{"id":"16da4e7d-461c-418f-9d66-7166467757c3","arxiv_id":"2504.13044","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"The paper labels aging as dissipative and uses gene-expression embeddings from a transformer model to sort genes into conservative and dissipative groups and to track entropy changes across tissues.","lead":"This paper proposes that aging is a dissipative process in which cells drift away from stable states, and it uses a transformer model trained on millions of single cells to measure that drift. A generalist might read it to see whether machine-learned cell embeddings can produce a quantitative cellular aging map with tissue- and gene-level resolution.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The dissipative-gene classification is definitional: high embedding drift defines dissipation, so 'aging is dissipative' is not empirically tested and no actual Hopf decomposition or null model is provided.","rationale":"The reader's REJECT verdict is appropriate, but the sharpest vulnerability is not the age-gap interpretation; it is that the central term 'dissipative' is operationalized as drift, so the headline result is definitional. The paper invokes ergodic theory and the Hopf decomposition, but the mathematical content does no work: Eqs. 10-12 define conservative/dissipative genes by a drift threshold, and Eq. 7 is not derived from the Hopf decomposition of a measured dynamical system. In cross-sectional scRNA-seq data there is no trajectory or map T, only age-labeled snapshots, so non-recurrence cannot be established. An honest empirical claim would need at least a null distribution for D_g(t) under shuffled age labels, or a direct estimate of the wandering set. The abstract/full-text age-group discrepancy (104 vs 171) and the lack of code/data further reduce confidence, but the definitional issue is sufficient for rejection. Because the concern reinforces the reader's verdict, no verdict change is needed.","tokens_in":14119,"tokens_out":3355,"duration_ms":34380,"concrete_test":"Permute the age labels across cells (or train an identical MLM without the age token), recompute D_g(t) from Eq. 11 and the entropy from Eq. 15, and compare the resulting drift distributions, dissipative/conservative assignments, and entropy trajectories with Figs. 4-5. If the permutation control yields comparable drift magnitudes, a similar fraction of genes classified as 'dissipative,' or similar entropy increases, then the dissipative conclusion is an artifact of age-token conditioning rather than a property of biological aging. Report the delta threshold and the distribution overlap, not just point estimates.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that aging is fundamentally a dissipative process. The empirical support consists of (i) classifying genes whose embeddings drift with age as 'dissipative' (Eqs. 10-12, D_g(t)=||E_{g,t}-E_{g,t0}||) and (ii) showing that some genes drift and that masked-token entropy rises with age (Fig. 5c). The problem is that step (i) makes the conclusion true by definition. The Hopf decomposition in §4.7.1 is never instantiated: no transformation T on the phase space is defined, no wandering set is measured, and the decomposition E(t)=E_C(t)+E_D(t) is asserted rather than derived from recurrence properties. 'Dissipative' is simply a relabeling of 'high drift'; with any nonzero drift, G\\G_C is nonempty, so the finding that aging involves dissipative genes is guaranteed in advance. The threshold delta is unspecified, and no null model shows that observed drift exceeds what random token geometry, batch effects, or age-label imbalance would produce. The entropy analysis has the same vulnerability: conditional entropy of a masked token can change with age because of vocabulary frequency or data sparsity even if no biological disorder changes. Because the entire CAM and the dissipative/conservative gene lists depend on this interpretive step, the central claim collapses if the drift is a model artifact.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a 'dissipation theory of aging' in which aging is described as the predominance of dissipative, non-recurrent dynamics in a biological dynamical system, formalized via the Hopf decomposition from ergodic theory. The authors train a transformer-based masked language model on a large single-cell RNA-seq corpus, including age as an input token, and analyze the resulting embeddings to construct a Cellular Aging Map (CAM). They classify genes as conservative or dissipative based on the temporal drift of their embeddings (Eqs. 10-12), report tissue- and cell-type-specific age predictions, and compute Shannon entropy of masked-token predictions conditioned on age (Eq. 15) as a measure of molecular disorder. The central claim is that these analyses demonstrate aging to be a fundamentally dissipative process and provide a quantitative framework for measuring age-related molecular changes.","tokens_in":14358,"tokens_out":4232,"duration_ms":39301,"significance":"If established, a