{"id":"1632dd06-0230-476e-8fc9-4ab770c84344","arxiv_id":"2608.06727","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A biology-aware Mixture-of-Recursions model, which routes gene and pathway tokens to different recursion depths using a graph of biological interactions, reports consistent accuracy and efficiency gains over biology-agnostic adaptive-recursion baselines.","lead":"bioMoR adds gene and pathway interaction knowledge to an adaptive Transformer, so it spends deeper computation only on biologically important tokens. Across single-cell and cancer benchmarks it reports higher accuracy with fewer parameters and less compute than a standard Transformer.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The biological-knowledge matrix B is not restricted to training folds; if B is estimated on all cells, all three injection sites leak held-out expression into training, inflating the reported gains over biology-agnostic MoR.","rationale":"The reader's weakest assumption is exactly the most load-bearing concern I can identify. I read the full manuscript and supplement looking for internal inconsistencies, leakage paths, and unsupported claims. The paper is generally careful: it provides a controlled architecture ladder, a unified five-fold protocol, per-dataset results in the supplement, a Wilcoxon significance test, ablations of each injection site, and a biological case study consistent with known cancer pathways. The abstract's eight-dataset headline is a favorable subset of the full 12-dataset suite, but the supplement discloses the complete table and the significance test, so this is a presentation issue rather than a hidden result. The auxiliary losses in Eq. (9) are not defined and no code or data are released, but those are reproducibility gaps, not proof of an invalid central claim. The single condition that could invalidate the comparison against biology-agnostic MoR is the construction of the biological-knowledge graph. If B is estimated on the full dataset, then test cells influence the fixed graph used in all three injection sites during training, making the evaluation transductive and inflating gains in a way that no amount of internal consistency can fix. The concrete test of fold-local B directly settles whether the claim is valid as stated. Since the manuscript does not currently state that B is fold-local, the appropriate verdict remains conditional pending this check.","tokens_in":27836,"tokens_out":4048,"duration_ms":40159,"concrete_test":"Recompute B separately within each of the five cross-validation folds: for each training split, estimate the inverse-covariance-based partial correlations using only training cells, restricted to markers selected on that same training split, and leave all test cells out of the covariance estimate entirely. Rerun Table 1 and the supplementary 12-dataset suite with this fold-local B. If the average macro-F1 gain over the strongest MoR baseline drops materially below the reported 8.2 points, or if per-dataset wins change (e.g., Xin or STAD), the current results are partly transductive. An even cleaner check is to replace B with a graph estimated on an independent single-cell reference dataset and confirm the gains persist.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that biology-guided routing, not architecture alone, drives the large gains over biology-agnostic MoR. A load-bearing condition for that claim is that the fixed graph A = D^{-1}B used in Sites 1-3 is constructed without held-out information. Section 'Biological-knowledge construction' says B entries are partial correlations computed from the inverse covariance of the gene-expression matrix and restricted to the selected markers, but it never states whether this computation is performed on training folds only. Appendix C explicitly says 'per-gene z-scoring fit on the train fold only,' yet is silent on B. Because A enters embedding smoothing (Eq. 6), attention bias (Eq. 7), and the router (Eq. 8), a B estimated on the full dataset including test cells makes the evaluation transductive: the router can exploit test-cell co-expression structure during training, and the three biological-injection sites are confounded with direct access to held-out expression. The five-fold protocol does not resolve this unless B is recomputed per fold. The magnitude of the risk is visible in the supplement: the largest gains are on T-cell, Baron, and Oesophagus, exactly the datasets where a full-data covariance graph could most help. If B is fold-local or derived from an independent reference, the concern disappears; if not, the headline 8.2/7.1 percentage-point improvements over the strongest biology-agnostic MoR baseline are not a valid estimate of biology-guided routing under held-out evaluation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes bioMoR, a Mixture-of-Recursions (MoR) framework for genomic classification in which a fixed biological-knowledge graph (gene co-expression for single-cell data, Reactome pathway relations for multi-omics data) is injected at three sites: embedding smoothing, attention bias, and router-score computation. A shared recursive Transformer block is reused for up to K steps, and a biology-aware router decides each token's recursion depth under expert-choice or token-choice policies. The paper reports that, across eight benchmarks under five-fold cross-validation, bioMoR improves average macro-F1 by 8.2 points and balanced accuracy by 7.1 points over the strongest biology-agnostic MoR baseline, while using 75% fewer parameters and up to 58% fewer FLOPs than a non-recursive Transformer. The supplement adds four further datasets, a