{"id":"93c9410b-d0c1-49a7-b5aa-7c920375c25d","arxiv_id":"2607.23607","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"Querying a fingerprint-conditioned molecule-language model with a calibrated band of spectrum-induced posteriors beats point-threshold fingerprint decoding on NPLIB1 and MassSpecGym.","lead":"MS-GPT turns noisy mass-spec fingerprint predictions into a band of queries for a pretrained molecule generator, instead of one hard threshold. It sets new exact-match records on two public de novo structure benchmarks and ships code.","discovery_kind":"new_method","skeptic_critique":null,"referee_report":{"model":"moonshotai/kimi-k3","summary":"The paper addresses the training–inference mismatch in fingerprint-mediated MS/MS de novo structure elucidation: decoders are pretrained on oracle Morgan fingerprints but queried at inference with a single thresholded spectrum-induced posterior. MS-GPT recasts decoding as posterior querying of a fingerprint–formula-conditioned molecule-language model (a cross-attention-augmented SAFE-GPT, pretrained on ~100M molecules). Three mechanisms are proposed: (i) active-bit density band calibration, which maps a density-ratio interval K=[0.95,1.50] (matched to encoder-training-set active-bit density) to a threshold band; (ii) fixed-size group querying, which spreads a candidate pool B=M×N over M=20 interpolated thresholds; and (iii) posterior-aligned adaptation, a rank-4 LoRA on the query-reading pathway trained with recoverability-grouped CE+KL losses. On NPLIB1 and MassSpecGym under the known-formula protocol, MS-GPT reports Top-1/Top-10 exact match of 29.76%/41.07% and 23.91%/28.65%, exceeding FRIGID by 4–8 points. Ablations (Tables 2–3), posterior-fidelity and pretraining-proximity stratifications, allocation and pool-scaling sweeps, pretraining-data scaling, and failure-mode case studies support the claims.","tokens_in":23932,"tokens_out":4441,"duration_ms":172105,"significance":"If the results hold, this is a useful contribution to an actively contested benchmark area. The exact-match gains over the strongest diffusion baseline (FRIGID) are large (+4.7/+7.7 points Top-1/Top-10 on NPLIB1; +5.6/+6.7 on MassSpecGym) and the method is inference-cheap. Notable strengths that weigh positively: the pretraining corpus explicitly excludes connectivity-equivalent (InChIKey-14) validation/test structures, addressing the most obvious leakage channel; source code and checkpoints are released, making the pipeline reproducible; the ablations cleanly separate the three components across two benchmarks; the stratified analyses (posterior fidelity, pretraining proximity) directly test and largely rule out the memorization explanation; and the case studies honestly document a hard representational ceiling (radius-2 Morgan non-identifiability, Fig. 14(c)/Table 9) rather than hiding it. The reframing itself — treating the fingerprint posterior as a query distribution rather than a point estimate — is conceptually simple but, on this evidence, effective, and the candidate-pool scaling results give the field a falsifiable, easy-to-replicate direction.","major_comments":[{"comment":"The SOTA claim rests on comparing MS-GPT at candidate-pool size B=100 against FRIGID's numbers transcribed from [2], but the manuscript never states the inference budget (number of diffusion steps / candidate evaluations per spectrum) underlying FRIGID's Table 1 entries. This matters because FRIGID's central contribution is inference-time scaling, and MS-GPT's own Table 4 shows Top-10 gains of ~6 points from B=100 to B=1500. Figure 1 partially addresses this by plotting FRIGID's compute point, but the x-axis provenance ('timings taken from FRIGID') does not establish that the accuracy marker corresponds to FRIGID's best-budget configuration rather than a budget-matched one. Please state, in Table 1 or its caption, the per-spectrum evaluation budget of each baseline, and ideally include a FRIGID-at-matched-compute point on the Figure 1 frontier.","section":"Table 1 / §4.2 / Fig. 1"},{"comment":"The density band K=[0.95,1.50] is selected 'empirically with reference to validation-set metrics' and shared across benchmarks, and Appendix A.6 itself concludes the band 'may have been more conservative than necessary.' The component ablation brackets the band's value (outer-band and point-query controls in Tables 2–3), but there is no sensitivity analysis over (κ_min, κ_max) itself. Given that the headline margin over the strongest point-query control is ~4 Top-1 points (Table 2), and that Figures 11–12 show non-trivial val/test divergence in fidelity shares on NPLIB1, the robustness of the headline numbers to the band choice is load-bearing for the claim that calibration (rather than fortuitous endpoint tuning) drives the gain. A small sweep over band endpoints on the validation split, reported for both benchmarks, would close this. Relatedly, please clarify how much of the band selec","section":"§3.3 / App. A.6 / App. B.4"}],"minor_comments":[{"comment":"The uniform grid average in Eq. (6) is presented as approximating the posterior expectation in Eq. (1), but no weighting connects the linear-in-threshold grid to the actual posterior over fingerprints; the approximation is heuristic. A sentence clarifying that Eq. (6) is a coverage/ensemble device rather than a Monte Carlo estimate of Eq. (1) would prevent over-reading.","section":"Eq. (6) / §3.3"},{"comment":"No