Spectral foundation models (DreaMS, PRISM), the MassSpecGym benchmark and its audit, CASMI 2022 — plus a likelihood-ratio test of "Bayes vs rank-average"

TL;DR. DreaMS (116 M-parameter transformer, pre-trained on up to 700 M GNPS spectra, MIT weights on Hugging Face) is the only open spectral foundation model; Enveda's PRISM (1.2 B spectra, half proprietary) has a blog post, no weights, no code and — as far as I could verify — no preprint. MassSpecGym is the leaderboard that ranks everyone, but its June-2026 audit found evaluation problems in at least 17 of 26 papers, with "state-of-the-art" retrieval numbers collapsing from 82 % to 7 % hit@1 once a SMILES-canonicalisation shortcut was removed; honest formula-known retrieval today tops out around 54 % hit@1 and formula-free around 27–28 %. A simulation calibrated to those numbers says: never rank-average two scorers; sum calibrated log-likelihood-ratios instead (+0.03–0.06 MRR@25), weight them by predicted compound class (+0.02–0.03 more), and expect every fusion rule except a jointly-trained one to stop helping once scorer errors are ≥ 60 % correlated.

Part A — Facts

A1. DreaMS

A2. PRISM (Enveda + Microsoft)

What is public is a single Enveda blog post (2024-05-28; authors Healey, Domingo-Fernández, Taylor, Krettler, Lightheart, Park, Kind, Allen, Colluru) plus press coverage. Claims: 1.2 B MS/MS spectra, 85 B tokens; ≈ 600 M from public repositories (GNPS, MetaboLights, Metabolomics Workbench) + ≈ 600 M from Enveda's own platform; BERT-family encoder; "masked peak modeling" masking 20 % of peak masses; downstream: 7–16 % relative R² improvement across 12 chemical properties vs a baseline, and "23 % relative improvement" in spectral library matching vs "standard ML models". Parameter count, layer count, tokeniser, benchmarks used and baselines are not disclosed. I found no weights, no code and no preprint (the search budget ran out before I could exhaust bioRxiv; treat "no preprint" as unverified, confidence low). Note that Rose Lightheart (PRISM author) is a co-author of the MassSpecGym audit (Liu, Bushuiev, Lightheart et al.) — the challenge host knows the shortcut literature intimately, which is consistent with the RDKit tautomer-canonicalisation step in the Kaggle metric.

A3. MassSpecGym (NeurIPS 2024 Datasets & Benchmarks; arXiv 2410.23326 v1 2024-10-30, v3 2025-02-14)

Dataset. 231,104 spectra (231 k), ≈ 29 k unique structures (≈ 10 k, 33 %, newly measured in-house); sources MoNA, MassBank, GNPS, matchms-filtered; adducts only [M+H]+ and [M+Na]+; Orbitrap or QTOF; m/z < 1,000; spectra with > 50 % unexplained intensity removed. HF roman-bushuiev/MassSpecGym, MIT, 52.8 GB. Split. Agglomerative clustering of the 29 k molecules on MCES distance, single linkage, threshold 10 — no two molecules in different folds are < 10 edge edits apart; folds stratified by instrument, collision energy, adduct, molecule frequency. The paper shows that the usual InChIKey-disjoint split leaks Tanimoto > 0.85 analogs. Retrieval candidates. |C| ≤ 256 per spectrum, sampled with the same mass ("within an acceptable experimental error range" — tolerance not given in the main text) first from 1 M bio/environmental molecules, then 4 M diverse, then all 118 M PubChem (2024-05-31 download). Bonus track: same formula instead of same mass. Metrics: hit@1/5/20, MCES@1, 99.9 % bootstrap CIs (20,000 resamples).

Original baselines (arXiv v3 Table 3; %):

De novo (Table 2): every baseline (SMILES/SELFIES transformer, random generation) scores 0.00 % top-1 and top-10 exact accuracy; best MCES@top-10 was random generation (18.25 formula track). Simulation (Table 4): FraGNNet cosine 0.52, forward-model retrieval hit@1 46.64 % (mass track).

