{"id":"39017472-31e1-4d88-a4a5-062a572e3e67","arxiv_id":"2608.08148","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A central-dogma-guided Transformer-MoE with directed attention and masked reconstruction improves cross-cohort multi-omics performance in cancer subtyping, metastasis detection, and survival prediction.","lead":"DoGMA is a machine-learning model that makes DNA, RNA, and protein data communicate in the direction biology does, from genes to proteins. The authors report better cancer subtyping, metastasis detection, and survival prediction than several existing multi-omics models, and argue that this directional prior helps models transfer across hospitals and incomplete datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cancer-type label in supervised contrastive pretraining confounds attribution of gains to central-dogma prior; no ablation isolates this label, so the central claim is under-determined.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the 32-category cancer-type label in the supervised contrastive loss (Eq. 10) is a confound that no ablation removes. This is the most critical point because the paper's central claim is about the causal role of the central-dogma prior in producing robust transfer. The reported representation-geometry improvement, the cross-omics reconstruction gains, and the downstream transfer improvements are all measured from a model pretrained with this label. The 'w/o Contra' ablation removes the entire contrastive term, conflating the label with the contrastive mechanism, and the 'w/o Adv-GRL' ablation keeps the label. Neither can distinguish whether the benefit comes from the directional prior or from the cancer-type supervision itself. The DDI and edge-intervention analyses are thoughtful, but they measure whether the model uses dogma-consistent edges after training; they do not establish that those edges are what cause the downstream gains, because the supervised label could independently drive the representation organization that the attention then exploits. The paper is otherwise technically sound: the directionality prior is clearly defined, the reconstruction objectives are reasonable, and the experiments are extensive. However, without a label-free contrastive ablation, the central claim remains underdetermined. The reader's CONDITIONAL verdict already asks for exactly this ablation, so no change to the verdict is needed. The proposed concrete test, replacing the cancer-type label with random or unsupervised labels while keeping all other components fixed, would settle whether the central-dogma prior, rather than the cancer-type supervision, is the active ingredient.","tokens_in":21319,"tokens_out":6127,"duration_ms":64199,"concrete_test":"Retrain the full DoGMA pretraining pipeline with the cancer-type label in L_supcon replaced by (i) a random label per sample or (ii) an unsupervised contrastive loss (e.g., SimCLR on input-feature augmentations), keeping L_mask, L_cross, L_adv, directed attention, MoE, and all hyperparameters identical. Then compute the pretrained silhouette score, DDI, METABRIC average test accuracy, and COAD/READ C-index for this variant. If these metrics fall to the level of the 'w/o Contra' or 'Standard Attn.' rows in Table 2, while DDI remains high, the cancer-type label is the primary driver of the reported gains and the central-dogma attribution is not supported. If metrics remain comparable to Full DoGMA, the label is not a confound and the central-dogma prior is the operative ingredient.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that central-dogma-guided directed attention and reconstruction-based pretraining produce robust transfer. However, the pretraining objective in Eq. (10) includes a supervised contrastive loss L_supcon whose labels are a 32-category cancer-type label (Methods: 'Cancer type was encoded as a 32-category semantic label during representation learning'). This label explicitly encourages same-cancer samples to cluster in the latent space. The headline representation-geometry result (silhouette 0.368 to 0.628, Figure 2) is therefore exactly the kind of improvement a supervised cancer-type contrastive loss is designed to produce; it cannot be attributed to the central-dogma prior without an ablation that removes only this label. The existing ablations do not isolate it: 'w/o Contra' removes the entire contrastive term (label plus mechanism), and 'w/o Adv-GRL' retains the label while removing the adversarial term. No variant trains the same architecture and losses with the same contrastive mechanism but with labels other than cancer type (e.g., unsupervised or batch-domain contrastive). Because the cancer-type label is a strong, phenotype-correlated supervision signal available only during pretraining, it could independently organize the representation space and drive both the silhouette gain and downstream transfer (e.g., METABRIC test accuracy 0.955) by making each cancer type's molecular state more separable. If so, the paper's conclusion that 'domain-specific inductive bias' (the central dogma) is the key to robust transfer would be unsupported; the effective prior could instead be cancer-type supervision. The absence of a label-removed ablation is a load-bearing gap because it is the minimal experiment needed to distinguish these explanations.