test-set sizes or confidence intervals are reported anywhere. With Top-1 margins of 4.7–5.6 points over FRIGID this is likely not decisive, but standard errors (or Wilson intervals) for the Table 1 entries should be added, particularly for MassSpecGym.","section":"§4.2 / Table 1"},{"comment":"Top-k Tanimoto is reported for all methods, but only MCES gets a comparability appendix (B.6). Please state the fingerprint type/radius used for the Tanimoto metric and confirm baseline papers used compatible definitions; otherwise add a comparability caveat analogous to the MCES one.","section":"Table 1 / §4.1 metrics"},{"comment":"Table 5 shows a one-epoch MS-GPT-Base trained on 500M–1B structures substantially outperforming the main model's Base variant (31.55/37.80 vs 28.42/33.78 Top-1/Top-10 on NPLIB1), yet the main results use two epochs over 100M. Please explain why the larger corpus was not used for the headline model, or clarify the chronology; as written it leaves open whether the main numbers are pretraining-limited.","section":"§4.1 Implementation / App. A.7 (Table 5)"},{"comment":"The rationale that extending below κ=1.00 'maps to the flat, high-threshold region of A(t)' is plausible but unsupported; a small plot of A(t) per dataset (or a pointer to where it can be found in the supplement) would help.","section":"§3.3 / App. B.4"},{"comment":"LaTeX artifacts: '29.8\\%/41.1\\%' and '23 .9%/28.7%' have stray spacing/escaping in the abstract; similar spacing issues appear in Table 1 ('25.037.100.58', '29.767.380.61'). Please proof the compiled tables.","section":"Abstract / Table 1"},{"comment":"D_enc is defined as a mean active-bit count over (x,C,m) drawn from T_enc (Eq. 3), but T_enc is a set of spectrum–structure pairs; please clarify whether multiple spectra per structure are counted with multiplicity and whether D_enc weights spectra or unique molecules.","section":"§3.3, Eq. (3)"}],"recommendation":"minor_revision","confidential_remarks":"The empirical comparisons depend on baseline numbers transcribed from other papers and on MIST checkpoints released by the FRIGID team; both are standard practice here and the paper is unusually transparent about MCES non-comparability (App. B.6), but I would suggest verifying that the promised code/checkpoint release is complete before final acceptance. The manuscript is a good fit for the journal's AI-for-science scope."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful takeaway is simple: fingerprint-mediated de novo has been training decoders on oracle bits and then feeding them one thresholded guess at test time. MS-GPT treats the spectrum encoder’s posterior as a query distribution—density-band calibration, multi-threshold group sampling, frequency ranking, plus a small LoRA on the query path—and posts clear Top-1/Top-10 jumps over FRIGID on both NPLIB1 and MassSpecGym (roughly +5 and +6–8 points). That is the result that matters in this subfield.\n\nWhat is actually new is not SAFE-GPT or MIST; it is the inference framing and the three pieces that implement it. The ablations back the story: group querying beats strong single-threshold controls, the band beats outer-band noise, and adaptation mostly lifts Top-10 recall. They also do the right diagnostic work—posterior-fidelity bins, pretraining-proximity bins, pool scaling, and case studies that admit radius-2 non-identifiability and “confidently wrong” posteriors. Code and checkpoints are promised. Circularity looks low: public splits, connectivity-equivalent structures stripped from pretraining, calibration from encoder train density rather than test labels.\n\nSoft spots are real but proportionate. Everything still rides on a frozen MIST radius-2 Morgan posterior and a hand-chosen band K=[0.95,1.50]; when fidelity is low or the fingerprint cannot separate isomers, more samples will not save you. Evaluation is known-formula only. MCES numbers across papers are not strictly comparable (they say so). Hyperparameters for the band and recoverability-weighted CE/KL are free knobs. None of that undoes the main claim; it bounds how far the method can go without a better spectrum→structure interface.\n\nThis is for people who build or use MS/MS structure generators, not for general ML theory. I would bring it to reading group, cite the posterior-querying setup and the numbers if I work nearby, and send it to referees. Solid empirical methods paper with honest limits.","headline":"Real empirical SOTA on de novo MS/MS via a clean fix for the oracle-vs-posterior mismatch; the fingerprint interface is still the ceiling.","tokens_in":24251,"tokens_out":527,"would_cite":true,"duration_ms":18839,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"De novo MS/MS structure recovery improves when a noisy fingerprint posterior is turned into a calibrated band of queries for a molecule-language model, not one thresholded fingerprint.","keywords":["tandem mass spectrometry","de novo structure elucidation","spectrum-induced posterior","molecule-language model","fingerprint-mediated generation","active-bit density calibration","AI for science"],"falsifier":"On held-out spectra whose posterior already agrees well with the true fingerprint inside the calibrated band, multi-query pooling plus adaptation would fail to beat strong single-threshold controls on Top-10 exact match; that result would collapse the posterior-querying