The audit — "MassSpecGym in the Wild: Uncovering and Correcting Evaluation Pitfalls in AI-Driven Molecule Discovery" (Liu, Bushuiev, Lightheart et al., arXiv 2606.19624, 2026-06-17). 34 papers in the first year; 26 report benchmark numbers; ≥ 17 of those 26 have evaluation issues in three classes: (i) data leakage (decoder pre-training sets excluding only exact InChIKeys, not MCES ≥ 10 analogs; ICEBERG simulators trained on test molecules; formula annotators trained on test), (ii) shortcut learning from candidate artefacts, (iii) implementation bugs and metric divergence.

Corrected / updated retrieval leaderboard (audit Table 3; formula-known track, |C| ≤ 256, %):

(Column alignment of the MCES values for FRIGID/vFRIGID in the PDF text is ambiguous; the 5.16/6.28 pair belongs to the two FRIGID rows, 6.61/8.73 to ICEBERG/MolForge — treat MCES assignments among those four rows as ± one row.)

Mass-based (formula-free) track after the audit (Table 2, 8; data-safe MIST retrained per FRIGID recipe): Random 0.43 %; MIST with data-safe MIST-CF top-1 formula 26.64 % hit@1 (38.47 / 54.90 @5/@20, MCES 16.41); with MIST-CF top-5 element-wise-max 28.23 %; with test-contaminated MIST-CF 31.39–35.15 % (the leak); with ground-truth formula 51.45 % (66.25 / 80.58). MIST-CF itself: 50.5 % top-1 / 73.2 % top-5 formula accuracy when clean, 66.5 / 84 % when trained on test. Simply filtering mass-candidates to the true formula lifts random from 0.43 % to 12.98 % hit@1 — a 30× likelihood ratio from the formula alone.

De novo (audit Table 9, formula track; top-1 acc / top-10 acc): MIST+MSNovelist 0 / 0; Spec2Mol 0 / 0; MIST+Neuraldecipher 0 / 0; MS-BART 1.07 / 1.11; MADGEN 1.31 / 1.54; DiffMS 2.30 / 4.25 (MCES 18.45 → audit prints 27.47 under its pinned MCES; see Table 10); FOAM 1.50 / 10.28; MIST+MolForge corrected 10.73 / 14.48 (inflated 31.75 / 40.55 from the batching bug); FRIGID 18.29 / 22.00. Metric divergence (Table 10, identical DiffMS predictions): Tanimoto 0.297 (Morgan r2, 1024 b) vs 0.214 (r3, 4096 b); MCES 21.21 vs 15.87 depending on solver bound. v1.5 pins InChIKey14, 2048-bit ECFP4, MCES threshold 15.

Unaudited claims to treat with suspicion: GLMR (Zhang et al., arXiv 2511.06259, Nov 2025) reports 64.17 % recall@1 mass-track / 68.48 % formula-track on MassSpecGym, with re-run MIST at 18.46 % and JESTR 17.62 % (mass track); the text never mentions candidate canonicalisation and the paper predates v1.5 — it is absent from the audited leaderboard. MARLIN (arXiv 2607.04774, July 2026), MS-GPT (2607.23607), SpecBridge (2601.17204), FlowMS (2603.18397) report on NPLIB1 or MassSpecGym with numbers I could not open; unverified. The live leaderboard massspecgym.onrender.com returned HTTP 503 on three attempts.

A4. CASMI 2022 (Fiehn lab, UC Davis)

500 compounds: 250 "priority" (not in public spectral libraries) + 250 "bonus"; raw LC-HRMS/MS in +ESI and −ESI (each challenge in one polarity only); inclusion list, 0.8 s dynamic exclusion; MS1, MS2 and RT provided; top-1 answer only, scored as exact InChIKey first-block match; deadline 2022-06-06; 16 groups. Instrument vendor/type not stated on the pages I could read (unverified).

2D-structure results (priority / bonus; correct / attempted):

Formula: Duhrkop 94 % (priority) / 95 % (bonus), Nothias 93 %; class: Duhrkop 69 % priority, Wang 77 % bonus; adducts: 93–96 % for the best. Tools per participant beyond Duhrkop are not stated on the results page (MetFrag and CFM-ID entries are not identifiable there; unverified). Take-away: on "not-in-any-library" compounds the best 2022 top-1 rate was ≈ 26 % — roughly where the 2026 public-LB MRR@25 of ≈ 0.42–0.47 sits once you account for partial credit at ranks 2–25.

