{"id":"9e2e622c-72b9-49df-b251-7dfa4732172c","arxiv_id":"2502.02630","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":10,"one_line_summary":"A new cross-modal deep learning model, scBIT, pairs single-cell gene expression with fMRI to improve Alzheimer's diagnostic accuracy, though the reported gains may be inflated by target-label leakage.","lead":"This paper introduces a computational framework that pairs single-cell RNA sequencing data with fMRI to predict Alzheimer's disease diagnosis, reporting accuracy gains up to 26.6% in five-class classification. The method is a possible template for combining molecular and imaging data, but its evaluation includes a likely data leakage step that could inflate the reported gains.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Test-set fMRI features appear to enter pre-training of the cross-modal attention encoders, so the reported snRNA gain may reflect transductive leakage rather than modality value.","rationale":"The central claim is an empirical performance gain. Everything else—novelty, interpretability—is secondary. The single assumption that carries the claim is that the 10-fold evaluation isolates the contribution of snRNA. Section 4.2's sentence only masks disease-state similarity labels; it does not state that test fMRI scans are excluded from the pretraining of the four cross-modal encoders. Since those encoders produce the attention features used by the classifier, and since the no-snRNA comparator does not receive this pretraining, the comparison is not controlled. This is a concrete, testable flaw rather than a disagreement with consensus. The reader's weakest_assumption points to the same sentence and reaches the same conclusion, so I agree. A strict fold-split rerun would settle it. If the gains persist under that protocol, the central claim would be supported; if they collapse, the current reported improvements are evaluation artifacts.","tokens_in":15407,"tokens_out":4539,"duration_ms":44865,"concrete_test":"Rerun the 10-fold cross-validation with a strict fold-aware protocol: pretrain E_rm^{pt,s} for all similarity types and compute snRNA attention only on the training folds, then evaluate on the held-out fold. Report binary and five-class ACC with and without snRNA under this protocol. If the snRNA gain falls materially below the reported 3.39%/26.59% or vanishes, the central claim is not supported. As a secondary check, inspect the released code to confirm that the pretraining data loader excludes all test-fold fMRI samples for all four similarity encoders.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's evidence for the central claim—that snRNA improves fMRI-based AD prediction by 3.39% (binary) and 26.59% (five-class)—rests on the evaluation protocol stated in Section 4.2: 'To avoid the data leakage problem, we mask the disease state similarity corresponding to the testing set during the training of the pre-trained encoder.' Masking the state-similarity label is not sufficient to prevent leakage. In Stage 2 (Section 3.3), four pre-trained fMRI encoders E_rm^{pt,s} are built using the fMRI data for similarity types s ∈ {age, sex, state, gene}. The description does not say test fMRI examples are excluded from pre-training; it only masks one label type for one encoder. Thus the age, sex, and gene encoders can be fit with test-fMRI inputs, and their outputs enter the attention scores a_i^s = csim(P, E_rm^{pt,s}(X_i)) that are fed into the Stage 3 classifier. The 'without snRNA' baseline, by contrast, uses only the Stage 3 ROI encoder trained on training folds. This asymmetry, not snRNA content, can explain part or all of the reported accuracy gains. The claim that masking prevents leakage is therefore load-bearing and is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes scBIT, a three-stage framework that integrates single-nucleus RNA-seq (snRNA) with resting-state fMRI for Alzheimer's disease (AD) diagnosis. Stage 1 converts snRNA data into cell-type-specific gene subgraph embeddings using a self-explainable graph neural network; Stage 2 trains four fMRI encoders on age, sex, clinical-state, and genetic similarities to compute cross-modal attention scores against the snRNA prototypes; Stage 3 fuses those attention scores with fMRI embeddings for classification. The authors report ten-fold cross-validation on ADNI, claiming that incorporating snRNA improves binary classification accuracy by 3.39% and five-class classification accuracy by 26.59%, and they release code on