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces DoGMA, a Transformer-Mixture-of-Experts foundation model for pan-cancer multi-omics analysis. Its central design idea is to impose a directional attention bias that follows the central dogma of molecular biology (DNA to RNA to protein, with weak feedback), and to pretrain with masked hierarchical reconstruction, cross-omics reconstruction, adversarial alignment, and supervised contrastive learning. Downstream evaluations include pan-cancer representation geometry, cross-cohort breast cancer subtyping from TCGA to METABRIC, metastasis prediction on a TCGA MetaCancer cohort, and survival analysis on an institutional COAD/READ cohort. The paper reports improved silhouette scores, an average METABRIC test accuracy of 0.955, a metastasis F1 of 0.975, and better C-index values, together with ablations and attention-flow analyses, and attributes the gains to a synergy between the directed attention prior and reconstruction-based pretraining.","tokens_in":21453,"tokens_out":4941,"duration_ms":52191,"significance":"If the empirical claims hold, the paper offers a practically useful demonstration that a biologically motivated directional bias can improve transfer in multi-omics foundation models. The evaluation has notable strengths: five-seed downstream runs, case-level splitting in the metastasis cohort, an external METABRIC cohort that was held out from model development, parameter-matched ablation variants, and an attempted mechanistic analysis through DDI and edge interventions. However, the central attribution is underdetermined because pretraining includes a strong cancer-type semantic label whose contribution is not isolated. In addition, two of the three headline transfer results rest on a small test set and a private cohort, limiting independent verification. With the missing ablation and additional statistical evidence, the contribution would be significant for the multi-omics representation learning community.","major_comments":[{"comment":"The central claim that the gains come from the central-dogma prior is not supported, because the pretraining objective in Eq. (10) includes L_supcon, a supervised contrastive loss using a 32-category cancer-type semantic label (Methods: 'Cancer type was encoded as a 32-category semantic label during representation learning'). A supervised contrastive loss is specifically designed to cluster same-label samples and separate different labels, so the silhouette gain from 0.368 to 0.628 in Figure 2 and part of the downstream transfer gain in Table 1 (average test accuracy 0.955) could be driven by this label rather than by the directed attention prior. The existing ablations do not isolate the label: 'w/o Contra' removes the entire contrastive term, confounding label supervision with the contrastive mechanism, and 'w/o Adv-GRL' retains the label while removing only the adversarial term. The paper needs an ablation that removes only the cancer-type label, for example by replacing L_supcon with an unsupervised or batch-level contrastive objective while keeping all other components unchanged; without such an ablation, the attribution of the representation-geometry and transfer gains to the central-dogma prior is underdetermined.","section":"Eq. (10) and Table 2"},{"comment":"The DDI analysis and counterfactual edge intervention are presented as evidence that DoGMA 'functionally relies' on prior-consistent routes, but the link is weaker than stated. DDI is computed from the same edge taxonomy that defines the directional prior, and the variants in Table 3 differ in ways that directly change how much attention can flow along those edges (standard attention, scratch training, removing reconstruction objectives), so the positive association between DDI and performance is partly by construction and is based on only five points without confidence intervals. The edge-knockout result in Table 4 is suggestive, but applying a hard test-time mask to every attention block can induce a large distribution shift relative to the training-time computation, so the performance drop does not by itself establish learned mechanistic reliance. Please provide control interventions with random edge sets of matched cardinality beyond what is reported, bootstrapped confidence intervals for DDI, and ideally a control model with a shuffled topology.","section":"Attention-flow Alignment, Tables 3 and 4"},{"comment":"Two of the three headline external-transfer results rest on limited or inaccessible test sets. The metastasis test set contains 79 specimens from 64 unique cases (Table 7); Figure 4 reports means over five seeds but no confidence intervals or significance tests against the baselines, and with this sample size the reported F1 gap (0.975 for pretrained DoGMA versus 0.916 for pretrained TMO-Net) may not be stable. The survival experiment uses a private institutional cohort whose