{"id":"373e4a49-9de3-4572-b438-091caca06abb","arxiv_id":"2505.23776","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":6,"one_line_summary":"The paper claims a convergent ASD signature: an fMRI classifier with AUC 0.89 on an independent test set, whose salient brain regions spatially match four of five ASD-associated gene expression modules.","lead":"This thesis reports a deep learning model that distinguishes autistic from typically developing brains using resting-state fMRI with an AUC of 0.89, and says the brain regions it relies on overlap with four of five autism-related gene expression modules. The wider interest is the promise of a non-invasive, biologically anchored biomarker for a disorder currently diagnosed through behavioral observation alone.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The cross-modal convergence claim rests on treating AHBA control-brain module expression as a spatial proxy for ASD pathology; until a spatial-null and independent-replication test is done, the 'significantly high absolute expression' overlap is not evidence of biological convergence.","rationale":"Only Ch. 1-5 and the table of contents are available; the methods and results chapters (Ch. 10-12) that substantiate the central claim are absent. The reader's UNVERDICTED status is therefore appropriate. Within the available material, the most load-bearing assumption is the imaging-transcriptomics premise: the abstract's third claim equates classifier-relevant brain regions with regions where ASD-associated modules have high absolute expression, as determined from AHBA control-brain expression. For this equation to carry biological meaning, the normative spatial distribution of module expression must indicate where the module is dysregulated in ASD. This is a standard but untested proxy in imaging transcriptomics, and the word 'proved' in the abstract overstates what a spatial overlap can establish. The absolute-value construction is a red flag because it discards the sign of dysregulation, and the absence of any described spatial-autocorrelation null makes it likely that the overlap significance is inflated. The untested BR-NN loss design (Sec. 3.4.2) is a second unresolved issue, but it is secondary because the later confounder-resilience chapters may address it. A concrete check using an independent ASD expression dataset and a spin null would settle whether the transcriptomic half of the convergence is genuine; until such a test is reported, the claim remains unverified.","tokens_in":45355,"tokens_out":6258,"duration_ms":66884,"concrete_test":"Compute the Ch. 12 overlap with module eigengenes re-estimated from an independent post-mortem ASD RNA-Seq dataset (e.g., PsychENCODE or the expanded Voineagu 2016 cohort) instead of AHBA controls, and assess significance with a spin-permutation null that preserves the spatial autocorrelation of the saliency map. If the 4-of-5 module overlap does not replicate or fails to beat the null (e.g., p > 0.05 after FDR), the convergence claim collapses. Also report the overlap using signed module eigengenes, not absolute values, to check whether the result depends on the absolute-value transform.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Central to the abstract's third claim is the transfer of ASD-associated gene modules onto a brain-wide expression map from the Allen Human Brain Atlas, as outlined in Ch. 9. The imaging-transcriptomic comparison (Ch. 12) presumably colocalizes classifier saliency with regions where 4/5 modules have 'significantly high absolute value of expression.' This is only meaningful if the spatial distribution of module expression in healthy control brains indicates where the module is pathologically dysregulated in ASD. The reviewed chapters provide no validation of this premise. Two specific problems follow. First, the use of absolute value is ambiguous: if a module is down-regulated in ASD, the relevant regions may show low, not high, expression; using |eigengene| conflates sign and requires a biologically motivated baseline that has not been stated. Second, both neuroimaging saliency maps and gene-expression maps are spatially autocorrelated; a naive overlap test can be significant by chance. The available text also notes (Sec. 3.4.2) that the BR-NN's adversarial loss design was never experimentally tested, so the imaging map itself may carry confounder signal; but even granting the imaging map, the transcriptomic half of the convergence is unsupported without a spatial null and independent replication.