{"id":"44e63566-43f5-4b59-a2f6-966c2e6d8799","arxiv_id":"2608.08926","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"GeneFuse fuses genomic language model embeddings with MRI via multi-scale FiLM modulation and uncertainty-gated residuals, improving dementia screening AUROC by 0.12 over image-only on the ADNI cohort.","lead":"This paper presents GeneFuse, a method that combines DNA sequence representations from a genomic language model with brain MRI features to help diagnose Alzheimer's disease and mild cognitive impairment. In tests on 182 people from the ADNI dataset, it reached AUROCs of 0.77 and 0.83, slightly better than using images alone or simpler fusion baselines.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The GLM's added value over imaging may be an artifact of comparing one SNP against a 1024-bp window containing multiple linked variants; the claim that sequence context helps is not yet established.","rationale":"The reader identified the load-bearing premise: frozen NT-v2 embeddings of a 1024 bp APOE window must carry AD-relevant signal beyond variant allele counts. My stress-test agrees but sharpens this into a specific confound in the experimental comparison. The Image + SNP scalar baseline in Table 1 encodes genetic information as hard-coded labels, but the GLM window encompasses multiple variants in the APOE region, including both APOE-defining SNPs. Thus the observed improvement over Image + SNP scalar is consistent with a simpler explanation: the GLM representation includes more alleles, not richer sequence context. This is a correctness risk for the paper's mechanistic claim, not merely for the empirical AUROC. The proposed multi-SNP control and background-shuffle control would settle whether the frozen GLM's nucleotide context adds anything beyond allele identity. I do not recommend changing the verdict because the reader's CONDITIONAL assessment already captures the need for stronger evidence; this concern makes the required evidence more specific. The text/table mismatch on significance (Section 3 states both improvements are significant, but Table 1 marks only NC vs AD with a dagger) remains a separate reporting issue that should be fixed regardless.","tokens_in":8051,"tokens_out":3299,"duration_ms":34619,"concrete_test":"Recompute Table 1 with an ablated genomic encoder that maps the same 1024-bp window to a fixed-length vector of allele dosages for every polymorphic site in the window (or a one-layer MLP on those dosages), keeping GCFM, U-GRF, and all experimental settings identical. If this multi-SNP control matches the NT-v2 AUROC within error bars, the additional signal attributed to GLM sequence context is actually multi-variant coverage. In addition, run a background-shuffle control that randomly permutes non-variant nucleotide positions while preserving all variant alleles; if NT-v2 performance is unchanged, frozen GLM context contributes nothing beyond allele identity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.1 encodes a 1024 bp patient-specific window centered on the target variant with frozen NT-v2 and compares against 'Image + SNP scalar' in Table 1. For the APOE locus, the two defining variants rs429358 and rs7412 lie within roughly 400 bp, so both fall inside the window, and the window can contain additional nearby polymorphic sites. The GLM embedding can therefore separate patients by a multi-SNP haplotype, whereas the scalar baseline encodes only one variant call. The reported AUROC gain (0.77 vs 0.70 for NC vs MCI; 0.83 vs 0.73 for NC vs AD) may reflect adding more variant alleles rather than nucleotide context learned by the frozen GLM. The central claim that 'GLM-derived genomic embeddings provide additional information to imaging' depends on ruling out this confound. A second, related issue is that if the non-polymorphic positions of the window are irrelevant, a frozen language model contributes nothing beyond allele identity, and the proposed 'preserves nucleotide context' mechanism is unsupported. The ablation and GLM-comparison results do not resolve this, because all GLM baselines use sequence windows of comparable length and therefore inherit the same multi-variant coverage.