data and partition manifests will be released only upon acceptance, so reviewers cannot verify the survival preprocessing, endpoint definitions, or the Cox results. Please report case-level bootstrapped confidence intervals and significance tests for the metastasis metrics, and either release the survival cohort or validate on a public colorectal cancer dataset (for example TCGA-COAD/READ with appropriate splits) to make the survival claim independently checkable.","section":"Metastasis Prediction on TCGA; Institutional COAD/READ"}],"minor_comments":[{"comment":"The breast-cancer fine-tuning protocol uses a learning rate of 1e-5 for pretrained DoGMA but 1e-4 for scratch and TMO-Net baselines; this creates a small confound, and the paper should justify or test sensitivity to this choice.","section":"Hyperparameter Development"},{"comment":"The checkpoint interpolation described in the Pretraining Configuration is not fully specified: the interpolation coefficient is not reported, so the exact initialization used for downstream adaptation is not reproducible.","section":"Pretraining configuration"},{"comment":"The 'Breast subtype' columns in Table 2 appear to report averages over the four breast-cancer tasks, but the caption does not state this; please clarify to match the 'Avg.' convention used in Table 1.","section":"Table 2 caption"},{"comment":"The text does not state which split (training, validation, or test) is used for the t-SNE visualization and silhouette scores in Figure 2; please specify the split in the caption.","section":"Figure 2"},{"comment":"The in-silico intervention results in Table 5 are described as hypothesis-generating in the appendix, but the main text refers to 'causal-chain analysis'; please include an explicit statement in the main text that these results are model-generated hypotheses, not experimentally validated causal mechanisms.","section":"Table 5 and causal-chain analysis"},{"comment":"The survival C-index results in Figure 5 are shown as distributions, but no confidence intervals or pairwise significance tests are reported; adding these would strengthen the survival claim.","section":"Figure 5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope for the journal, but the missing label ablation is the key issue and should be decisive for the revision. The private survival cohort is a reproducibility concern, though the stated plan to release it upon acceptance is reasonable. The DDI analysis should be toned down or re-framed as descriptive rather than mechanistic evidence. I would not recommend acceptance without the ablation and the additional statistical evidence described in the major comments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: DoGMA is a legitimate architectural contribution—a directed attention bias that encodes the central dogma as a fixed penalty plus learnable residuals, married to masked hierarchical reconstruction. The results are consistent across several tasks, and the mechanistic analyses (DDI, edge knockout/rescue) are genuinely thoughtful. If I worked in multi-omics fusion, I would want this idea in front of me.\n\nThe strongest part is the attention-prior ablation: Standard Attn. with the same pretraining package (including the cancer-type label) performs clearly worse than Full DoGMA, so the directional bias is doing real work. The paper deserves credit for that comparison.\n\nThe soft spots are real but not fatal. The supervised contrastive loss uses a 32-category cancer-type label, and no ablation removes just that label. That means the representation-geometry result (silhouette 0.368→0.628) is partly or largely explained by the label—this is exactly what that loss is designed to do. The paper's language overstates the attribution of the geometry gain to the central-dogma prior. However, the core claim about the attention prior doesn't collapse, because the Standard-Attn. vs Full comparison holds the label fixed. Still, the authors should add an ablation that replaces the cancer-type label with unsupervised or batch-label contrastive learning to cleanly isolate the prior.\n\nOther concerns: the survival cohort is private until acceptance, so the most clinically interesting result is not independently checkable; the metastasis test set is 79 specimens; baseline tuning is under-reported; mosGraphGPT is discussed but not compared; no code or weights released. All are fixable or explainable in revision.