{"id":"e301ac3f-4841-4e97-af98-3ceb2bf38518","arxiv_id":"2605.26192","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"AIMS-Fold integrates XL-MS and HDX-MS data as differentiable potentials to guide diffusion sampling for more accurate protein complex structures than unguided models like Boltz-2 on induced-proximity targets.","lead":"AIMS-Fold is a guided diffusion framework that steers protein complex structure generation using spatial restraints from XL-MS and solvent accessibility from HDX-MS. A smart generalist might read it because the approach targets improved modeling for induced-proximity drug modalities such as PROTACs and antibodies.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Conversion of sparse XL-MS/HDX-MS data into differentiable potentials may introduce unquantified bias","rationale":"The reader's weakest_assumption directly identifies the same point. Because the supplied abstract contains no quantitative validation of the potential construction and the full text is referenced but not reproduced here, the concern cannot be ruled out; the verdict therefore remains UNVERDICTED pending the concrete test above.","tokens_in":1708,"tokens_out":352,"duration_ms":31134,"concrete_test":"Re-implement the potential functions exactly as described in the methods; on a benchmark set of 10 induced-proximity complexes with both known crystal structures and withheld XL-MS/HDX-MS data, run AIMS-Fold with (a) real potentials and (b) potentials computed from randomized but same-density restraints. If the accuracy advantage over Boltz-2 disappears or reverses under randomization, the guidance effect is not robust to the data-to-potential mapping.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline claim requires that the derived potentials steer the diffusion trajectory to biologically correct states on induced-proximity targets without systematic error or degradation of the base model's priors. The abstract states that potentials are 'derived from XL-MS spatial restraints and HDX-MS solvent accessibility profiles' but supplies no functional form, weighting scheme, or conflict-resolution rule for heterogeneous/sparse inputs. If the mapping from measurement to potential is misspecified (e.g., treating cross-link distances as hard harmonic restraints when they are ensemble averages), the guidance can favor non-native conformations even while improving nominal metrics. This assumption is the single least-secured step between the experimental data and the reported accuracy gain over Boltz-2.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript introduces AIMS-Fold, an inference-time guided-diffusion framework that steers pretrained diffusion models for protein complex structure prediction by incorporating differentiable physical potentials derived from XL-MS spatial restraints and HDX-MS solvent accessibility profiles. The central claims are that the individual data types improve accuracy, their combination produces synergistic gains, and the resulting model outperforms unguided state-of-the-art methods such as Boltz-2 on challenging induced-proximity targets relevant to antibody and PROTAC design. Public release of evaluation code is promised.","tokens_in":1866,"tokens_out":404,"duration_ms":23026,"significance":"If the quantitative results and potential formulations hold, the work could offer a practical route to hybrid experimental-computational modeling of conformationally dynamic complexes that current generative models handle poorly. The emphasis on inference-time guidance rather than retraining preserves the base model's learned priors while adding experimental constraints, and the promised code release aids reproducibility.","major_comments":[{"comment":"Abstract: The headline claim of synergistic improvement and superiority over Boltz-2 is asserted without any quantitative metrics, dataset sizes, validation protocols, or numerical comparisons. This absence prevents evaluation of whether the reported accuracy gains are statistically meaningful or merely nominal.","section":"Abstract"},{"comment":"Abstract (framework description): No functional form, scaling factors, weighting scheme, or conflict-resolution procedure is supplied for converting sparse XL-MS cross-link distances and HDX-MS accessibility profiles into differentiable potentials. Because the guidance mechanism is the sole link between the experimental inputs and the claimed accuracy gains, this omission is load-bearing for the central thesis.","section":"Abstract"}],"minor_comments":[{"comment":"The abstract refers to 'challenging induced proximity targets' without naming the specific complexes or providing even a high-level description of the test set.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for their constructive comments regarding the abstract. We address each point below and will revise the abstract accordingly to improve clarity and support for the central claims.","responses":[{"response":"We agree that the abstract would benefit from inclusion of key quantitative results to substantiate the claims of synergistic improvement and outperformance over Boltz-2. The full manuscript reports these details, including specific accuracy metrics on a dataset of induced-proximity targets, validation protocols, and statistical comparisons. We will revise the abstract to incorporate concise numerical highlights of the main results while preserving its length constraints.","revision_made":"yes","referee_comment":"[Abstract] Abstract: The headline claim of synergistic improvement and superiority over Boltz-2 is asserted without any quantitative metrics, dataset sizes, validation protocols, or numerical comparisons. This absence prevents evaluation of whether the reported accuracy gains are statistically meaningful or merely nominal."},{"response":"The functional forms of the differentiable potentials (harmonic restraints for XL-MS distances and accessibility-derived terms for HDX-MS), along with scaling, weighting, and multi-potential conflict handling during guidance, are fully specified in the Methods section. We acknowledge that a high-level description of this mechanism would strengthen the abstract. We will add a brief clause to the abstract summarizing the guidance approach.","revision_made":"yes","referee_comment":"[Abstract] Abstract (framework description): No functional form, scaling factors, weighting scheme, or conflict-resolution procedure is supplied for converting sparse XL-MS cross-link distances and HDX-MS accessibility profiles into differentiable potentials. Because the guidance mechanism is the sole link between the experimental inputs and the claimed accuracy gains, this omission is load-bearing for the central thesis."}],"tokens_in":1343,"tokens_out":375,"duration_ms":18199,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"AIMS-Fold turns XL-MS cross-link distances and HDX-MS accessibility profiles into inference-time potentials that steer a pretrained diffusion model toward conformations relevant for induced-proximity targets. The main new element is the joint use of those two specific experimental sources as differentiable terms in the same sampling run.