Part B — Forced analogy: medical differential diagnosis with likelihood ratios

The analogy, precisely. A clinician starts from a pre-test probability (prevalence given the presentation), then multiplies the odds by the likelihood ratio of each test result, LR = P(result | disease) / P(result | no disease), and only multiplies when tests are conditionally independent given the disease. Mapping: a candidate structure = a diagnosis; the prior = 1/|C| or a database-frequency prior; each scorer (fingerprint cosine, spectral-library similarity, forward-simulation match, formula agreement) = a test; a scorer's output for a candidate = a test result; LR = how much more often that score arises for the true structure than for a decoy. Rank-averaging corresponds to the clinically discredited habit of "counting positive tests".

(i) Crude per-class sensitivities and LRs from published numbers. The only per-class, two-method table I could reach was none (MIST's bioRxiv full text returned HTTP 429 on every attempt; the Nature version is paywalled to the fetcher; paperclip is not installed on this box). So I calibrated the marginal sensitivity of two real methods on identical candidate sets and modelled class-dependence as an explicit assumption: MassSpecGym formula track, |C| ≤ 256 — MIST 9.57 / 22.11 / 41.12 % and JESTR 11.82 / 33.48 / 61.46 % hit@1/5/20. Fitting an equal-variance Gaussian score model (true ~ N(μ,1), decoys ~ N(0,1)) gives μ_MIST = 1.30 and μ_JESTR = 1.60. Then three latent "classes" with priors 0.5/0.3/0.2 and multiplicative signal factors (A: 1.35/1.0/0.55; B: 0.6/1.0/1.45) — i.e. A is good where B is weak — assumed, not measured. At 1,000 candidates the implied sensitivities P(true in top-25 | class) are A: 0.415 / 0.251 / 0.105 and B: 0.157 / 0.355 / 0.637; against a random top-25 rate of 2.5 % these are LRs of 16.6 / 10.1 / 4.2 (A) and 6.3 / 14.2 / 25.5 (B). A genuinely measured LR from the audit: a correct formula filter lifts random hit@1 from 0.43 % to 12.98 %, LR ≈ 30.

(ii) Simulation. 400 molecules × 1,000 candidates, two scorers with shared noise fraction ρ, calibration fitted on a separate 400-molecule simulated "training" set (logistic regression of P(true | score), global or per class), 5 seeds. Code (lr_sim.py, numpy + scipy only):

# score model: decoys ~ N(0,1); true ~ N(mu_scorer[class],1); shared noise fraction rho
def simulate(n_mol, n_cand, mu_A, mu_B, class_prior, rho, rng):
    cls = rng.choice(len(class_prior), size=n_mol, p=class_prior)
    shared = rng.normal(0, 1, (n_mol, n_cand))
    eA = np.sqrt(rho)*shared + np.sqrt(1-rho)*rng.normal(0, 1, (n_mol, n_cand))
    eB = np.sqrt(rho)*shared + np.sqrt(1-rho)*rng.normal(0, 1, (n_mol, n_cand))
    sA, sB = eA.copy(), eB.copy()
    sA[:, 0] += mu_A[cls]; sB[:, 0] += mu_B[cls]       # candidate 0 is the truth
    return cls, sA, sB

def mrr_at_25(score):
    rank = 1 + np.sum(score[:, 1:] >= score[:, [0]], axis=1)   # pessimistic ties
    return np.mean(np.where(rank <= 25, 1.0/rank, 0.0))

def rank_avg(sA, sB):      # lower mean rank = better
    rA = (-sA).argsort(1).argsort(1); rB = (-sB).argsort(1).argsort(1); return -(rA+rB)/2
def z_avg(sA, sB):
    z = lambda s: (s - s.mean(1, keepdims=True)) / s.std(1, keepdims=True); return z(sA)+z(sB)
def naive_bayes(sA, sB, cls, models, per_class, joint=False):
    # models[c] = (wA, wB, wAB): logistic fits of P(true|score) on a held-out simulated set.
    # logit = log LR + log prior-odds; the prior is constant within a query, so sum of logits ranks
    out = np.empty_like(sA)
    for i in range(len(cls)):
        wA, wB, wAB = models[cls[i] if per_class else None]
        out[i] = (wAB[0] + wAB[1]*sA[i] + wAB[2]*sB[i]) if joint \
                 else (wA[0] + wA[1]*sA[i]) + (wB[0] + wB[1]*sB[i])
    return out
# mu fitted so that hit@1/hit@20 at |C|=256 match MIST (9.57/41.12 %) and JESTR (11.82/61.46 %);
# classes: priors (0.5,0.3,0.2), factors A=(1.35,1.0,0.55), B=(0.6,1.0,1.45)