GitHub and Zenodo.","tokens_in":15712,"tokens_out":7910,"duration_ms":78909,"significance":"If the empirical results were sound, scBIT would be a genuinely novel cross-modal diagnostic pathway that brings single-cell transcriptomics into fMRI-based AD prediction, with interpretable brain-region–gene associations. The paper has clear strengths: it uses public datasets, releases code and detailed instructions, and builds the interpretability machinery (ProtGNN-style subgraph extraction and cell-type-level attention) directly into the model. However, the central empirical claim rests on an evaluation protocol that appears to include test-set fMRI features during pre-training, which undermines the reported accuracy gains. The contribution is valuable in conception, but the current evidence does not establish the claimed snRNA benefit.","major_comments":[{"comment":"The statement that 'to avoid the data leakage problem, we mask the disease state similarity corresponding to the testing set during the training of the pre-trained encoder' is not sufficient to prevent leakage. The text does not say that the test fold's fMRI inputs are excluded from pre-training the four encoders E_rm^{pt,s} in Eq. (12); it only masks one label type. Consequently, the age, sex, and gene encoders can be trained with test fMRI features and their labels, and even the state encoder can see test fMRI inputs. Because the attention scores in Eq. (14) are computed with these encoders and then fed into the classifier in Eq. (16), test-set statistics can enter the predictions. The load-bearing claim of snRNA-driven improvement therefore is not established; the evaluation must be redone with the test fold excluded from all pre-training.","section":"Section 4.2 and Stage 2 of Section 3.3"},{"comment":"The 'without snRNA' baseline is not a controlled ablation of the snRNA modality. It omits the entire Stage 2 pretrained-encoder and attention pathway, so the reported 3.39% and 26.59% gains cannot be attributed specifically to the snRNA-derived prototypes; they may instead come from the additional pretrained features, particularly the state-similarity encoder. The paper should compare scBIT against a baseline that includes the same Stage 2 encoders and attention features but with the snRNA prototypes P replaced by a non-informative or randomly permuted prototype set, so that the snRNA contribution is isolated.","section":"Section 4.2, Fig. 3(a)"},{"comment":"The state-similarity encoder is trained directly on the clinical-state labels, and its attention score a_i^{state} is fed into the classifier predicting those same clinical states. Even under a clean data split, this creates a feature that is essentially a learned encoding of the diagnostic label, and the MoE weight for the state similarity is reported as 11.11% (Fig. 3(h)). The paper should include an ablation that removes the state-similarity attention or replaces it with a non-diagnostic similarity, and should also validate the linear mapping from SEA-AD 'Low/Intermediate/High' to ADNI 'EMCI/MCI/LMCI' categories; without such an ablation, the snRNA modality itself may not be responsible for the observed improvement.","section":"Section 3.3, Eqs. (13) and (16), and Section 4.3"}],"minor_comments":[{"comment":"The subsection title contains a typo: 'Data and materlals' should read 'Data and materials'.","section":"Section 4.1"},{"comment":"Equation (3) writes d_q = [sim(p_i, h_q) | k = 1 … 90], but the index in the expression should be consistent with the loop variable; presumably it should be i = 1 … 90 or p_k.","section":"Eq. (3)"},{"comment":"The symbol ∤(·) is used for embedding normalization but is never defined; please use a standard notation such as ‖·‖-normalized or define the operation explicitly.","section":"Eq. (13)"},{"comment":"The paper does not report the number of ADNI fMRI subjects or the class counts for the binary and five-class tasks; these numbers are essential for interpreting the reported accuracies and standard deviations.","section":"Section 4.2"},{"comment":"The comparison with prior methods in Table 1 relies on best results reported in the literature, which may use different data splits, patient selections, and preprocessing; the text should clarify whether metrics are matched on the same subjects and folds.","section":"Table 1"},{"comment":"The subgroup accuracies for gender and age are reported without standard deviations or sample sizes; please include these to support the subgroup comparison.","section":"Fig. 3(b) and