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GeneFuse, a multimodal classifier that combines 3D T1w MRI features from a 3D TransUNet with genomic embeddings from the frozen Nucleotide Transformer v2 (NT-v2). For each subject, a 1024 bp patient-specific window centered on an AD risk variant (APOE in the main analysis) is encoded by the GLM; a Genotype-Conditioned Feature Modulation (GCFM) block applies channel-wise FiLM conditioning to multi-scale imaging features, and an Uncertainty-aware Genomic Residual Fusion (U-GRF) module gates the genomic contribution using imaging-derived predictive entropy. The authors evaluate on ADNI (N=182; 52 NC, 53 MCI, 77 AD) for NC vs MCI and NC vs AD, reporting AUROCs of 0.77 and 0.83 for the full APOE-based GeneFuse, compared with 0.69 and 0.71 for image-only, and claim that GLM-derived genomic embeddings provide additional information to imaging. The paper also ablates GCFM and U-GRF, compares several GLM encoders, and reports multi-locus results.","tokens_in":8373,"tokens_out":7292,"duration_ms":70287,"significance":"Strengths: the architecture is clearly specified; subject-level out-of-fold predictions with paired DeLong tests are the right protocol for correlated AUROC comparison; the use of a frozen pre-trained GLM with no fine-tuning makes the comparison to scalar SNP encoding interpretable in principle; ablations of GCFM insertion scales and GLM backbones are informative. I agree with the reader that the evaluation is not circular: the frozen GLM and out-of-fold predictions make the comparison legitimate in principle. If the reported effect is real, the paper would be among the first to show that pre-trained genomic language model embeddings of non-coding sequence context can add diagnostic signal beyond SNP allele counts when fused with volumetric imaging. The uncertainty gating idea is novel and clinically sensible. However, because the central comparison is potentially confounded by multi-variant coverage in the sequence window, and because only one of the two headline significance claims is supported by Table 1, the contribution is not yet established with the confidence required for publication.","major_comments":[{"comment":"The text in Section 3 states that GeneFuse's AUROC improvements over the image-only backbone 'are statistically significant under paired DeLong tests on subject-level out-of-fold predictions (p < 0.05).' However, Table 1 places the dagger marker only on the NC vs AD row of GeneFuse (Full); the NC vs MCI row (0.77 vs 0.69) carries no dagger or double-dagger marker. As written, the central abstract claim covers both tasks, so the NC vs MCI significance assertion is contradicted by the manuscript's own table. Please report exact DeLong p-values for all comparisons to image-only and correct the text, or explicitly qualify the NC vs MCI result as a non-significant trend.","section":"Section 3, Table 1"},{"comment":"The headline comparison between GeneFuse and 'Image + SNP scalar' is confounded by the amount of genetic information each input carries. A 1024 bp window centered on the APOE locus can contain both APOE-defining variants (rs429358 and rs7412) as well as nearby polymorphic sites, so the frozen NT-v2 embedding can separate patients by multi-SNP haplotype; the scalar baseline by definition encodes a single variant call. The observed AUROC gain (0.77 vs 0.70 for NC vs MCI; 0.83 vs 0.73 for NC vs AD) may therefore reflect the addition of more allele information rather than nucleotide context learned by the GLM. To make the central claim ('GLM-derived genomic embeddings provide additional information to imaging') load-bearing, add a control that encodes all variants in the same 1024 bp window in an allele-only form (e.g., one-hot or allele counts per polymorphic site) under the same fusion framework, and a control that masks or shuffles non-variant positions in the NT-v2 input. If the allele-only control matches or outperforms the GLM, the conclusion should be revised.","section":"Section 2.1 and Table 1"},{"comment":"The abstract says GeneFuse 'outperforms existing imaging–genetics fusion methods,' but no significance test is reported for comparisons against MADDi, MMDL, stage-wise DNN, Late Concat, Cross-Attention, or Residual in Table 1. Several baselines have overlapping standard deviations with GeneFuse (e.g., NC vs MCI Cross-Attention 0.74±0.07 vs 0.77±0.05; NC vs AD MADDi 0.77±0.12 vs 0.83±0.05), so the mean AUROC differences are not evidently significant. Report DeLong p-values for all pairwise baseline comparisons, or temper the claim to mean-AUROC advantage.","section":"Abstract and Section 3, Table 1"}],"minor_comments":[{"comment":"The genomic tokenization step is underspecified: the paper does not state whether NT-v2 is applied with 6-mer tokenization, whether all 1024 positions are used as a single sequence, or how the forward/reverse strand is chosen; please add these details for reproducibility.","section":"Section 2.1"},{"comment":"The multi-locus comparison reports mean AUROC gains over