(iii) Results — MRR@25, mean of 5 seeds (SD ≤ 0.012 everywhere):

* At ρ = 0.9 the joint model exploits the difference A − B, whose noise variance is 2(1 − ρ) = 0.2 while its signal μ_A − μ_B stays class-specific; that "common-mode rejection" is a property of my Gaussian shared-noise model, not something to expect from two real spectral scorers. Discount it.

Verdict. Bayesian (calibrated log-LR) combination beats rank averaging by 0.03–0.05 MRR@25 at every correlation, and class-weighting the LRs adds another 0.02–0.03; rank-averaging two correlated scorers is worse than the better scorer alone once ρ ≥ 0.3, and every independence-assuming combiner (rank, z, naive Bayes) stops beating the single best scorer around ρ ≈ 0.6. Z-averaging and global naive Bayes are indistinguishable here because, for equal-variance Gaussian scores, log LR is linear in the score — so the honest message is "calibrate, then add log-odds", not "Bayes is magic". The gain of class-weighting survives a 20 % class-misassignment rate (0.229 vs 0.213 at ρ = 0). Absolute values are lower than published hit rates because 1,000 candidates were used, not 256.

Does the analogy hold? Partially. It holds for the core prescription (prior × LR, not vote counting), for the role of the formula/mass filter as a high-LR "screening test" (LR ≈ 30 measured), and for class-conditional calibration (a test's LR depends on the patient subgroup = compound class). It breaks on conditional independence: fingerprint-cosine scorers trained on the same libraries share failure modes (same missing substructures, same adduct confusions), exactly the ρ ≥ 0.6 regime where the naive product over-counts evidence. It also breaks on the prior: clinicians have measured prevalences; here the natural prior (1/|C|) ignores that PubChem decoys are not drawn from "chemistry that gets measured" — the audit's spectrum-blind PubChem-order baseline reaching 49.98 % hit@1 is a prior, not a test, and should enter the fusion as log prior-odds, not as another LR.

Gotchas — what could silently fool us

1. Canonicalisation leak in any locally built candidate set. If your true-structure SMILES come from the training parquet (normalized_smiles) and decoys from COCONUT/PubChem canonicalised differently, a learned re-ranker will reach 80–99 % locally and ≈ 7 % on the hidden test (audit Table 3). Canonicalise everything with the same RDKit version and tautomer routine as the metric (RDKit 2026.03.3).
2. Formula leakage dressed as "mass-based". Any formula annotator (SIRIUS, MIST-CF, BUDDY) trained or tuned on molecules that appear in your validation set inflates hit@1 by 9 points absolute (26.6 → 35.2 %); a correct formula alone is a 30× LR. The hidden test molecules are "none previously published" — a formula model trained on all public libraries is fine for the test but will lie on any validation set drawn from those libraries.
3. Spectrum-blind priors masquerade as skill. PubChem default order gives 50 % hit@1 on MassSpecGym; number of chiral centres gives 4.9 %. On a COCONUT-sourced pool, "is in COCONUT at all", NP-likeness, or deposition frequency will carry similar spurious signal — fine to use as a prior, dangerous to mistake for a spectral model.
4. Batched-inference masks. The MIST/DiffMS attention mask added +1 instead of excluding padding; a reused checkpoint tripled de novo accuracy (10.7 → 31.8 %). Check that batched and single-spectrum outputs agree before trusting any borrowed encoder.
5. Metric parameterisation. Tanimoto moves 0.30 → 0.21 across Morgan radius/bit choices and MCES 21.2 → 15.9 across solver bounds on the same predictions. The Kaggle metric is InChIKey14 after tautomer canonicalisation — a prediction that differs only by a tautomer is correct, one that differs by a stereocentre is also correct, one that differs by a double-bond position is wrong. Local MRR must use the same routine.
6. "Formula track" is harder than "mass track" for fingerprint methods. MIST drops from 14.64 % (mass candidates) to 9.57 % (formula candidates) because same-formula isomers have near-identical fingerprints. A COCONUT pool with a tight mass window is effectively a formula-track problem for most queries.
7. Collision energies missing. Roughly half of the MassSpecGym test split lacks CE; ICEBERG's numbers required imputing them. The hidden test gives CE lists (merged multi-CE spectra); models trained on single-CE library spectra see a distribution shift.
8. Foundation-model benchmarks are mostly on Orbitrap. 97 % of GeMS-A is Orbitrap; the hidden test is all timsTOF. DreaMS's Fourier m/z features assume ~4-decimal accuracy; QTOF spectra are in GeMS-C only.
9. The leaderboard you cite may be the uncorrected one. Published papers (GLMR 64 %, FRIGID 45 %, MIST vFRIGID 54 %) sit above the audited clean frontier (FLARE 22.7 %, ICEBERG 36.2 %, formula track). The clean formula-free number to beat is MIST + data-safe MIST-CF ≈ 27–28 % hit@1 at |C| ≤ 256.
10. CASMI 2022 was top-1 only, one polarity per challenge, with RT and MS1 — its 26 % is not an MRR@25 and was obtained with information the 2026 challenge does not give (RT) and without the information it does (multi-CE merged spectra).
11. Simulation caveats. Equal-variance Gaussian scores, a single shared-noise ρ for truth and decoys, three fictitious classes, calibration on simulated rather than real data; class factors are assumptions. The ordering of methods is robust across the three scenarios; the magnitudes are not to be quoted as predictions.