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a PhD thesis posted on arXiv that claims three results: (1) a bias-resilient deep learning classifier for ASD using resting-state fMRI, with an AUC of 0.89 on an independent test set; (2) five ASD-associated gene co-expression modules identified from post-mortem transcriptomics; and (3) an imaging-transcriptomics comparison that, in the abstract's words, \"proved\" that the brain regions relevant to the classifier are those where four of the five modules take a significantly high absolute expression value. The text available for review contains the abstract, Introduction, and background chapters on ASD, neuroimaging methods, genetics, and imaging genetics, but the Materials and Methods, Results, and Discussion chapters announced in the table of contents are not present in the provided text. The three central claims are therefore stated but their supporting statistical and experimental details are absent.","tokens_in":45511,"tokens_out":5280,"duration_ms":52718,"significance":"If the results hold, the claimed cross-modal convergence would be of genuine interest: it would provide a spatial link between an fMRI-derived biomarker and ASD-associated gene expression, with potential implications for early diagnosis and for interpreting neuroimaging signals in terms of molecular pathology. The study design has notable strengths: the imaging and transcriptomic analyses rely on largely independent data sources, the manuscript demonstrates awareness of confounder problems in large multi-site MRI datasets, and the literature review is thorough. However, because the actual methods and results are not present in the reviewed text, the significance cannot currently be assessed beyond the plausibility of the design. The claimed convergence, if substantiated, would be a valuable contribution, but it is not yet a verifiable contribution.","major_comments":[{"comment":"The posted text contains the abstract and background chapters only; the Materials and Methods, Results, and Discussion chapters that would support the three central claims are absent. The abstract reports an AUC of 0.89, five gene modules, and a 'significantly high absolute value of expression' overlap, but none of the underlying performance numbers, module definitions, test statistics, thresholds, or significance corrections are available in the reviewed text. This is a load-bearing omission: the central claims cannot be checked, replicated, or fairly criticized at the statistical level. A revision must supply the complete methodological and results sections, or the claims must be withdrawn.","section":"Abstract, third bullet; TOC Parts III–V"},{"comment":"The text explicitly states that the adversarial correction design of the BR-NN, based on the combined LC1/LC2 losses, was not experimentally tested: 'Despite no experimental tests have been made to assess the efficacy of this choice'. This is an author-acknowledged absence of support for the very component that the classifier's confounder resilience depends on. Since the saliency map used in the imaging-transcriptomics comparison is generated by this classifier, an ineffective adversarial correction would directly undermine the biological interpretation of the overlap. This point needs either empirical validation of the adversarial scheme or a clear statement of what the imaging results would mean if the scheme is only partially effective.","section":"Section 3.4.2"},{"comment":"The convergence claim rests on transferring ASD-associated module expression from post-mortem ASD/TDC transcriptomics to the spatial expression distribution of Allen Human Brain Atlas data, which is derived from control brains. The reviewed text does not define the test statistic used for 'significantly high absolute value of expression', nor the threshold for calling a region 'relevant', nor a spatial null model that accounts for the strong spatial autocorrelation of both neuroimaging saliency maps and gene-expression maps. Without such a spatial null and an independent replication, the reported overlap can be expected to be significant by chance and cannot be interpreted as biological convergence.","section":"Chapter 9 / Chapter 12"},{"comment":"The use of the absolute value of the module eigengene is ambiguous and biologically unmotivated as described. If a module is down-regulated in ASD, the relevant brain regions might show low, not high, expression in control-brain data, so the sign of the eigengene carries information that the absolute value discards. The abstract's claim of 'significantly high absolute value of expression' therefore requires a clear baseline and a biological rationale for why high absolute expression, rather than direction-specific dysregulation, is the relevant quantity. This issue is central to the third claim and is not addressed in the provided text.","section":"Chapter 9 / Eq. (3.5) and abstract"},{"comment":"The abstract uses 'proved' twice to describe the spatial overlap result. Given that the analysis involves multiple free thresholds (saliency threshold, overlap significance threshold, network cut height, soft-threshold power, adversarial weighting) and no causal link is