\n\nBottom line: the paper is a solid conditional-accept candidate. The mechanism is new, the experiments are broadly consistent, and the internal analyses are above average. It deserves a serious referee. The main requests: release code and the survival dataset, add a label-removed ablation, and tighten the attribution language around the silhouette result.","headline":"A genuinely new directed-attention mechanism for multi-omics with consistent results; the missing cancer-label ablation and private survival cohort weaken the attribution, but the core idea holds up.","tokens_in":22255,"tokens_out":3605,"would_cite":true,"duration_ms":34833,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"DoGMA argues that making multi-omics attention follow the central dogma’s DNA-to-RNA-to-protein direction improves cancer representations and transfer.","keywords":["multi-omics","central dogma","directed attention","foundation model","cancer transfer learning","mixture of experts","masked reconstruction","survival prediction"],"falsifier":"An ablation that removes only the cancer-type semantic label from the supervised contrastive loss, keeping the directional attention and all reconstruction terms, should be run and compared with full DoGMA on METABRIC test accuracy and the Dogma Directionality Index. If the label-free variant loses most of the 0.955 test accuracy or the DDI gap between DoGMA and standard attention, then the label, rather than the directional prior, is carrying the reported gains.","tokens_in":20986,"feed_emoji":"🧬","tokens_out":7448,"duration_ms":70237,"temperature":0.7,"pith_summary":"The paper's central claim is that multi-omics foundation models should not treat molecular modalities as interchangeable channels: information in a cell flows directionally, from DNA-level states to RNA to protein, and a model that encodes this direction as an attention prior will transfer better across cancers, downstream tasks, and missing-modality settings. To test this, DoGMA adds a fixed penalty plus learnable residual bias to cross-omics attention so that reverse-direction edges are discounted, and pretrains with masked self-reconstruction, cross-omics reconstruction, and invariant alignment. In the paper's experiments the pretrained model reaches the best average external-cohort breast-cancer test accuracy (0.955), the best metastasis F1 (0.975), and the best survival C-index among compared methods, while ablations show the gains depend on the combination of the directional prior and reconstruction pretraining. A sympathetic reader would take the paper to show that domain-specific biological inductive bias is a principled route to more reliable multi-omics foundation models.","feed_headline":"Central-dogma attention lifts cancer multi-omics transfer","feed_subtitle":"Biasing attention toward DNA-to-RNA-to-protein flow plus reconstruction pretraining beats bidirectional baselines on external cohorts","key_machinery":"The load-bearing object is the omics-specific residual Central-Dogma attention bias: for each head, logits get an additive term $B_{\\text{dogma}}$ that is a fixed negative constant $A$ on weak-feedback (reverse-direction) edges plus a zero-initialized learnable residual $R$ per omics pair, with a binary topology matrix $T$ marking which source-to-target edges follow the dogma. A second piece is the masked hierarchical reconstruction pretraining: a VQ-VAE tokenizer maps each modality to discrete tokens, and the backbone is trained to reconstruct masked tokens within a modality and to reconstruct each modality's latent from the fused representation, so the model must actually use cross-omics routes to recover missing molecular states. A third piece is the modality-specific Mixture-of-Experts fusion that lets each omics layer be transformed by its own expert with learned routing weights. These pieces together give the model a directional prior, a training pressure to use it, and a mechanism to adapt it per pair and per sample.","core_discovery":"On its own terms, DoGMA establishes that the central dogma can be operationalized as a structural prior for attention: each omics pair gets a directed communication edge, forward (DNA to RNA to protein) edges are favored by a shared negative bias, reverse edges are penalized, and a zero-initialized learnable residual keeps pair-specific flexibility. The paper shows that this prior, when combined with reconstruction objectives that force cross-omics inference, yields representations that separate cancer types more clearly (silhouette 0.628 versus 0.368 for the strongest baseline), reconstruct gene expression from non-expression omics with higher correlation, infer driver mutation status with higher AUC, and transfer to an external breast-cancer cohort with higher average accuracy than bidirectional baselines. The paper also reports that attention-flow directionality, measured by a Dogma Directionality Index, tracks downstream performance, and that inference-time knockout of forward edges is the most damaging intervention while restoring forward edges is the most effective rescue. The intended upshot is that the performance gain is not capacity but biologically consistent information-flow structure learned through pretraining.","pith_inferences":["The paper never ablates the 32-category cancer-type label used in the supervised contrastive loss, so the cleanest test of its attribution would be an ablation that removes only that label; if the silhouette jump and transfer gains mostly vanish, the label rather than the directional prior would carry the result.","The same directed-attention-plus-reconstruction recipe could be applied to other biological or physical systems with an ordering constraint, such as signaling cascades, metabolic pathways, or time-ordered multimodal data, where the directionality