Implications for the challenge

1. Fuse scorers by summing calibrated log-odds, never by averaging ranks. Fit per-scorer P(true | score) with logistic/isotonic regression on an MCES-disjoint, novel-molecule validation set; rank by Σ logit. Expected +0.03–0.05 MRR@25 over rank averaging in simulation.
2. Add a class-conditional calibration layer. Predict NPClassifier/ClassyFire superclass from the spectrum (CANOPUS-style head or DreaMS fine-tune, 80 % accuracy suffices) and fit calibration per class; worth ≈ +0.02–0.03 in simulation and it is exactly where "scorer A is good on terpenoids, B on alkaloids" lives.
3. Measure ρ between your scorers on validation (correlation of per-query true-candidate z-scores or of rank errors). If ρ ≥ 0.6 for two fingerprint models, drop one or replace the product with a jointly trained 2-feature logistic re-ranker; seek genuinely orthogonal evidence (forward-simulated spectrum match, spectral-library analog similarity, formula/adduct consistency, isotope pattern) rather than a third fingerprint.
4. Treat formula/mass filtering as the highest-LR test and get it right first. Use a formula annotator trained on all public data (no leakage risk against unpublished test molecules), take its top-k formulas with calibrated probabilities, and feed them as log prior-odds over candidates, not as a hard filter — 51 % vs 27 % hit@1 is the stake.
5. Use DreaMS as a feature extractor, not as a solved component: embedding_model_massspecgym.ckpt or the v1.5 data-safe DreaMS, 1,024-d embeddings, single-GPU; fine-tune to fingerprints with the Murcko-histogram split protocol. Its MassSpecGym retrieval number (6.7 % corrected, formula track) says the published alignment recipe is not a shortcut to a strong ranker.
6. Validate exactly like the audit's v1.5: uniform RDKit canonicalisation of truth and decoys, MCES ≥ 10 (or at least Tanimoto < 0.7) between train and validation molecules, pinned metric code, and a spectrum-blind baseline (database order, NP-likeness, chiral-centre count) reported next to every model — if the blind baseline is within a few points of the model, the model has learned the pool, not the spectrum.
7. Do not benchmark against PRISM: nothing is reproducible; its only transferable lessons are "masked-peak pre-training works at 20 % masking" and that Enveda's half of the corpus is proprietary — expect the hidden test to resemble enveda-180 acquisition conditions more than GNPS.
8. Set expectations with the audited numbers: clean formula-free hit@1 ≈ 27 % at 256 candidates; at a COCONUT-scale pool with 10 ppm windows the public LB's ≈ 0.42–0.47 MRR@25 is consistent with that, so headroom is in candidate coverage (recall@25) and calibrated fusion, not in a bigger single encoder.