demonstrated, 'proved' overstates the evidential quality of a spatial association. This is a substantive concern because the proof claim is the headline of the paper; the wording should be revised to 'consistent with' or 'suggests' unless a formal, validated statistical proof is supplied.","section":"Abstract"}],"minor_comments":[{"comment":"The phrase 'Effective therapies to reduce all the heterogeneous symptoms of the disorder do not exists yet' contains a grammar error: it should be 'do not exist yet'.","section":"Abstract"},{"comment":"'Trascriptomics' should be spelled 'Transcriptomics' in the chapter title listing.","section":"Table of Contents"},{"comment":"The phrase 'the few higher-performing results are never replied in subsequent studies' should likely read 'never replicated' rather than 'replied'.","section":"Section 3.3.2"},{"comment":"There is a typo in the sentence 'minimizes th ability of C to predict the confounders'; 'th' should be 'the'.","section":"Section 3.4.2"},{"comment":"'In addiction' is used where 'In addition' is intended.","section":"Section 2.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript as provided to the referee is essentially only the front matter and background of a PhD thesis; the results, methods, and statistical details for all three central claims are in the missing chapters. This is not a situation where the science can be assessed. I recommend that the editor request the complete version of the thesis or a journal-style article containing the full methods and results before further review. In addition, the abstract's use of 'proved' is too strong for a spatial overlap analysis and should be softened in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: based on what is actually in the file—abstract, front matter, background chapters, and the table of contents—I cannot verify the central claims. If the full thesis delivers what the abstract promises, it is a significant result. Based only on what I have, this is an unverdictable thesis, not a paper I can endorse or reject.\n\nWhat is genuinely new is the chain: a confounder-resilient rsfMRI classifier with an AUC of 0.89 on an independent test set, WGCNA modules from post-mortem ASD transcriptomics, and a spatial overlap between classifier saliency and AHBA-derived module expression. That specific convergence claim is not in the prior literature described in the background chapters. The background prose is also careful and honest. The discussion of MRI being non-quantitative, the need for validation on truly unseen data, and the explicit note in Sec. 3.4.2 that the BR-NN's dual-loss design was never experimentally tested all suggest a serious thinker who is not overstating what was done. That honesty counts, though it cuts against the imaging half of the convergence claim.\n\nThe soft spots are large, but they are soft because evidence is missing, not because the argument is internally incoherent. The abstract uses the word \"proved\" for a spatial overlap, yet no test statistic, multiple-comparison correction, or spatial null appears in the reviewed portion. The stress-test concern about using AHBA control-brain expression as a spatial proxy for ASD pathology is legitimate but not confirmed; the relevant methods and results chapters are absent. The use of absolute eigengene expression is genuinely ambiguous: if a module is down-regulated in ASD, regions with low expression may be the relevant ones. And both saliency maps and expression maps are spatially autocorrelated, so a naive overlap test can look significant by chance. The AUC of 0.89 also needs site-level split details and leakage checks, especially since the background itself documents that large-dataset ASD classifiers usually plateau around 70%.\n\nMy take: this deserves a careful referee if and only if the full methods and results are supplied. The question is important, the thesis looks serious, and the novelty is real. I would ask for the complete chapters, and specifically for a spatial permutation null, a sign-aware treatment of module expression, and independent replication of the overlap. Until then, the right verdict is \"not yet verifiable,\" not accept or reject.","headline":"Readable, honest thesis front-matter with a potentially important cross-modal claim; the actual evidence is not in the files I was given, so it is unverdictable rather than wrong.","tokens_in":46179,"tokens_out":2013,"would_cite":false,"duration_ms":23221,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This thesis claims that the brain regions driving an fMRI-based ASD classifier are the same regions where four of five ASD-associated gene modules are most strongly expressed.","keywords":["autism spectrum disorder","resting-state fMRI","imaging transcriptomics","gene co-expression network","deep learning","adversarial debiasing","module eigengene","biomarker"],"falsifier":"Compute