APOE alone (0.79 vs 0.77; 0.84 vs 0.83) without significance tests; given the small sample, please report DeLong p-values or label the result as exploratory.","section":"Section 3 and Table 3"},{"comment":"Figure 2 is described only qualitatively; please state how many subjects were used, whether the maps are averaged across the cohort and across five folds, and whether the reported periventricular/medial temporal pattern was consistent across subjects.","section":"Section 3, Figure 2"},{"comment":"There is a typo ('adddtional') in the conclusion, and the paper lacks a data/code availability statement; even a brief statement would improve reproducibility.","section":"Section 4"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline: GeneFuse is a real methods contribution, but its central claim about sequence context is not yet supported by the experiments as reported.\n\nWhat is new: the paper adapts FiLM conditioning to multi-scale 3D MRI features using frozen GLM embeddings, and adds an uncertainty-gated residual fusion. That combination is not in the cited baselines, and the evaluation is more careful than much of this literature: subject-level cross-validation, paired DeLong tests on out-of-fold predictions, ablations for each module, and a comparison across several GLM backbones. The descriptions are clear enough to reproduce.\n\nWhere the problems are: first, there is a factual mismatch. The text says both AUROC gains over image-only are statistically significant, but Table 1 marks only the NC vs AD row with a dagger. The NC vs MCI gain (0.77 vs 0.69) is not flagged. That has to be fixed. Second, N=182, single cohort, no external validation, no code or data released. Minor, but it matters for interpreting the absolute numbers.\n\nThird, and most important, the stress-test concern is on target. The 1024 bp window around APOE contains both rs429358 and rs7412 and other linked variants. The scalar baseline encodes one variant only. So the GLM embedding has access to a multi-SNP haplotype while the baseline does not. The AUROC gain could come from adding that allele information, not from any learned nucleotide context. The mechanism claim—that GLM context representations are what help—is not established. This is not a fatal flaw: the fusion architecture could still be useful. But the paper needs a multi-SNP scalar or haplotype baseline, and ideally a window-shift or permutation control, before claiming that context matters.\n\nWho this is for: anyone working on imaging-genetics fusion or GLM-based biomarkers. It deserves review, but with the expectation of a revision that addresses the confound and corrects the significance claim.\n\nMy recommendation: accept for peer review, with the confound and the reporting error as the main conditions.","headline":"Genuinely new fusion architecture that deserves refereeing, but the claim that GLM sequence context adds signal is currently confounded by a multi-SNP baseline gap.","tokens_in":8894,"tokens_out":3370,"would_cite":true,"duration_ms":31334,"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":"A fusion framework that conditions multi-scale brain-MRI features on sequence embeddings of a 1024-base-pair DNA window around APOE improves early cognitive decline and dementia screening over imaging alone, reporting AUROCs of 0.77 and…","keywords":["Alzheimer's disease","genomic language models","imaging-genetics fusion","multimodal learning","APOE","uncertainty gating","feature-wise linear modulation","brain MRI"],"falsifier":"Retrain GeneFuse after randomly shuffling the nucleotide order within each subject's 1024-base-pair window while keeping imaging, labels, and all other settings fixed; if AUROC remains at 0.83 for NC vs AD, the reported gain is not attributable to DNA sequence context. A simpler check is comparing against a model that uses only APOE ε2/ε3/ε4 allele counts; matching 0.83 would mean the extra sequence context is unnecessary.","tokens_in":7827,"feed_emoji":"🧬","tokens_out":14030,"duration_ms":125110,"temperature":0.7,"pith_summary":"The paper argues that representing a patient's DNA as a learned sequence embedding—the kind of representation produced by a genomic language model trained on whole genomes—carries diagnostic signal for Alzheimer's disease that conventional single-letter variant (SNP) genotypes and brain MRI alone miss. GeneFuse feeds a 1024-base-pair window around the APOE gene through a frozen pre-trained Transformer, uses the resulting embedding to recalibrate multi-scale MRI feature maps, and lets the image's own uncertainty decide how much genomic information enters the final