quantitative, data-driven demonstration that aging is fundamentally dissipative would be a significant contribution, potentially unifying disparate aging theories under a dynamical-systems framework. The paper has notable strengths: it leverages an exceptionally large corpus of over 65 million single cells, integrates age as a learnable token into a transformer-based model, and proposes concrete embedding-based metrics (drift, similarity, entropy) that could serve as descriptive tools for aging research. However, as presented, the central claim is not empirically established. The classification of 'dissipative genes' is definitionally tied to embedding drift, the Hopf decomposition is never instantiated with a concrete map or measure, and the entropy and age-gap analyses lack the controls needed to distinguish biological signal from model artifacts. The paper does not provide reproducible code for the new analyses, and no falsifiable predictions are extracted beyond the relabeling of drift as dissipation.","major_comments":[{"comment":"The central conclusion that aging is dissipative is not derived from an independent dynamical-systems quantity; it is imposed by definition. In §4.7.2, conservative genes are defined as those with max_t D_g(t) ≤ δ (Eq. 10), and dissipative genes are simply the complement GD = G \\ GC (Eq. 12), where D_g(t) is the Euclidean drift of the gene embedding from a baseline (Eq. 11). Therefore the finding in §2.4 that 'aging is a dissipative process' because dissipative genes exist is a tautology, guaranteed for any nonzero drift. The Hopf decomposition in §4.7.1 (Eq. 7) is asserted but never instantiated: no transformation T on the embedding space is specified, no measure μ is defined, and no wandering set is identified or measured. Without an actual decomposition into conservative and dissipative parts based on recurrence properties, the theoretical framework does no empirical work.","section":"§4.7.1-§4.7.2, Eqs. (7), (10)-(12)"},{"comment":"The threshold δ in Eq. (10) is left unspecified: the text only says it is 'determined through statistical analysis, such as a percentile of the D_g(t) distribution.' No percentile value, no stability analysis, and no null model are provided. In particular, there is no comparison of the observed drift against what would be expected from random token geometry, age-label imbalance, batch effects, or model training variability. Without such a null model, the specific lists of dissipative and conservative genes (e.g., ALDH3B1, NR2C2, HERPUD1 versus MKRN1, SESN1, ATR in §2.4.1) and the disease-related changes in Fig. 5b lack statistical support.","section":"§4.7.2, Eq. (10), and Fig. 5a-b"},{"comment":"The interpretation of z-scored age gaps as evidence of accelerated or decelerated biological aging is under-justified. The paper states (end of §2.2.1): 'These deviations suggest that the model captures underlying variations in gene expression that reflect processes moving alongside chronological aging.' However, no evidence is presented that the residuals are not dominated by prediction noise, batch effects, or tissue-specific sampling imbalances. The correlation coefficients in Fig. 2 are reported without confidence intervals, and no held-out validation or comparison to a null distribution of age gaps is provided. The nonlinear age-gap dynamics in Fig. 3c are therefore not distinguishably biological rather than model artifacts, which undermines the CAM interpretation.","section":"§2.2.1 and §2.2.3"},{"comment":"The entropy analysis is subject to the same confounding issue as the drift analysis. Eq. (15) computes the Shannon entropy of a masked token's predicted probability distribution conditioned on context that includes the age token. Changes in H(x) across age can arise from vocabulary frequency shifts, data sparsity, or class imbalance across age groups, without any change in biological disorder. The assertion in §2.4.2 that 'higher entropy indicates greater uncertainty and suggests a loss of molecular specificity' is not supported without controlling for these factors. Additionally, the entropy results are shown only for kidney tissue and endothelial cells (Fig. 5c-e), with no confidence intervals, effect sizes, or significance tests.","section":"§2.4.2 and §4.7.3, Eq. (15)"}],"minor_comments":[{"comment":"The number of age groups is inconsistent: the abstract states 104 age groups, while the Results and Methods sections state 171 distinct age groups; please reconcile.","section":"Abstract; §2.1; §4.4"},{"comment":"The phrase 'autounity' appears to be a typo for 'autoimmunity'; please correct.","section":"§2.2.1"},{"comment":"The caption contains 'the different between model's prediction and chronological age labels'; this should be 'the difference between.'","section":"Figure 2 caption"},{"comment":"The notation x_C(t) is used in Eq. (8) without a prior definition, and the sentence 'we shown that the modeling is Lipschitz continuous' contains a grammatical error; please revise.","section":"§4.7.1, Eq. (8)"},{"comment":"Code