SmoothGrad saliency maps from the same rsfMRI data using classifiers trained on non-ASD labels such as sex or acquisition site; if those maps overlap the four gene modules as strongly as the ASD map does, the overlap is not ASD-specific. Alternatively, recompute module eigengenes in an independent post-mortem ASD dataset with multiple regions and check that the same four modules still show high absolute expression in the classifier-relevant regions.","tokens_in":45054,"feed_emoji":"🧠","tokens_out":5465,"duration_ms":55033,"temperature":0.7,"pith_summary":"Across the heterogeneity of autism, this thesis tries to establish one shared, measurable signature: the brain regions that let a machine-learning classifier separate ASD from typical controls using resting-state fMRI are the same regions where four of the five ASD-associated gene-expression modules have significantly high expression. To get there, the thesis builds a confounder-robust deep classifier reaching an AUC of 0.89 on an independent test set, identifies five gene co-expression modules from post-mortem ASD and control transcriptomes, and then projects module eigengenes onto the same brain space as the classifier's saliency map. The reason this matters is that a non-invasive imaging biomarker validated at the transcriptomic level would be usable for early diagnosis, while transcriptomic data alone can only be obtained post-mortem.","feed_headline":"Autism's fMRI signature matches four gene modules' brain expression","feed_subtitle":"A confounder-proof classifier hits AUC 0.89; its salient regions overlap where four ASD gene modules are strongly expressed.","key_machinery":"The argument is carried by three coupled objects. First, a two-stream CNN treats a 4D resting-state fMRI and its optical flow as separate input streams, so that spatial and temporal-motion information are learned separately. Second, a Bias-Resilient Neural Network, an adversarial tripartite architecture of feature extractor, predictor, and confounder predictor, forces the learned features to be uninformative about specified confounders such as sex, age, and acquisition site. Third, SmoothGrad relevance maps average input-gradient explanations over noisy perturbations to give a fuzzy region of interest per subject. On the transcriptomic side, the load-bearing objects are weighted gene co-expression network modules and their first principal component, the module eigengene. The thesis computes eigengene values for each module across brain regions using Allen Human Brain Atlas expression data mapped to the AAL atlas, then tests whether the classifier-relevant regions show significantly high absolute eigengene expression. SmoothGrad provides the imaging pattern, the eigengene provides the transcriptomic pattern, and the spatial comparison is what establishes convergence.","core_discovery":"The central claim, stated in the abstract, is that the regions characterizing the ASD brain at the neuroimaging level are precisely those in which four of the five gene modules take a significantly high absolute expression value. The neuroimaging component is a two-stream convolutional network that processes resting-state fMRI and its optical flow, made resilient to confounders by an adversarial bias-resilient network, reaching an AUC of 0.89 on an independent test set; the transcriptomic component is a weighted gene co-expression network analysis on post-mortem ASD and control brains yielding five ASD-associated modules; the imaging-genetics component compares the classifier's SmoothGrad relevance regions with Allen Human Brain Atlas module eigengene expression in AAL regions. The convergence of the two independent modalities is taken as mutual validation, showing that a neuroimaging biomarker of ASD can be identified and confirmed by transcriptomics despite the disorder's heterogeneity.","pith_inferences":["Because the AHBA expression data come from control brains, the convergence likely maps regions of intrinsic vulnerability rather than active pathology; a natural follow-up is to test whether the same regions show elevated dysregulation in ASD post-mortem samples at single-cell resolution.","The paper's use of absolute expression leaves open whether up- and down-regulated modules mark the same or different regions; separating directions could sharpen the imaging-transcriptomic match.","A testable extension is to apply the same spatial-overlap pipeline to sex, age, or symptom-severity labels, predicting that only the ASD classifier yields a significant match.","The pipeline is portable to other polygenic conditions with both fMRI cohorts and brain transcriptome atlases, so the convergence claim can be probed outside ASD."],"forward_implications":["If correct, the rsfMRI classifier's saliency map is not an arbitrary artifact of deep learning but a spatial readout of ASD-related