classifier. On a cohort of 182 subjects with paired MRI and whole-genome sequencing, the model reaches AUROCs of 0.77 for normal versus mild cognitive impairment and 0.83 for normal versus Alzheimer's disease, compared with 0.69 and 0.71 for image-only and higher than existing imaging-genetics fusion baselines. The authors flag that all experiments use a single cohort, which may limit generalizability.","feed_headline":"DNA language models boost MRI-based dementia screening to AUROC 0.83","feed_subtitle":"Genomic embeddings of a 1024-base-pair APOE window add diagnostic signal beyond MRI and SNP counts for early cognitive decline.","key_machinery":"The argument is carried by two fusion modules plus a frozen genomic encoder. GCFM (Genotype-Conditioned Feature Modulation) adapts feature-wise linear modulation (FiLM) by predicting channel-wise scale $\\gamma^s$ and shift $\\beta^s$ from the genomic embedding and applying $F' = \\gamma \\odot F + \\beta$ to image feature maps at the 1/2, 1/4, and 1/8 encoder scales, so DNA context recalibrates MRI features before global pooling. U-GRF (Uncertainty-aware Genomic Residual Fusion) computes predictive entropy of the image-only logits, maps it together with the image latent to a gate $\\alpha \\in [0,1]$, and forms $z_{\\mathrm{final}} = (1-\\alpha) z_{\\mathrm{img}} + \\alpha z'_{\\mathrm{gene}}$, suppressing genomic input when the image is confident. The genomic embedding itself comes from a frozen pre-trained Nucleotide Transformer v2 (100M) applied to a patient-specific 1024-bp window centered on the risk variant, with mean pooling over loci in multi-locus settings.","core_discovery":"The paper's central claim is that genomic language model (GLM) embeddings provide additional information to imaging for early cognitive decline and dementia screening. In the APOE-centered setting, GeneFuse improves AUROC by 0.08 (NC vs MCI) and 0.12 (NC vs AD) over the image-only backbone, with the gains statistically significant at $p<0.05$ by paired DeLong tests on subject-level out-of-fold predictions; scalar SNP encoding gives only a small gain, and late concatenation, FiLM-style conditioning, cross-attention, and fixed residual fusion all underperform. Extending to multiple AD-risk loci raises AUROC to 0.79 and 0.84, which the authors read as evidence that loci beyond APOE contribute complementary diagnostic information. They conclude that preserving nucleotide context around disease-associated variants, rather than reducing variants to allele counts, is what makes the fusion work.","pith_inferences":["The uncertainty-gating principle is generic: any weak or loosely coupled auxiliary modality could be added this way, letting the confident primary modality decide how much weight the auxiliary signal receives; testing GeneFuse with fluid biomarkers or clinical scores would show whether the mechanism transfers.","The multi-scale conditioning design suggests imaging-genetics fusion is best applied at feature-map level rather than at the final prediction, a principle that could transfer to other pairings of genomic embeddings with volumetric imaging, such as cardiac or oncological MRI.","Because the cohort is small and single-source, external replication on another MRI and whole-genome-sequencing cohort using the same frozen encoder would show whether the AUROC gains transfer before any clinical use."],"forward_implications":["If GeneFuse's result is right, pre-trained genomic language models can be used as parameter-free encoders that add signal to MRI-based diagnosis without needing to fine-tune the DNA model on medical data.","The comparison with scalar SNP encoding implies that nucleotide context around risk variants contains phenotype-relevant information beyond allele dosage, so future imaging-genetics pipelines may replace hard-coded genotype labels with sequence windows.","The ablation results indicate that conditioning multiple encoder scales beats single-scale or late conditioning, so fusion should happen at the feature-map level rather than only at the final latent.","The uncertainty-gated residual design implies that genomic features should be weighted by the confidence of the stronger modality, not by a fixed rule, when the genotype-phenotype link is loose.","Multi-locus integration improving over APOE alone suggests the same framework generalizes beyond a single well-established risk gene to collections of risk loci."],"supporting_citations":[{"why":"This reference supplies the frozen Nucleotide Transformer v2 backbone that produces the sequence embeddings used by both fusion modules.","marker":"[3]"},{"why":"This reference