arXiv shortlist: MS/MS to 2D structure for Enveda CASMI 2026
Checked 2026-09-23. Discovery/abstract-level screening, not a full replication or license audit. Priorities below are our project-specific judgement, not paper claims of CASMI performance.
Immediate ranking and validation
Small molecule retrieval from tandem mass spectrometry: what are we optimizing for? (2026-02), https://arxiv.org/abs/2602.16507. Fingerprint prediction quality and candidate retrieval can trade off. FP-01/02: compare BCE plus fingerprint-logit scoring with candidate-ranking objectives on identical pools; choose by exact molecule MRR@25.
JESTR (2024-11), https://arxiv.org/abs/2411.14464. Joint spectrum/molecule embeddings and contrastive training; ranks candidate structures without a fingerprint reconstruction requirement. Challenger to FP-01, not a replacement for candidate coverage.
MassSpecGym in the Wild (2026-06), https://arxiv.org/abs/2606.19624. Reports evaluation issues in at least 17/26 surveyed papers, involving leakage, shortcuts and implementation/metric divergence. Relevant to AUDIT-01 and VAL-01/02; no claim here that every shortlisted paper is implicated.
MassSpecGym benchmark (2024-10), https://arxiv.org/abs/2410.23326. Retrieval, de novo generation and forward simulation with generalization-oriented splits. Its benchmark metrics do not replace CASMI scorer validation.
De novo generation
MS-BART (2025-10), https://arxiv.org/abs/2510.20615. Fingerprint/molecule pretraining and transfer using MIST predictions. Candidate for GEN-03: learn structure decoding from structure-only pairs, then adapt to predicted fingerprints.
MARLIN (2026-07), https://arxiv.org/abs/2607.04774. Predicted fingerprints plus precursor mass condition block-diffusion generation without a supplied formula. Promising formula-free GEN-03 research direction; resources, checkpoints, rights and pretraining exposure not yet verified.
DiffMS (2025-02), https://arxiv.org/abs/2502.09571. Formula-restricted graph diffusion with fingerprint/structure pretraining. Known heavy-atom composition is a central assumption; CASMI application requires an eligible predicted-formula route and accounting for its misses.
MADGEN (2025-01), https://arxiv.org/abs/2501.01950. Retrieve scaffold then spectrum-conditioned generation. Evaluate predicted scaffold retrieval, not oracle scaffold results; complementary to analogue work.
MSFlow (2026-02), https://arxiv.org/abs/2602.19912. Formula-restricted spectrum encoder plus flow-matching decoder. Abstract explicitly describes trained version for non-commercial users; do not approve assets for competition use without separate rights verification.
Forward reranking
GLACIER (2026-06), https://arxiv.org/abs/2606.29161. Molecular graph fragment detection to predict spectra. RERANK-01 candidate: score a fixed high-recall shortlist via forward-spectrum agreement. Not direct de novo identification.
Use molecule-grouped OOF outputs, scorer-compatible identity deduplication, full-cohort and in-pool MRR@25, candidate retention and runtime measurements. Separate unavailable structure, formula/scaffold prediction failure and ranking failure. Published benchmark percentages are not expected CASMI scores. No training, kernel launch, submission or TODO completion is implied by this literature screen.