availability only points to the model from reference [14]; scripts specific to the CAM construction, drift classification, and entropy analysis are not provided, which limits the reproducibility of the paper's quantitative claims.","section":"§5-§6"}],"recommendation":"reject","confidential_remarks":"The central problem is definitional rather than local: classifying genes by embedding drift and then concluding that aging is dissipative is a tautology, and the Hopf decomposition is never operationalized. A revision that reframes the contribution as a descriptive atlas of embedding drift with proper null models, uncertainty quantification, and held-out validation might be suitable for a specialized bioinformatics audience, but the theoretical claim in the title is not supported by the present analysis. I also note the inconsistency in the reported number of age groups and the lack of publicly released code for the new analyses."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Farhan, here's my take on arXiv:2504.13044.\n\nWhat's actually new: the paper applies the authors' previously published multimodal transformer to CellxGene's 65M-cell corpus and generates a Cellular Aging Map – age-gap distributions, tissue and cell-type similarity rankings, gene-level drift measures, and entropy trajectories. That is a lot of descriptive output, and some of it may be useful to people mapping transcriptomic aging, e.g., which tissues cluster as age-associated. The authors also acknowledge interpretability limits and dataset bias in the Discussion. Credit where due: it's an honest extension of their modeling program to a large, public dataset.\n\nThe soft spot is not small. The paper claims to show that aging is fundamentally dissipative, but the empirical support is definitional. Eqs. 10-12 classify genes as dissipative if their embedding drift exceeds an unspecified threshold delta; then aging is 'dissipative' because some genes drift. The Hopf decomposition is invoked in Section 4.7.1 but never instantiated: there is no transformation T, no wandering set, no measure. The E(t)=E_C(t)+E_D(t) decomposition is asserted, not derived from recurrence properties. So the theory is an analogy, not a test. The entropy analysis has the same vulnerability: conditional entropy of masked tokens conditioned on age can rise due to vocabulary sparsity or label imbalance, and no null model or permutation test is provided. There are no confidence intervals, and the abstract says 104 age groups while the results say 171 – a minor but telling inconsistency.\n\nThe age-gap 'accelerated/decelerated aging' interpretation also conflates prediction error with biology; residuals from an age-token-conditioned model are not automatically biological aging rates. This could have been addressed with held-out validation, calibration, or comparison against known aging clocks.\n\nSo my bottom line: as a quantitative scientific claim, the central thesis fails as written. It's a suggestive analogy plus a large descriptive analysis. I would not cite it as evidence for dissipation in aging, but I would send it to referees – a good reviewer could push the authors to either drop the theory framing or actually test it with a null model and a real Hopf decomposition. That would be a better paper. It deserves serious peer review, not desk rejection.","headline":"A large-scale descriptive embedding analysis of aging, but the central dissipative claim is definitionally true rather than empirically discovered.","tokens_in":14902,"tokens_out":2095,"would_cite":false,"duration_ms":18970,"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":"The paper argues that aging is the growing dominance of dissipative forces in cellular dynamics, and that transformer-based gene embeddings can quantify this as drift and rising entropy.","keywords":["aging","dissipative dynamics","Hopf decomposition","single-cell RNA sequencing","transformer embeddings","entropy","cellular aging map","biomarkers of aging"],"falsifier":"Train the same transformer architecture on age labels shuffled across samples or with the age token removed, and check whether the conservative/dissipative gene structure and the entropy increase with age persist; if they do, the dissipation signal is an artifact of the training objective. In parallel, follow a longitudinal cohort with repeated single-cell transcriptomes and test whether the model's age gaps track within-person changes and later health outcomes; if the gaps are uncorrelated with future disease or mortality, the deviations are not biological.","tokens_in":13903,"feed_emoji":"🧬","tokens_out":9770,"duration_ms":88727,"temperature":0.7,"pith_summary":"The paper proposes that aging should be understood as a dissipative dynamical process rather than merely the passage of time: biological systems drift out of their recurrent, homeostatic states, and entropy increases as a result. To make this quantitative, the authors train a transformer-based masked language model on more than 65 