paired Wilcoxon significance test, survival endpoints, ablations, and a pathway-prioritization case study.","tokens_in":28027,"tokens_out":10791,"duration_ms":87358,"significance":"If the central claim holds, this is a solid contribution: it demonstrates that structured biological priors can serve as a control signal for adaptive per-token computation rather than only as static feature enrichments. The controlled architecture ladder (Vanilla, Recursive, MoR, bioMoR), the unified five-fold protocol, fixed hyperparameters by modality, and the supplementary 12-dataset paired Wilcoxon test are methodological strengths. The paper also honestly reports the negative Xin delta in the supplement. The main obstacle is the construction protocol for the single-cell biological-knowledge matrix B: the manuscript never states whether B is estimated on the training folds only or on all cells, and all three injection sites depend on A = D^{-1}B. This must be resolved before the headline claims can be accepted as valid estimates of biology-guided routing.","major_comments":[{"comment":"The manuscript never states whether the single-cell biological-knowledge matrix B is estimated on the training folds only or on the full dataset including held-out test cells. Appendix C explicitly says that per-gene z-scoring is fit on the train fold only, but it is silent on B. Because A = D^{-1}B enters all three injection sites (embedding smoothing in Eq. 6, attention bias in Eq. 7, and the router in Eq. 8), a B computed from all cells would make the five-fold evaluation transductive: the router and the attention/embedding injections could exploit test-cell co-expression structure during training. In that case, the reported 8.2/7.1 percentage-point gains over biology-agnostic MoR would not be an unbiased estimate of the benefit of biology-guided routing. Please state the fold-local construction of B, or recompute B per fold from training data only and report whether the conclusions change.","section":"Section 3, 'Biological-knowledge construction', Eq. (5)"},{"comment":"The headline 8.2/7.1 percentage-point improvements are averages over the eight datasets shown in Table 1, which exclude Oesophagus, Baron, Xin, and STAD. The supplement's per-dataset deltas (Table 5) show these four excluded datasets have mixed results, with Xin at -1.3 for macro-F1. The supplement's own 'Avg' column in Tables 3 and 4 appears to reproduce the 8-dataset average rather than the mean of the 12 displayed columns, so the full-suite average is not directly readable from the tables. Please report the complete 12-dataset average explicitly (or state the selection rule for the eight benchmarks), and correct the supplement's Avg column so it matches the displayed columns.","section":"Abstract and Section 5, RQ1 (Table 1)"},{"comment":"The marker-token queries are trained as soft weighted mixtures of genes, but the text says that at test time each query collapses to its top-1 gene. This is a train/test mismatch that applies to all models in the ladder, so it is not the source of the bioMoR-vs-MoR difference, but it could affect the absolute numbers and the interpretability claims. Please report whether all evaluations use the hard top-1 collapse, and ideally include an ablation comparing soft-mixture and hard-collapse tokenization at test time.","section":"Section 3, 'Genomic Tokenization', Eq. (2)"}],"minor_comments":[{"comment":"Figure 7 states that the router sends the top 75% of active pathways to the next depth at each recursion, but Appendix L reports a capacity schedule of (1.00, 0.75, 0.5625, 0.5625). Please reconcile these descriptions.","section":"Figure 7 and Appendix L"},{"comment":"The table header says 'all 11 classification datasets' but the table contains 12 dataset columns; the count should be corrected.","section":"Supplementary Table 3"},{"comment":"The survival-prediction comparison is presented only as boxplots without numerical C-index values in the main text; please report the mean C-index and standard deviation for each model and cohort, at least in the supplement.","section":"Section 5, RQ1 and Figure 2"},{"comment":"The FLOPs formula uses a_t as the mean number of active tokens, but the text also refers to 'the mean number of active tokens at recursion step t'; please clarify whether the reported FLOPs are per-sample averages or per-batch totals, since the formula as written scales with a_t^2.","section":"Section 3, Eq. (10)"}],"recommendation":"major_revision","confidential_remarks":"The central architecture and experimental ladder are sound, and the supplement is unusually honest about per-dataset failures such as Xin. The key question for the revision is the fold-local construction of B; if the authors cannot confirm that B is estimated from training folds only, the headline comparison should be re-run with fold-local B. The mismatched Avg column in the supplement tables should also be corrected before publication. I see no citation-practice concerns; the Genomap and PATH self-citations are appropriate because those are the sources of the biological knowledge used here."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is actually new: no one has combined Mixture-of-Recursions with a fixed biological graph to decide token depth in genomics. The paper does a genuinely controlled architecture ladder (Vanilla to Recursive to MoR to bioMoR) with the same folds and operating point, reports per-dataset numbers in the supplement, and is candid about the Xin loss. The 12-dataset paired Wilcoxon test (p=0.0007) gives real statistical weight. If the reported gains hold, bioMoR is useful for omics-ML.