claim.","tokens_in":24423,"feed_emoji":"🧪","tokens_out":939,"duration_ms":39982,"temperature":0.7,"pith_summary":"This paper targets the training–inference mismatch in fingerprint-mediated de novo structure elucidation from tandem mass spectra. Decoders are trained on perfect fingerprints computed from molecules, but at inference they usually receive a single noisy fingerprint obtained by thresholding a spectrum-induced posterior. MS-GPT instead treats that posterior as a source of many queries: it calibrates a band of active-bit densities near the oracle-fingerprint regime, samples a fixed candidate pool across the band, pools the generated structures, and ranks them by how often they appear. A small adapter then adjusts how posterior-derived queries are read without rewriting the pretrained molecular prior. On two public known-formula benchmarks the method raises exact structure match well above prior decoders, and larger candidate pools keep lifting recall at near-linear cost. A sympathetic reader cares because this is a practical route into chemical space that library search cannot reach, while keeping the speed of autoregressive molecule generation.","feed_headline":"Query a fingerprint band, not one threshold, from mass spectra","feed_subtitle":"Calibrated posterior queries to a molecule-language model set new exact-match marks on two de novo benchmarks.","key_machinery":"Spectrum-induced posterior querying: active-bit density band calibration chooses thresholds so query fingerprints stay near oracle active-bit density; fixed-size group querying spreads a fixed candidate pool across those thresholds; posterior-aligned LoRA adapts fingerprint encoding and cross-attention while freezing the molecule backbone.","core_discovery":"Fingerprint-mediated MS/MS de novo elucidation should be cast as spectrum-induced posterior querying of a conditional molecule-language model. Converting the continuous fingerprint posterior into a density-calibrated band of queries, distributing a fixed candidate budget across that band, ranking by generation-frequency consensus, and lightly adapting only the query-reading path yields new state-of-the-art Top-1/Top-10 exact-match accuracy of 29.8%/41.1% on NPLIB1 and 23.9%/28.7% on MassSpecGym.","pith_inferences":["Fixed-radius fingerprints can make true isomers indistinguishable even under a perfect oracle query, so some failures are representational ceilings, not posterior noise.","The same band-query pattern could apply to other inverse problems that force a continuous sensor posterior through a discrete pretrained interface.","Gains that concentrate at Top-10 more than Top-1 suggest the main product is a richer shortlist for expert review, not only a single automatic ID.","The fidelity analyses point next toward training or refining the spectrum encoder against the decoder’s query manifold rather than freezing it."],"forward_implications":["Fingerprint-mediated pipelines should keep the full posterior and query multiple operating points instead of one threshold.","A molecule-only pretrained decoder can be shared across spectral datasets, with only the spectrum encoder and a tiny adapter made domain-specific.","Autoregressive generation can raise structure recall by enlarging the candidate pool at near-linear extra inference cost.","Exact-match gains track spectrum-induced posterior fidelity; better encoders transfer directly into better elucidation.","Ranking by generation-frequency consensus across the query band is a workable substitute for collapsing uncertainty into one fingerprint."],"fun_headline_variants":["MS-GPT queries a calibrated fingerprint band, not one threshold","Cast MS/MS de novo as posterior querying of a molecule model","Density-calibrated fingerprint bands lift exact-match de novo accuracy","Pool candidates across a spectrum-induced fingerprint query band","LoRA-adapted molecule LM sets new MS/MS exact-match marks"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The method assumes that a fixed-bit fingerprint predicted from the spectrum, after population-level density calibration, still carries enough identifying structural signal for the language model to recover the true molecule.","fun_headline_variants_meta":{"raw":{"variants":["MS-GPT queries a calibrated fingerprint band, not one threshold","Cast MS/MS de novo as posterior querying of a molecule model","Density-calibrated fingerprint bands lift exact-match de novo accuracy","Pool candidates across a spectrum-induced fingerprint query band","LoRA-adapted molecule LM sets new MS/MS exact-match marks"]},"model":"grok-4.5","effort":"low","cost_usd":0.002684,"raw_usage":{"total_tokens":1097,"prompt_tokens":920,"num_sources_used":0,"completion_tokens":72,"cost_in_usd_ticks":26844000,"prompt_tokens_details":{"text_tokens":920,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":105,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":920,"tokens_out":72,"duration_ms":3551,"temperature":1.0,"reasoning_tokens":105,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-30T17:42:43.509843+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On held-out spectra whose posterior already agrees well with the true fingerprint inside the calibrated band, multi-query pooling plus adaptation would fail to beat strong single-threshold controls on Top-10 exact match; that result would collapse the posterior-querying claim.","supporting_citations":[],"review_version":1}