Open questions

• MIST paper per-class (NPClassifier) MIST-vs-CSI:FingerID retrieval tables (Fig. 2/3 and SI) — needed to replace my assumed class factors with measured ones; bioRxiv 10.1101/2022.12.30.522318 and Nature 10.1038/s42256-023-00708-3 could not be opened here.
• Whether a PRISM preprint exists (search budget exhausted); whether PRISM's "library matching +23 %" was measured on timsTOF data.
• MassSpecGym candidate mass tolerance (ppm/mDa) — stated only in the SI; and whether v1.5 candidate JSONs change |C|.
• Whether FRIGID/MIST-vFRIGID's "few leaked test structures" flag was resolved in a later revision; the live leaderboard (503) would show it.
• CASMI 2022 instrument and per-participant tools beyond Duhrkop (results page lists names only).
• Real ρ between fingerprint-cosine, DreaMS-similarity and forward-simulation scorers on timsTOF natural products — the single number that decides whether naive LR fusion or a joint re-ranker is right.

Sources

• Bushuiev R et al. Self-supervised learning of molecular representations from millions of tandem mass spectra using DreaMS. Nat Biotechnol 44:630–640 (2026, online 2025-05-23). https://doi.org/10.1038/s41587-025-02663-3 (PDF read; Europe PMC record PMC13090125 https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=DOI:%2210.1038/s41587-025-02663-3%22&format=json)
• DreaMS code/README: https://raw.githubusercontent.com/pluskal-lab/DreaMS/main/README.md ; docs https://dreams-docs.readthedocs.io/en/latest/ ; weights https://huggingface.co/roman-bushuiev/DreaMS/tree/main ; GeMS https://huggingface.co/datasets/roman-bushuiev/GeMS ; Zenodo 10997887
• Enveda. PRISM: A foundation model for life's chemistry (2024-05-28). https://www.enveda.com/prism-a-foundation-model-for-lifes-chemistry/
• Bushuiev R et al. MassSpecGym: A benchmark for the discovery and identification of molecules. arXiv:2410.23326 (v3, 2025-02-14), NeurIPS 2024 D&B. https://arxiv.org/abs/2410.23326 (HTML v3 and PDF read); README https://raw.githubusercontent.com/pluskal-lab/MassSpecGym/main/README.md ; dataset https://huggingface.co/datasets/roman-bushuiev/MassSpecGym
• Liu, Bushuiev, Lightheart et al. MassSpecGym in the Wild: Uncovering and Correcting Evaluation Pitfalls in AI-Driven Molecule Discovery. arXiv:2606.19624 (2026-06-17). https://arxiv.org/abs/2606.19624 (HTML and PDF read)
• Bohde M et al. DiffMS: Diffusion Generation of Molecules Conditioned on Mass Spectra. arXiv:2502.09571 (ICML 2025). https://arxiv.org/html/2502.09571v2
• Zhang Y et al. Breaking the Modality Barrier: Generative Modeling for Accurate Molecule Retrieval from Mass Spectra (GLMR). arXiv:2511.06259 (2025-11-09). https://arxiv.org/abs/2511.06259
• Che X, Du X, Xu D. MARLIN. arXiv:2607.04774 (2026-07-06). https://arxiv.org/abs/2607.04774 (abstract only)
• De Waele G et al. Small molecule retrieval from tandem mass spectrometry: what are we optimizing for? arXiv:2602.16507 (2026-02-18). https://arxiv.org/abs/2602.16507 (abstract only)
• Goldman S et al. Annotating metabolite mass spectra with domain-inspired chemical formula transformers. Nat Mach Intell (2023). https://doi.org/10.1038/s42256-023-00708-3 (abstract/summary only — full text not accessible); code README https://raw.githubusercontent.com/samgoldman97/mist/master/README.md
• Fiehn Lab CASMI 2022: https://fiehnlab.ucdavis.edu/casmi ; results https://fiehnlab.ucdavis.edu/casmi/casmi-2022-results ; FAQ https://fiehnlab.ucdavis.edu/casmi/casmi-faq
• Simulation code and raw output: lr_sim.py, lr_sim_results.json (session scratchpad; code reproduced above; numpy 2.5.3, scipy 1.18.1).
• Not reachable in this session: massspecgym.onrender.com (HTTP 503 ×3), ChemRxiv DreaMS preprint page (403), bioRxiv MIST full text (429 ×4), casmi-contest.org (TLS certificate mismatch).