retrieval accuracy
Baselines: ICEBERG ICEBERG (w/ collision energy) MassFormer SCARF NEIMS 3DMolMS FixedVocab FraGNNet Hungarian matching RRWM
Builds on: ICEBERG RRWM Sinkhorn set transformer Pygmtools
Closest papers by meaning
- On the Scalability of GNNs for Molecular Graphs (NeurIPS 2024 · similarity 0.86)
- Bridging Molecular Graphs and Large Language Models (AAAI 2025 · similarity 0.85)
- Expressivity and Generalization: Fragment-Biases for Molecular GNNs (ICML 2024 · similarity 0.85)
- Neural Atoms: Propagating Long-range Interaction in Molecular Graphs through Efficient Communication Channel (ICLR 2024 · similarity 0.85)
- GLACIER: Rethinking Mass Spectrum Prediction as an Object Detection Problem (NeurIPS 2026 · similarity 0.85)
- A Simple and Scalable Representation for Graph Generation (ICLR 2024 · similarity 0.85)
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.