← Back
📝 Notes / research / jestr-prep-20260924

📑 Contents

JESTR-PREP — independent spectrum/structure retrieval challengerHypothesisFixed comparison contractRepresentation contract (to be pinned before training)Spectrum encoderStructure encoderObjective ablationsAcceptance gatesOpen questions before trainingExplicit non-claims

JESTR-PREP — independent spectrum/structure retrieval challenger


Status: preparation contract only; no paper-code, checkpoint, tokenizer, pretrained representation or dataset reuse. Full-paper method extraction remains pending because the document extractor was unavailable on 2026-09-24.


Hypothesis


A jointly trained spectrum/structure embedding objective may improve candidate ordering over FP-01 fingerprint-logit scoring while using the same candidate pool and molecule-grouped exposure policy. It does not solve candidate recall: report pool coverage and candidate retention separately from ranking.


Fixed comparison contract


  • Identity: scorer-compatible scorer_identity14, tautomer-canonicalized InChIKey14, RDKit 2026.03.3.
  • Pool: competition-training structure universe only; PubChem and external pretrained representations excluded.
  • Split: manifests/rebuild-20260917/molecule-grouped-5fold-scorer.json; development folds 1–3; fold 0 diagnostic only; fold 4 reserved stress/repeat, not an untouched claim.
  • Query unit: molecule-level aggregation over permitted spectra. Preserve adduct, polarity, instrument, collision-energy value/missingness and precursor mass; do not merge incompatible acquisitions into one synthetic metadata row.
  • Positive pair: spectrum view and its scorer identity's structure representation. Multiple spectra of one identity are grouped for exposure and scored at molecule level.
  • Negatives: in-batch and deterministic hard negatives selected by declared mass/formula compatibility, excluding scorer-equivalent positives and all held-out identities. Every negative pool gets a content hash.
  • Candidate score: cosine similarity (or temperature-scaled dot product) between query and candidate embeddings, deterministic identity tie-break. Score the full permitted pool before top-K reporting.
  • Metrics: full-cohort MRR@25, Recall@25, Top-1, Top-5, candidate Recall@100/1000, pool coverage, covered-pool conditional MRR, valid/usable query rate, runtime and RAM/VRAM. Preserve per-molecule ranks and candidate membership.

    • Representation contract (to be pinned before training)


    Spectrum encoder


    Independent implementation. Input is a set/sequence of cleaned MS/MS peaks plus observed metadata. The implementation must record peak cleaning, max peaks, m/z/intensity normalization, precursor-mass/adduct encoding, metadata missingness and pooling. No target-derived precursor error or hidden formula may enter the query representation.


    Structure encoder


    Independent implementation from competition-eligible structures. The exact graph atom/bond features, chirality policy, pooling and parameter count are not yet fixed; they must be selected on development folds and written into the run manifest before training. Scorer identity is for grouping/scoring, not a substitute for graph representation.


    Objective ablations


    1. Spectrum-to-fingerprint FP-01 control.

    2. Spectrum/structure contrastive loss with in-batch negatives.

    3. Same objective plus mass/formula hard-negative term, with scorer-equivalent exclusion.

    4. Optional structure-encoder freeze/control only if the full paper method review justifies it.


    Do not call a loss “JESTR” until the full paper method is checked and the implemented differences are recorded. The name here means a paper-inspired joint-embedding challenger.


    Acceptance gates


    1. Static schema/mask fixtures pass without Torch or GPU.

    2. Positive/negative masks are molecule-grouped and scorer-identity safe.

    3. Candidate pool and fold hashes are frozen.

    4. A bounded GPU run completes one train step, one validation batch and writes an inspectable checkpoint.

    5. The trained challenger is compared with FP-01 and the resolved retrieval control on exactly the same molecule cohort/pool.

    6. Promotion requires paired molecule-level improvement, no severe intended-regime regression, repeatability evidence and an offline package pass.


    Open questions before training


  • Exact paper architecture and objective terms, after full-method extraction.
  • Structure representation and whether any paper pretraining is used; default is no external pretraining.
  • Temperature, batch size, hard-negative definition and compute budget.
  • Whether candidate embeddings can be cached without changing pool-relative scoring.
  • Rights/provenance of every library dependency and data-derived artifact.

    • Explicit non-claims


    This preparation document reports no paper reproduction, no trained result, no CASMI score and no evidence that joint embeddings outperform fingerprint logits.