Fig. 3(c)"}],"recommendation":"major_revision","confidential_remarks":"The reader's central concern about data leakage is well grounded: the text explicitly says that the disease-state similarity for the test set is masked during pre-training, which implies the testing fold is present in pre-training; masking one label type does not eliminate leakage from the other three encoders or from the test fMRI features themselves. This is a load-bearing flaw in the current evaluation, and the reported numerical gains should not be accepted as evidence for the snRNA contribution until the evaluation is redone with a strict per-fold split. The idea and the released code are promising, so I believe a major revision requiring re-running the experiments is the appropriate outcome rather than outright rejection, provided the authors can re-establish the central claim under a leakage-free protocol."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline: this is the first paper I've seen that integrates single-nucleus RNA-seq with fMRI for AD diagnosis, and the architecture is thoughtfully put together. But the reported accuracy gains are not credible as presented, because the evaluation leaks test-set information into the models that produce the diagnostic features.\n\nWhat's actually new and good: the bagging-to-gene-subgraph representation via ProtGNN is a clever way to get cell-type-level embeddings, and the four-similarity cross-modal attention with MoE fusion is coherent. They also ship code and data, which is more than most papers in this space.\n\nThe soft spot is load-bearing. Section 4.2 says they 'mask the disease state similarity corresponding to the testing set during the training of the pre-trained encoder.' That only masks one of four similarity labels. The age, sex, and gene pre-trained encoders are still fit on all fMRI data, including the test fold, with their test labels. Even the state encoder likely sees test fMRI features, just without the state label. Those encoders then produce attention scores that are concatenated with the fMRI embedding and fed into the classifier. The 'without snRNA' baseline has no access to these pre-trained encoders. So the 3.39% binary and 26.59% five-class gains may reflect the leakage or the asymmetry, not the biological value of snRNA. The paper's central claim is therefore unestablished.\n\nThere's also a mild circularity: the state-similarity encoder is trained on clinical state and its output is used to predict the same clinical state. If the encoder is restricted to training folds, this is a legitimate feature-engineering choice, but it deserves scrutiny.\n\nMinor: the five-class improvement is reported only as deltas with no baseline values; Table 1 compares against literature results that are not re-run on the same folds; and the attention mechanism is described as both cosine similarity and cross-attention in different places.\n\nWho is this for? People working on imaging transcriptomics or multimodal AD diagnosis will want to know this architecture exists. But no one should take the numerical claims at face value.\n\nRecommendation: send it to peer review with a clear request for major revision. The novel modality pairing and the public code justify referee time. If the authors redo the evaluation with encoders pre-trained on training folds only, and report the five-class baseline, the results may be interesting. In the current form, I would not accept it.","headline":"A genuinely novel snRNA+fMRI architecture whose reported accuracy gains are likely inflated by test-set leakage in the pre-trained encoders; worth reviewing, but not credible as presented.","tokens_in":16232,"tokens_out":4091,"would_cite":false,"duration_ms":37157,"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":"Adding single-cell RNA data to an fMRI-based model improves Alzheimer's diagnosis accuracy by 3.39 percentage points in binary classification and 26.59 points in five-stage severity classification.","keywords":["single-cell imaging transcriptomics","Alzheimer's disease diagnosis","cross-modal contrastive learning","graph neural network","functional MRI","brain region-gene association","snRNA","mixture-of-experts"],"falsifier":"Re-run scBIT with the test fold's fMRI data completely excluded from every pre-training stage; if the reported gains of 3.39% and 26.59% collapse, the improvement is an artifact of test-set statistics rather than the snRNA modality.","tokens_in":15192,"feed_emoji":"🧠","tokens_out":6644,"duration_ms":58604,"temperature":0.7,"pith_summary":"This