million single-cell transcriptomes, encode age as an input token, and read out a Cellular Aging Map from the learned embeddings. They report tissue- and cell-type-specific age gaps, genes whose embeddings either remain stable or drift strongly over the lifespan, and entropy trends that rise with age in kidney tissue and in diseased endothelial cells. If the framework holds, aging becomes a measurable loss of molecular specificity, and the gap between predicted and chronological age becomes a candidate biomarker for accelerated or decelerated aging.","feed_headline":"Aging is dissipation: gene maps reveal it cell by cell","feed_subtitle":"A transformer trained on 65 million cells turns aging into a measurable loss of molecular specificity.","key_machinery":"The central object is the Hopf decomposition, an ergodic-theory partition of the phase space into a conservative (recurrent, measure-preserving) part and a dissipative (non-recurrent, wandering) part, applied to gene embeddings. The transformer-based masked language model's embedding function plays the role of the unknown manifold flow map $H(t)$; because the embeddings are Lipschitz continuous, small changes in state or time produce small changes in embedding, so drift in embedding space is interpreted as drift in the biological manifold. Drift $\\lVert E_{g,t}-E_{g,t_0}\\rVert$ from the baseline age separates conservative from dissipative genes, and entropy of the model's masked-token predictions, conditioned on the age token, is used as the dissipation metric. This machinery converts the abstract conservative/dissipative split into a computable, per-gene, per-tissue quantity.","core_discovery":"The paper's central claim is that aging is a dissipative dynamical process: in the language of ergodic theory, biological dynamics split by the Hopf decomposition into a conservative part (recurrent, measure-preserving) and a dissipative part (non-recurrent, wandering), and aging is the growing dominance of the dissipative part, which pushes the system away from recurrent states and raises entropy. The authors make this measurable by replacing the unknown flow map of the gene-expression manifold with the embedding function of a transformer-based masked language model trained on 65 million single-cell transcriptomes, with age as an input token. In the resulting Cellular Aging Map, they report that predicted molecular age deviates from chronological age in a tissue- and cell-type-specific, nonlinear way; that genes split into conservative (low drift) and dissipative (high drift) classes independently of biological function; and that entropy computed from masked-token predictions rises progressively with age in healthy kidney tissue and in diseased endothelial cells, while healthy endothelial cells show transient entropy spikes. The paper concludes that dissipation, measured as embedding drift and entropy, is a cellular-level signature of aging and that chronological age is an inadequate metric for it.","pith_inferences":["An implication the authors leave implicit: if aging is the growth of the dissipative part of the Hopf decomposition, then interventions that restore recurrence — for example, reprogramming or resetting transcriptional states — should measurably reduce embedding drift; this transforms the framework into a quantitative readout for rejuvenation experiments.","The entropy-spike pattern in healthy endothelial cells suggests that aging in some cell types is episodic rather than monotonic; a natural extension is to test whether these spikes align with known stressors, hormonal cycles, or immune challenges.","The similar conservative/dissipative structure across diverse pathways hints that dissipation is an aggregate dynamical property independent of gene function; one could test this by asking whether the same gene sets classify as dissipative in non-human primate or mouse aging data.","Because the framework depends on distributional training data, its tissue rankings should be validated against longitudinal within-individual measurements; the paper's cross-sectional design cannot by itself prove that its age gaps track individual aging rates."],"forward_implications":["Tissues and cell types can be ranked by aging intensity using similarity to the age token, creating a cellular-resolution aging clock that does not rely on any single molecular marker.","Conservative and dissipative gene sets give candidate biomarkers: dissipative genes are the ones whose molecular context changes most with age, so they are natural targets for age-related functional decline.","Disease changes the classification of specific genes (ATR and SAFB2 become more dissipative, TDP2 more conservative), so the framework can track how pathology rewires aging dynamics.","Entropy provides a scalar readout of aging-related information loss, rising progressively in healthy kidney tissue and in diseased endothelial cells, which could serve as a quantitative endpoint for aging studies.","The nonlinear age-gap trajectory, with acceleration early