is known a priori.","A stronger falsification would come from comparing the Dogma Directionality Index against experimentally validated perturbation effects: if high-DDI models do not predict regulatory responses better than low-DDI models on an independent gene-perturbation benchmark, the directionality metric is not actually tracking biological consistency.","Because the pretraining cohort is TCGA-only, the promised pan-cancer generalization across platforms would be better supported by a second external cohort beyond METABRIC and the institutional colorectal set, testing whether the transferred representations survive platform shifts elsewhere."],"forward_implications":["If the claim is right, future multi-omics architectures should replace symmetric cross-modal attention with topology-aware attention that encodes known biological directionality, rather than leaving interactions fully unconstrained.","Reconstruction-based pretraining becomes a necessary companion to the prior: the ablations show that removing cross-omics reconstruction is the single largest drop in external-cohort accuracy, so the prior only pays off when training forces the model to use it.","Missing modality settings benefit directly: a missing omics layer is treated as an unobserved molecular state that can be inferred through forward routes, which explains robust performance under incomplete assay coverage.","Inference-time edge interventions provide a testable signature: if a model has internalized the prior, knocking out forward edges should degrade prediction more than knocking out reverse or random edges, which is what DoGMA shows.","The Dogma Directionality Index offers a model-level diagnostic that correlates with downstream performance and could be used to audit whether a multi-omics model is relying on biologically plausible information flow."],"supporting_citations":[{"why":"supplies the central-dogma directionality principle that motivates the attention topology.","marker":"Crick 1970"},{"why":"provides the main pretrained multi-omics baseline and the TCGA pan-cancer data archive used for pretraining.","marker":"Wang et al. 2024"},{"why":"supplies the VQ-VAE tokenization that maps each omics into a discrete latent vocabulary.","marker":"Van Den Oord, Vinyals et al. 2017"},{"why":"motivates the masked self-reconstruction objective for robust latent denoising.","marker":"He et al. 2022"},{"why":"provides the supervised contrastive loss used in invariant semantic alignment.","marker":"Khosla et al. 2020"},{"why":"supplies the gradient-reversal adversarial alignment used to suppress cohort and platform artifacts.","marker":"Ganin and Lempitsky 2015"},{"why":"provides the attention-flow quantification method behind the Dogma Directionality Index.","marker":"Abnar and Zuidema 2020"},{"why":"provides the independent METABRIC breast-cancer cohort used for external transfer evaluation.","marker":"Curtis et al. 2012"},{"why":"provides the MetaCancer distant-metastasis cohort and endpoint definitions.","marker":"Albaradei et al. 2021"}],"fun_headline_variants":["Central dogma steers attention for better cancer omics","Biology's one-way flow boosts multi-omics AI","DNA-to-protein attention improves cancer transfer","DoGMA: Directed attention follows life's logic","Central-dogma prior sharpens cancer multi-omics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the performance gains are caused by the central-dogma directionality prior and the reconstruction objectives, not by the 32-category cancer-type label that is included in the supervised contrastive loss during pretraining; the paper provides no ablation that removes only that label.","fun_headline_variants_meta":{"raw":{"variants":["Central dogma steers attention for better cancer omics","Biology's one-way flow boosts multi-omics AI","DNA-to-protein attention improves cancer transfer","DoGMA: Directed attention follows life's logic","Central-dogma prior sharpens cancer multi-omics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000174,"raw_usage":{"total_tokens":1316,"prompt_tokens":1012,"completion_tokens":304,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":628,"completion_tokens_details":{"reasoning_tokens":230}},"tokens_in":628,"tokens_out":304,"duration_ms":3719,"temperature":1.0,"reasoning_tokens":230,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T00:20:22.020337+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"An ablation that removes only the cancer-type semantic label from the supervised contrastive loss, keeping the directional attention and all reconstruction terms, should be run and compared with full DoGMA on METABRIC test accuracy and the Dogma Directionality Index. If the label-free variant loses most of the 0.955 test accuracy or the DDI gap between DoGMA and standard attention, then the label, rather than the directional prior, is carrying the reported gains.","supporting_citations":[{"cited_title":"Quantifyingattentionflow in transformers","cited_arxiv_id":null,"evidence_quote":"provides the attention-flow quantification method behind the Dogma Directionality Index."}],"review_version":1}