\n\nThe paper correctly identifies that current generative models handle single chains reasonably well but often miss the right interface or state in complexes, and that structural proteomics data exists but has not been routinely folded into the sampling process. Framing the task around practical drug-design needs is useful.\n\nThe soft spots sit in the missing mechanics. The abstract states that the potentials are derived from the measurements and that their combination produces synergistic gains, yet it gives no functional form, no scaling between the two data types, and no rule for handling sparse or conflicting restraints. The headline result—that the guided model beats Boltz-2 on challenging targets—is asserted without any reported metrics, test-set description, or comparison protocol. That gap makes the central claim impossible to assess from the text provided.\n\nThe stress-test point about possible systematic bias in the measurement-to-potential mapping is therefore live; without the actual equations or code, one cannot check whether the guidance preserves the base model’s priors or simply pulls toward an ensemble average that may not be native. If the full manuscript contains the implementation and the promised evaluation code, those concerns can be checked directly.\n\nThe work is aimed at computational groups already generating or using XL-MS and HDX data for PROTAC or antibody design. A reader who wants to test an experimental-guidance layer on their own targets would get value once the methods are reproducible.\n\nIt deserves peer review because the integration direction addresses a documented practical limitation and the idea is concrete enough to be worth referee scrutiny, even if the current write-up needs the missing quantitative and technical sections filled in.","headline":"AIMS-Fold adds XL-MS and HDX-MS guidance to diffusion sampling for induced-proximity complexes, but the abstract supplies no formulation details or accuracy numbers to evaluate the superiority claim over Boltz-2.","tokens_in":2368,"tokens_out":465,"would_cite":false,"duration_ms":21326,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Mass spectrometry data steers diffusion models to more accurate protein complex structures","keywords":["protein structure prediction","diffusion models","XL-MS","HDX-MS","induced proximity","structural proteomics","protein complexes"],"falsifier":"A benchmark of induced proximity complexes with available XL-MS and HDX-MS data on which AIMS-Fold fails to exceed the accuracy of Boltz-2 would falsify the central performance claim.","tokens_in":2633,"feed_emoji":"🧬","tokens_out":597,"duration_ms":32514,"temperature":0.7,"pith_summary":"Protein structure generative models routinely fail to capture the correct conformational states of protein complexes that matter for induced proximity modalities such as PROTACs and antibodies. AIMS-Fold bridges this gap by converting XL-MS spatial restraints and HDX-MS solvent accessibility profiles into differentiable physical potentials that actively steer the sampling trajectory of pretrained diffusion models at inference time. The framework shows that each data type improves accuracy on its own and that their combination produces further synergistic gains. On challenging induced proximity targets the guided model outperforms unguided state-of-the-art methods such as Boltz-2.","feed_headline":"Mass spec data steers diffusion models to better protein complexes","feed_subtitle":"AIMS-Fold converts XL-MS and HDX-MS measurements into potentials that guide sampling and beat unguided models on induced proximity targets","key_machinery":"AIMS-Fold guided-diffusion framework that turns XL-MS and HDX-MS measurements into differentiable physical potentials to steer the diffusion trajectory","core_discovery":"AIMS-Fold is an inference-time guided-diffusion framework that converts XL-MS spatial restraints and HDX-MS solvent accessibility profiles into differentiable physical potentials derived from structural proteomics measurements; these potentials steer the generative sampling trajectory of pretrained diffusion models and yield higher accuracy on induced proximity targets than purely computational unguided models.","pith_inferences":["The same steering approach could be tested with other sparse experimental restraints if differentiable potentials can be defined for them","Reliable complex structures from this method could directly support computational design of bifunctional molecules that induce proximity","The results suggest that inference-time guidance with experimental data is a general route to improve diffusion models on tasks where static sequence-to-structure prediction is insufficient"],"forward_implications":["XL-MS restraints alone improve predictive accuracy on protein complexes","HDX-MS data alone improves predictive accuracy on protein complexes","Combining the two data types produces synergistic gains in accuracy","The guided approach outperforms unguided models on induced proximity targets relevant to drug design"],"fun_headline_variants":["XL-MS and HDX-MS guide diffusion models to protein complexes","Structural proteomics potentials steer AIMS-Fold sampling","Mass spec data improves complex predictions in guided diffusion","AIMS-Fold uses XL-MS restraints for induced proximity models","HDX-MS and XL-MS enhance diffusion model accuracy on complexes"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"Sparse heterogeneous XL-MS and HDX-MS measurements can be turned into differentiable physical potentials that guide the model toward correct conformations without introducing systematic bias or degrading the pretrained priors.","fun_headline_variants_meta":{"raw":{"variants":["XL-MS and HDX-MS guide diffusion models to protein complexes","Structural proteomics potentials steer AIMS-Fold sampling","Mass spec data improves complex predictions in guided diffusion","AIMS-Fold uses XL-MS restraints for induced proximity models","HDX-MS and XL-MS enhance diffusion model accuracy on complexes"]},"model":"grok-4.3","cost_usd":0.003666,"raw_usage":{"total_tokens":1896,"prompt_tokens":644,"num_sources_used":0,"completion_tokens":78,"cost_in_usd_ticks":36662000,"prompt_tokens_details":{"text_tokens":644,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1174,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":644,"tokens_out":78,"duration_ms":10508,"temperature":1.0,"reasoning_tokens":1174,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-29T22:28:32.436422+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A benchmark of induced proximity complexes with available XL-MS and HDX-MS data on which AIMS-Fold fails to exceed the accuracy of Boltz-2 would falsify the central performance claim.","supporting_citations":[],"review_version":1}