Research map / Detection and segmentation / CNN features

GLACIER: Rethinking Mass Spectrum Prediction as an Object Detection Problem

NeurIPS 2026 · Detection and segmentation › CNN features · approach: Efficient DETR query design · 1 citations (OpenAlex)

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 GLACIER, a single-stage transformer that detects molecular fragments as subgraphs via breakpoint prediction (DETR-style queries, LinSATNet cardinality-constrained projections) to predict MS/MS spectra, reaching SOTA retrieval on MassSpecGym and NIST'20 with ~8x faster inference than ICEBERG 2.0.

Problem

Predicting tandem mass spectra (MS/MS) from molecular structures; existing high-accuracy fragment-based predictors use two-stage pipelines (candidate fragment enumeration then scoring) that are slow and rely on heuristic fragmentation.

Approach

Reformulates MS/MS fragmentation as object detection on molecular graphs: learnable queries predict breakpoint sets (with a predicted cardinality via differentiable top-k projection) matched to labels by Hungarian matching, and fragment intensities are predicted in one pass with shared Graphormer embeddings pooled over fragments, trained with a decaying MAGMa teacher-forcing schedule.

Key results

• On MassSpecGym mass challenge, GLACIER achieves 69.95% Recall@1 with contrastive finetuning vs 63.95% for ICEBERG 2.0 (69.69% without CF).
• On NIST'20 random split (all adducts), Top-1 retrieval rises to 52.5% with CF and 39.0% without, versus 33.5% for ICEBERG 2.0.
• Inference is about 1.3 ms/spectrum vs 10.8 ms for ICEBERG 2.0, a 7.95x speedup on 2 A5000 GPUs.

Surprising findings

• Contrastive finetuning provides little gain on MassSpecGym (69.69 to 69.95 Top-1) but a large gain on NIST'20 random split (39.0 to 52.5), which the authors attribute to in-distribution structure.
• Ablation shows removing external RDKit descriptors from fragment tokens does not hurt test loss, and training without MAGMa supervision fails to converge.

Limitations

• Not fully end-to-end: the fragment generation step truncates gradients and relies on MAGMa heuristic supervision.
• Does not handle complex rearrangements that change heavy-atom connectivity.
• Intensity predictor and breakpoints limited to k=3 breakpoints per fragment.

Experimental setup

Datasets: NIST'20 MassSpecGym

Benchmarks: MassSpecGym NIST'20

Metrics: top-k retrieval accuracy (Recall@k / Top-k) cosine similarity entropy (Jensen-Shannon) similarity coverage inference time (ms/spec)

Baselines: ICEBERG 2.0 MolSpecFlow FraGNNet MARASON MassFormer GrAFF-MS NEIMS FIORA

Builds on: DETR Graphormer ICEBERG MAGMa LinSATNet

Closest papers by meaning

• MassSpecGym: A benchmark for the discovery and identification of molecules (NeurIPS 2024 · similarity 0.89)
• MS-BART: Unified Modeling of Mass Spectra and Molecules for Structure Elucidation (NeurIPS 2025 · similarity 0.88)
• PepSpecBench: A Unified Evaluation Benchmark for Peptide Tandem Mass Spectrometry Prediction (NeurIPS 2026 · similarity 0.87)
• DiffMS: Diffusion Generation of Molecules Conditioned on Mass Spectra (ICML 2025 · similarity 0.86)
• Neural Graph Matching Improves Retrieval Augmented Generation in Molecular Machine Learning (ICML 2025 · similarity 0.85)
• Unraveling Molecular Structure: A Multimodal Spectroscopic Dataset for Chemistry (NeurIPS 2024 · similarity 0.85)