\n\nThe soft spots are proportionate. The one that matters is the biological-knowledge matrix B. Section \"Biological-knowledge construction\" says B contains partial correlations from the inverse covariance of the gene-expression matrix, and A = D^{-1}B feeds all three injection sites (embedding smoothing, attention bias, router). The supplement says per-gene z-scoring is fit on the train fold only, but is silent on whether B is recomputed per fold or estimated on the full dataset. If B is full-data, the router sees test-cell co-expression during training and the headline gains over biology-agnostic MoR are inflated. This is a load-bearing ambiguity, not a minor footnote. It is also easily fixable: state the split, or derive B from an independent reference.\n\nSecond, the abstract's 8.2/7.1 point gains are on the favorable 8-dataset subset. Over the full 12-dataset suite the average gain over the strongest MoR baseline is about 6 points, and Xin loses to Vanilla. That is still a solid result, but the headline should use the complete suite. Third, the auxiliary losses in Eq. (9) are named but never defined, and no code or data are shipped, so independent reproduction would require contacting the authors. Survival results appear as a figure without numeric C-index values. These are clarity and reproducibility issues, not fatal.\n\nThe comparison against biology-agnostic MoR is internally clean, and the case-study pathways (Wnt, EMT, PI3K-Akt) align with known cancer biology. My verdict is conditional: the method is likely useful, but the authors must clarify the graph split, re-report on all 12 datasets, define the losses, and release artifacts before the claim is accepted as stated. Yes, send this to peer review. The idea is novel, the architecture ladder is well-designed, and the main concern is addressable by the authors.","headline":"Biology-guided recursion depth for genomics is a genuinely new idea with a clean controlled comparison, but the paper leaves a load-bearing ambiguity about whether the biological graph avoids test-cell information.","tokens_in":28713,"tokens_out":2846,"would_cite":true,"duration_ms":25405,"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 fixed gene-knowledge graph injected into a shared recursive transformer lifts omics classification by 8.2 macro-F1 points over biology-agnostic adaptive computation.","keywords":["bioMoR","Mixture-of-Recursions","adaptive computation","biological knowledge graph","gene co-expression","pathway tokens","single-cell genomics","multi-omics classification"],"falsifier":"Take any single-cell benchmark, recompute the gene co-expression partial-correlation matrix using only each training fold, re-run the bioMoR expert-choice configuration with all other settings fixed, and compare macro-F1 with the reported numbers. If the margin over biology-agnostic MoR collapses, the headline gain came from test-set information in the biological-knowledge matrix rather than from biology-guided routing.","tokens_in":27457,"feed_emoji":"🧬","tokens_out":7909,"duration_ms":67898,"temperature":0.7,"pith_summary":"bioMoR is a Mixture-of-Recursions (MoR) model for genomic data in which one Transformer block is reused up to K times, and a router decides which gene-marker or pathway tokens continue to the next recursion. The paper's central claim is that biology should drive those routing decisions: a fixed matrix of gene co-expression or curated pathway relations is injected at three points—smoothing token embeddings toward biological neighbors, biasing self-attention toward related tokens, and augmenting the router with a neighborhood-consistency signal. Across eight classification benchmarks, the authors report that this biology-guided recursion improves average macro-F1 by 8.2 percentage points and balanced accuracy by 7.1 percentage points over the strongest biology-agnostic MoR baseline, while using 75 percent fewer parameters than a non-recursive Transformer and up to 58 percent fewer FLOPs. If correct, the result would mean biological structure is useful not only as a representation prior but as a control signal for deciding where adaptive computation is spent.","feed_headline":"Biology-steered recursion lifts omics F1 by 8.2 points","feed_subtitle":"Injecting a gene-knowledge graph into routing, embeddings, and attention beats generic adaptive transformers while using far less compute.","key_machinery":"The load-bearing object is the fixed, row-normalized biological-knowledge matrix $A = D^{-1}B$, where $B$ holds partial correlations among marker genes (from inverse covariance of expression) or undirected relations among curated pathways, and $D$ is the row-sum diagonal matrix. BioMoR applies the same $A$ at three sites: Site 1 smooths tokens as $(1-\\lambda)T + \\lambda A T$ before recursion; Site 2 adds $\\lambda_{\\text{attn}}\\mathbf{1}[A > 0]$ to attention logits; Site 3 augments MoR's routing score with a learned correction computed from $[A H^{(t)},\\, H^{(t)} - A H^{(t)}]$. The machinery is a shared recursive Transformer block $f_\\theta$ reused across depths, with token-specific recursion depth $d_m$ decided by expert-choice or token-choice routing, so depth can vary without adding parameters.","core_discovery":"The paper proposes bioMoR, a Mixture-of-Recursions model in which one Transformer block is reused up to K times and a router decides, at each step, which gene-marker or Reactome pathway tokens continue to the next recursion. Its central discovery is that biology can guide all three places where adaptive computation is decided: a learned mixture smooths token embeddings toward biological neighbors, a binary mask biases self-attention