paper introduces scBIT, a computational pipeline that uses single-nucleus RNA (snRNA) data to improve Alzheimer's disease prediction from resting-state fMRI alone. Its central claim is that adding this second modality, even though the snRNA and fMRI come from different people, raises classification accuracy by 3.39 percentage points in a binary AD-versus-control task and by 26.59 percentage points in a five-stage severity task. This matters because fMRI is widely collected while single-cell transcriptomes carry molecular detail that fMRI cannot see; if the pairing is genuine, existing fMRI-only cohorts could be re-analyzed for more accurate staging and for molecularly informed biomarker discovery. scBIT avoids the need for paired samples by linking the two datasets through age, sex, disease-state, and genetic similarity, and it makes its reasoning traceable by mapping brain regions to cell-type-specific gene subgraphs.","feed_headline":"Single-cell RNA data lifts fMRI Alzheimer's accuracy by 26.6%","feed_subtitle":"Pairing snRNA with fMRI improves five-class AD diagnosis by 26.59% and binary classification by 3.39%.","key_machinery":"The load-bearing machinery is a 'cell bagging' scheme that pools nuclei of the same cell type into gene networks, a self-explainable graph neural network that distills those networks into cell-type prototype subgraphs, and four pre-trained fMRI encoders whose cosine similarities to the subgraph embeddings serve as cross-modal attention scores. A mixture-of-experts layer then learns per-patient weights for the four attention scores and concatenates the fused score with the fMRI embedding for classification. The whole design exists to give the fMRI classifier access to single-cell molecular information without requiring paired multi-modal data.","core_discovery":"The paper's core discovery is that cross-modal attention can transfer information from single-cell gene-expression data into an fMRI classifier without requiring the same individual to have both measurements. Each snRNA sample is converted into 90 prototype gene subgraphs spanning 15 cell types, four fMRI encoders are trained against age, sex, clinical-state, and genetic similarity labels, and the resulting attention scores are fused with brain-region embeddings by a mixture-of-experts layer before diagnosis. In a ten-fold cross-validation study on a public Alzheimer's cohort, the full model reaches 0.958 accuracy for binary AD classification and sharply improves five-class severity classification, with the authors reporting that genetic similarity is the strongest single matching signal and that accuracy rises as more similarity types are included. The authors present this as the first computational model to integrate single-cell RNA data with MRI data for Alzheimer's prediction.","pith_inferences":["Because the test fold's fMRI features participate in pre-training, the reported accuracy gain could partly reflect the encoder learning test-set statistics; a fully held-out fMRI split would be needed to confirm the gain is caused by the snRNA information.","If the leakage concern is resolved, the same unpaired cross-modal matching scheme could be applied to other conditions that have both single-cell atlases and neuroimaging cohorts.","The specific brain-region–gene-subgraph pairs highlighted by the attention weights could be validated as biomarkers by measuring the same genes or regions in independent AD cohorts.","The large five-class improvement suggests that the auxiliary snRNA signal may help refine disease staging even when binary detection is already saturated."],"forward_implications":["Alzheimer's prediction from fMRI alone can be improved using an external single-cell transcriptomic atlas, even when no individual appears in both modalities.","The largest gains appear in five-stage severity classification, suggesting the snRNA signal is especially informative for disease staging rather than mere detection.","Genetic similarity is the dominant matching signal, so genotype-aware sample pairing may matter more than demographic matching for cross-modal transfer.","The learned attention maps give testable hypotheses about which brain regions and cell-type-specific gene subgraphs are implicated in AD.","Existing public fMRI datasets could in principle be re-analyzed with this method to obtain molecularly informed diagnostic scores without new data collection."],"supporting_citations":[{"why":"supplies the single-nucleus RNA