in life and deceleration after the fourth decade, implies that single linear aging-clock equations will miss stage-specific molecular changes; multi-stage models are needed."],"supporting_citations":[{"why":"Supplies the Hopf decomposition theorem used to split the dynamical system into conservative and dissipative parts, the theoretical basis of the dissipation claim.","marker":"[13]"},{"why":"Provides the previously published transformer-based multimodal model that supplies the embeddings reused as the gene-expression manifold.","marker":"[14]"},{"why":"Supplies the aggregated single-cell RNA-seq dataset of over 65 million cells spanning 171 age groups and 215 tissues, on which the model is trained.","marker":"[34]"},{"why":"Gives the Lipschitz-continuity result for transformer embeddings that justifies treating embedding drift as faithful to state and time drift.","marker":"[21]"},{"why":"Provides the ergodic-theory definitions of conservative (measure-preserving) and dissipative (wandering) dynamics used in the decomposition.","marker":"[7]"},{"why":"States the information-theory-of-aging hypothesis that aging is progressive loss of youthful information, which the paper's entropy metric operationalizes.","marker":"[39]"},{"why":"Documents nonlinear dynamics of multi-omics profiles during human aging, supporting the paper's nonlinear age-gap trajectory.","marker":"[32]"}],"fun_headline_variants":["Aging as dissipation: transformer maps cellular entropy","65M cells show aging is a dissipative process","Cellular aging map: entropy rises with age","Dissipation theory: gene drift signals aging","Aging = loss of molecular specificity, transformer reveals"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise, stated in Section 2.2.1, is that deviations between model-predicted and labeled chronological age reflect real biological aging dynamics rather than prediction noise or batch effects; if that premise fails, the Cellular Aging Map, the conservative/dissipative split, and the entropy trends lose their biological meaning.","fun_headline_variants_meta":{"raw":{"variants":["Aging as dissipation: transformer maps cellular entropy","65M cells show aging is a dissipative process","Cellular aging map: entropy rises with age","Dissipation theory: gene drift signals aging","Aging = loss of molecular specificity, transformer reveals"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000215,"raw_usage":{"total_tokens":1420,"prompt_tokens":927,"completion_tokens":493,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":543,"completion_tokens_details":{"reasoning_tokens":421}},"tokens_in":543,"tokens_out":493,"duration_ms":5160,"temperature":1.0,"reasoning_tokens":421,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:15:53.522513+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same transformer architecture on age labels shuffled across samples or with the age token removed, and check whether the conservative/dissipative gene structure and the entropy increase with age persist; if they do, the dissipation signal is an artifact of the training objective. In parallel, follow a longitudinal cohort with repeated single-cell transcriptomes and test whether the model's age gaps track within-person changes and later health outcomes; if the gaps are uncorrelated with future disease or mortality, the deviations are not biological.","supporting_citations":[{"cited_title":"Ergodentheorie","cited_arxiv_id":null,"evidence_quote":"Supplies the Hopf decomposition theorem used to split the dynamical system into conservative and dissipative parts, the theoretical basis of the dissipation claim."},{"cited_title":"& Edelman, E","cited_arxiv_id":null,"evidence_quote":"Provides the previously published transformer-based multimodal model that supplies the embeddings reused as the gene-expression manifold."},{"cited_title":"& Others Cellxgene: a performant, scalable exploration platform for high dimensional sparse matrices","cited_arxiv_id":null,"evidence_quote":"Supplies the aggregated single-cell RNA-seq dataset of over 65 million cells spanning 171 age groups and 215 tissues, on which the model is trained."},{"cited_title":"& Wang, L","cited_arxiv_id":null,"evidence_quote":"Gives the Lipschitz-continuity result for transformer embeddings that justifies treating embedding drift as faithful to state and time drift."},{"cited_title":"An introduction to ergodic theory","cited_arxiv_id":null,"evidence_quote":"Provides the ergodic-theory definitions of conservative (measure-preserving) and dissipative (wandering) dynamics used in the decomposition."},{"cited_title":"& Sinclair, D","cited_arxiv_id":null,"evidence_quote":"States the information-theory-of-aging hypothesis that aging is progressive loss of youthful information, which the paper's entropy metric operationalizes."},{"cited_title":"& Snyder, M","cited_arxiv_id":null,"evidence_quote":"Documents nonlinear dynamics of multi-omics profiles during human aging, supporting the paper's nonlinear age-gap trajectory."}],"review_version":1}