toward biologically related tokens, and a graph-aware router adds a neighborhood-consistency correction to the token's routing score. The knowledge enters as a fixed row-normalized matrix $A = D^{-1}B$ built from gene co-expression partial correlations for single-cell data or curated pathway relations for multi-omics data. The authors report that this biology-guided recursion beats both vanilla Transformers and biology-agnostic MoR across eight classification benchmarks, improves survival-risk concordance on four TCGA cohorts, and allocates deeper recursion to pathways such as Wnt signaling, epithelial-to-mesenchymal transition, and PI3K-Akt in a metastatic-cancer case study. The paper's claim is that biological structure should control which tokens are refined further, not merely which features are selected.","pith_inferences":["Beyond the paper: the single-cell knowledge matrix $B$ is built from the inverse covariance of expression, and the paper does not state whether $B$ is fit on training folds only. Re-fitting $B$ within each cross-validation fold is the decisive check on whether the headline gain is a genuine routing effect or transduction from test-set information.","Beyond the paper: because the router correction starts at zero and is gated by a learned sigmoid, bioMoR defines a smooth interpolation from biology-agnostic to biology-guided routing; the same neighborhood-augmented score could be wrapped around other adaptive-computation mechanisms, not only Mixture-of-Recursions.","Beyond the paper: the binary attention mask for pathways could be replaced by a soft bias proportional to partial correlation or pathway overlap, which might help datasets where the curated graph is dense and a hard mask is overly restrictive.","Beyond the paper: the recursion-depth lists could be used as a marker-discovery tool for new cohorts, treating tokens that consistently reach the deepest levels as candidate disease drivers before downstream validation."],"forward_implications":["If the reported gains hold, biology-agnostic adaptive computation is leaving a substantial accuracy margin unused in genomics: the same shared recursive block, routed by biology, reaches 74.6 average macro-F1 versus 66.4 for the strongest non-bioMoR baseline.","Token-specific recursion depth gives interpretability for free: because each token names a marker gene or curated pathway, routing decisions produce a per-task list of which biological programs received deeper computation, and the paper shows metastatic-cancer pathways such as Wnt signaling, EMT, and PI3K-Akt among the deepest.","The efficiency claim is load-bearing: expert-choice bioMoR at K=2 uses 0.42 times the normalized FLOPs of the vanilla Transformer, so the accuracy gains are not bought simply by spending more compute.","The mechanism transfers to time-to-event endpoints: bioMoR variants are competitive on C-index across four TCGA survival cohorts, suggesting biology-guided recursion is not classification-specific."],"supporting_citations":[{"why":"Defines Mixture-of-Recursions, the shared-weight recursive-block and routing baseline that bioMoR extends with biological knowledge.","marker":"Bae et al. 2025"},{"why":"Supplies the single-cell datasets and the inverse-covariance partial-correlation method used to build the gene co-expression knowledge matrix B.","marker":"Islam and Xing 2023"},{"why":"Supplies the curated pathway relations and pathway memberships used as multi-omics biological knowledge and as pathway tokens.","marker":"Gillespie et al. 2022"},{"why":"Provides the gene-plus-value tokenization scheme that bioMoR uses to form single-cell marker-gene tokens.","marker":"Cui et al. 2024"},{"why":"Supplies the TCGA cohorts that anchor the cancer classification and survival benchmarks.","marker":"The Cancer Genome Atlas Research Network et al. 2013"}],"fun_headline_variants":["bioMoR: biology-guided recursion boosts F1 by 8.2","Gene-guided recursion lifts omics F1 by 8.2 points","Bio-recursion: 8.2-point F1 boost, 75% fewer params","Bio-guided recursion: +8.2 F1, -75% params on omics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper never states whether the matrix of gene or pathway relationships that guides all three biology-aware components is computed from training data only or from the entire dataset including held-out test cells; if it is computed from the entire dataset, the reported accuracy gains would be inflated by information leaking from the test set.","fun_headline_variants_meta":{"raw":{"variants":["bioMoR: biology-guided recursion boosts F1 by 8.2","Gene-guided recursion lifts omics F1 by 8.2 points","Bio-recursion: 8.2-point F1 boost, 75% fewer params","Bio-guided recursion: +8.2 F1, -75% params on omics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001232,"raw_usage":{"total_tokens":5098,"prompt_tokens":1020,"completion_tokens":4078,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":636,"completion_tokens_details":{"reasoning_tokens":3990}},"tokens_in":636,"tokens_out":4078,"duration_ms":25869,"temperature":1.0,"reasoning_tokens":3990,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:14:40.590940+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take any single-cell benchmark, recompute the gene co-expression partial-correlation matrix using only each training fold, re-run the bioMoR expert-choice configuration with all other settings fixed, and compare macro-F1 with the reported numbers. If the margin over biology-agnostic MoR collapses, the headline gain came from test-set information in the biological-knowledge matrix rather than from biology-guided routing.","supporting_citations":[],"review_version":2}