LIT-02 — primary-method extraction notes (2026-09-24)
Scope
Primary arXiv HTML methods were inspected for JESTR (2411.14464), retrieval objectives (2602.16507), MassSpecGym (2410.23326), and MassSpecGym in the Wild (2606.19624). These are technical references only. No paper code, checkpoints, tokenizers, pretrained representations, or external datasets were adopted.
JESTR (2411.14464)
JESTR learns spectrum and molecular encoders into a shared embedding space and ranks candidates by cosine similarity. The spectrum representation bins peaks into 1-Da bins below 1000 m/z, sums intensities per bin, normalizes intensity to 999, and applies one-third log10 transformation before a 3-layer MLP. The molecule encoder is a graph GCN/MLP using atom and bond features. Training uses temperature-scaled InfoNCE/CMC contrastive loss over matching spectrum–molecule pairs and in-batch negatives. A second regularization loss pushes the spectrum embedding away from hard candidate molecules sharing the target formula; candidates are ordered by Tanimoto similarity. The paper reports adding this regularizer only in the final 3% of training, weighted 0.1 against 0.9 contrastive loss.
Competition adaptation: retain the shared-space and late hard-negative ideas as an independently implemented challenger, but do not copy the exact 1-Da representation without a domain review. Preserve adduct, polarity, instrument, collision energy, precursor mass and missingness; use the competition-only pool. Same-identity and held-out-target exclusions override the paper's external candidate construction.
Retrieval objectives (2602.16507)
The paper distinguishes bitwise losses (BCE/focal), vectorwise losses (cosine/IoU/Tanimoto), and listwise contrastive losses over the candidate set. Its central result is that fingerprint similarity and molecule retrieval are different objectives; optimizing fingerprint accuracy can worsen retrieval. Candidate-set similarity structure controls the trade-off. For this competition, final comparison must use molecule-level MRR@25 and paired ranks, not fingerprint similarity alone. FP-02 should compare BCE-only against a candidate-aware/listwise or hard-negative objective on identical pools and cohorts, with scorer-equivalent false negatives removed.
MassSpecGym (2410.23326)
The benchmark defines retrieval, de novo generation, and spectrum simulation tasks, with molecule-disjoint and structure-distance-aware evaluation intended to reduce leakage. It emphasizes standardized preprocessing, candidate sets, and task-specific metrics. Relevance here is methodological: freeze identity grouping, candidate membership, split hashes, and molecule-level ranking metrics before training; do not treat a visible test fixture as a generalization result.
MassSpecGym in the Wild (2606.19624)
The audit identifies three recurring failure classes: data leakage, shortcut learning, and implementation/metric divergence. It specifically warns that exact-structure splits can still leave local chemical-space exposure; candidate composition, SMILES canonicalization, and ranking biases can enable spectrum-independent shortcuts; and shared metric/encoder bugs can propagate. Competition adaptation: retain the existing exposure ledger, scorer identity14 binding, candidate-absence strata, and no-clean-untouched-claim guard. Audit every learned route's training exclusions and compare against spectrum-independent controls.
Decision
LIT-02 is complete at the method-extraction level. JESTR-PREP remains design-only; no training is authorized by this note. FP-02 is the next measured comparison after GPU execution access returns. Until the recorded Kaggle quota refresh, continue only safe local validation/documentation and do not retry remote launches.