provides the FiLM conditioning operation that GCFM extends from final-latent conditioning to multi-scale feature maps.","marker":"[15]"},{"why":"This reference establishes APOE as a well-known late-onset Alzheimer's risk locus, motivating the APOE-centered experimental setting.","marker":"[10]"},{"why":"This reference provides the paired MRI and whole-genome sequencing cohort on which all reported AUROCs are measured.","marker":"[7]"},{"why":"This reference supplies the variant filtering, phasing, and personal-genome construction pipeline and a prior genomic-language-model approach to imaging genetics that GeneFuse builds on.","marker":"[11]"},{"why":"This reference provides the 3D TransUNet image encoder whose multi-scale feature maps GCFM modulates.","marker":"[2]"},{"why":"This reference is a multimodal attention-based Alzheimer's diagnosis method used as a comparison baseline.","marker":"[6]"},{"why":"This reference is a multimodal deep learning baseline for early Alzheimer's detection that GeneFuse is compared against.","marker":"[19]"},{"why":"This reference is a stage-wise deep neural network baseline for dementia diagnosis used for comparison.","marker":"[21]"}],"fun_headline_variants":["GeneFuse: DNA language models + MRI score 0.83 AUC for dementia","Genomic language models outscore SNP counts in imaging-genetics fusion","MRI plus DNA language embeddings improves early dementia detection","Nucleotide context, not allele counts, boosts imaging-genetics diagnosis","Fusing DNA language models with neuroimaging hits 0.83 AUC"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The result depends on the assumption that the order of DNA bases in a 1024-base-pair window around APOE, as read by a pre-trained DNA language model, carries Alzheimer's-relevant information beyond the known risk variants themselves, and that this information transfers to a small single-cohort sample without fine-tuning.","fun_headline_variants_meta":{"raw":{"variants":["GeneFuse: DNA language models + MRI score 0.83 AUC for dementia","Genomic language models outscore SNP counts in imaging-genetics fusion","MRI plus DNA language embeddings improves early dementia detection","Nucleotide context, not allele counts, boosts imaging-genetics diagnosis","Fusing DNA language models with neuroimaging hits 0.83 AUC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000242,"raw_usage":{"total_tokens":1533,"prompt_tokens":959,"completion_tokens":574,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":575,"completion_tokens_details":{"reasoning_tokens":482}},"tokens_in":575,"tokens_out":574,"duration_ms":6656,"temperature":1.0,"reasoning_tokens":482,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:20:06.953726+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain GeneFuse after randomly shuffling the nucleotide order within each subject's 1024-base-pair window while keeping imaging, labels, and all other settings fixed; if AUROC remains at 0.83 for NC vs AD, the reported gain is not attributable to DNA sequence context. A simpler check is comparing against a model that uses only APOE ε2/ε3/ε4 allele counts; matching 0.83 would mean the extra sequence context is unnecessary.","supporting_citations":[{"cited_title":"Nature Methods22(2), 287–297 (2025)","cited_arxiv_id":null,"evidence_quote":"This reference supplies the frozen Nucleotide Transformer v2 backbone that produces the sequence embeddings used by both fusion modules."},{"cited_title":"Nature Reviews Neurology9(2), 106–118 (2013)","cited_arxiv_id":null,"evidence_quote":"This reference establishes APOE as a well-known late-onset Alzheimer's risk locus, motivating the APOE-centered experimental setting."},{"cited_title":"medRxiv pp","cited_arxiv_id":null,"evidence_quote":"This reference supplies the variant filtering, phasing, and personal-genome construction pipeline and a prior genomic-language-model approach to imaging genetics that GeneFuse builds on."},{"cited_title":"Journal of the American Medical Informatics Association29(12), 2014–2022 (2022)","cited_arxiv_id":null,"evidence_quote":"This reference is a multimodal attention-based Alzheimer's diagnosis method used as a comparison baseline."},{"cited_title":"Scientific Reports 11(1), 3254 (2021)","cited_arxiv_id":null,"evidence_quote":"This reference is a multimodal deep learning baseline for early Alzheimer's detection that GeneFuse is compared against."},{"cited_title":"Human Brain Mapping40(3), 1001–1016 (2019)","cited_arxiv_id":null,"evidence_quote":"This reference is a stage-wise deep neural network baseline for dementia diagnosis used for comparison."}],"review_version":1}