expression data used as the auxiliary modality.","marker":"[34]"},{"why":"supplies the fMRI data, demographics, and genetic information for the prediction task.","marker":"[35]"},{"why":"companion description of the same data source used to justify its choice.","marker":"[36]"},{"why":"provides the prototype-based self-explaining graph neural network strategy that generates interpretable gene subgraphs.","marker":"[30]"},{"why":"provides the tree-search algorithm used to project each prototype onto a concrete gene subgraph.","marker":"[31]"},{"why":"provides the SNP filtering and pruning steps that turn raw genotypes into pathway-level features.","marker":"[32]"},{"why":"provides the pathway-activity scoring method used to compute pathway activities from both SNP and snRNA data.","marker":"[33]"},{"why":"provides the mixture-of-experts architecture that fuses the four cross-modal attention scores.","marker":"[29]"}],"fun_headline_variants":["scBIT merges single-cell RNA and fMRI to boost AD diagnosis by 26.6%","Single-cell RNA plus fMRI lifts Alzheimer's five-class accuracy 26.6%","First AI fusion of single-cell RNA and fMRI improves AD diagnosis by 26.6%","Integrating single-cell RNA with fMRI boosts Alzheimer's classification 26.6%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes that masking the disease-state similarity label for the test set is enough to prevent leakage, even though the test fold's fMRI features are included in pre-training.","fun_headline_variants_meta":{"raw":{"variants":["scBIT merges single-cell RNA and fMRI to boost AD diagnosis by 26.6%","Single-cell RNA plus fMRI lifts Alzheimer's five-class accuracy 26.6%","First AI fusion of single-cell RNA and fMRI improves AD diagnosis by 26.6%","Integrating single-cell RNA with fMRI boosts Alzheimer's classification 26.6%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0006,"raw_usage":{"total_tokens":2818,"prompt_tokens":972,"completion_tokens":1846,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":588,"completion_tokens_details":{"reasoning_tokens":1763}},"tokens_in":588,"tokens_out":1846,"duration_ms":13352,"temperature":1.0,"reasoning_tokens":1763,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T11:43:16.052495+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run scBIT with the test fold's fMRI data completely excluded from every pre-training stage; if the reported gains of 3.39% and 26.59% collapse, the improvement is an artifact of test-set statistics rather than the snRNA modality.","supporting_citations":[{"cited_title":"Protgnn: Towards self -explaining graph neural networks","cited_arxiv_id":null,"evidence_quote":"supplies the single-nucleus RNA expression data used as the auxiliary modality."},{"cited_title":"A survey of monte carlo tree search methods,","cited_arxiv_id":null,"evidence_quote":"supplies the fMRI data, demographics, and genetic information for the prediction task."},{"cited_title":"PLINK: a tool set for whole -genome association and population-based linkage analyses,","cited_arxiv_id":null,"evidence_quote":"companion description of the same data source used to justify its choice."},{"cited_title":"Analysis of features of Alzheimer’s disease: Detection of early stage from functional brain changes in magnetic resonance images using a finetuned ResNet18 network,","cited_arxiv_id":null,"evidence_quote":"provides the prototype-based self-explaining graph neural network strategy that generates interpretable gene subgraphs."},{"cited_title":"Multimodal data fusion of deep learning and dynamic functional connectivity features to predict Alzheimer’s disease progression","cited_arxiv_id":null,"evidence_quote":"provides the tree-search algorithm used to project each prototype onto a concrete gene subgraph."},{"cited_title":"Classification of Alzheimer’s Disease and Mild - Cognitive Impairment Base on High -Order Dynamic Functional Connectivity at Different Frequency Band,","cited_arxiv_id":null,"evidence_quote":"provides the SNP filtering and pruning steps that turn raw genotypes into pathway-level features."},{"cited_title":"Mixture -of-experts with expert choice routing,","cited_arxiv_id":null,"evidence_quote":"provides the pathway-activity scoring method used to compute pathway activities from both SNP and snRNA data."},{"cited_title":"Improved brain age estimation with slice - based set networks","cited_arxiv_id":null,"evidence_quote":"provides the mixture-of-experts architecture that fuses the four cross-modal attention scores."}],"review_version":1}