transcriptional activity, giving the imaging biomarker a molecular grounding.","Four modules, not just one, converge on the same regions, suggesting that even a heterogeneous disorder shares a partly common regional pathology.","The bias-resilient design implies that the reported AUC of 0.89 is not driven by sex, age, or acquisition-site confounders, so the pattern should transfer to new multi-site datasets.","The method provides a template for validating deep-learning neuroimaging biomarkers in other disorders by comparing saliency maps with disease-gene expression atlases.","Clinically, such a convergent signature could eventually support earlier, objective diagnosis when behavioural assessment is inconclusive."],"supporting_citations":[{"why":"Grounds the imaging-transcriptomics comparison by showing that brain regions belonging to the same functional network share correlated gene expression.","marker":"[26]"},{"why":"Reports enrichment of ASD-associated genes among transcripts whose expression correlates with default-mode-network activity, motivating imaging-transcriptomics for ASD.","marker":"[28]"},{"why":"Supplies the two-stream CNN architecture that processes rsfMRI and optical flow separately for classification.","marker":"[155]"},{"why":"Supplies the adversarial Bias-Resilient Neural Network used to strip confounder information from the learned features.","marker":"[156]"},{"why":"Supplies SmoothGrad, the explainability method that produces the classifier relevance maps defining the brain regions of interest.","marker":"[157]"},{"why":"Provides the largest review of ASD transcriptomic studies whose convergent findings frame the gene module analysis.","marker":"[258]"},{"why":"Earlier multi-region ASD transcriptome study identifying neuronal and immune-glial co-expression modules that the present module selection builds on.","marker":"[259]"},{"why":"Expanded RNA-Seq re-analysis corroborating ASD cortical transcriptome attenuation and core signatures used to validate the modules.","marker":"[260]"}],"fun_headline_variants":["Autism fMRI model ties brain regions to four gene modules","ASD's neuroimaging biomarker confirmed by four gene modules","Deep learning finds ASD biomarker echoed in four gene networks","Imaging-transcriptomics convergence: four modules light up fMRI","AUC 0.89: fMRI spots autism, gene expression backs it"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The convergence result rests on the premise that gene-expression patterns measured in control brains are a valid map of where ASD pathology occurs, so the spatial overlap between classifier relevance and module expression counts as biological convergence rather than mere geography.","fun_headline_variants_meta":{"raw":{"variants":["Autism fMRI model ties brain regions to four gene modules","ASD's neuroimaging biomarker confirmed by four gene modules","Deep learning finds ASD biomarker echoed in four gene networks","Imaging-transcriptomics convergence: four modules light up fMRI","AUC 0.89: fMRI spots autism, gene expression backs it"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000713,"raw_usage":{"total_tokens":3256,"prompt_tokens":1040,"completion_tokens":2216,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":656,"completion_tokens_details":{"reasoning_tokens":2131}},"tokens_in":656,"tokens_out":2216,"duration_ms":14743,"temperature":1.0,"reasoning_tokens":2131,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:16:38.820897+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute SmoothGrad saliency maps from the same rsfMRI data using classifiers trained on non-ASD labels such as sex or acquisition site; if those maps overlap the four gene modules as strongly as the ASD map does, the overlap is not ASD-specific. Alternatively, recompute module eigengenes in an independent post-mortem ASD dataset with multiple regions and check that the same four modules still show high absolute expression in the classifier-relevant regions.","supporting_citations":[{"cited_title":"Variation in gene expression in autism spectrum disorders: an extensive review of transcriptomic studies,","cited_arxiv_id":null,"evidence_quote":"Provides the largest review of ASD transcriptomic studies whose convergent findings frame the gene module analysis."},{"cited_title":"Transcriptomic analysis of autistic brain reveals convergent molecular pathology,","cited_arxiv_id":null,"evidence_quote":"Earlier multi-region ASD transcriptome study identifying neuronal and immune-glial co-expression modules that the present module selection builds on."},{"cited_title":"Genome-wide changes in lncrna, splicing, and regional gene expression patterns in autism,","cited_arxiv_id":null,"evidence_quote":"Expanded RNA-Seq re-analysis corroborating ASD cortical transcriptome attenuation and core signatures used to validate the modules."}],"review_version":1}