{"id":"4b822200-6198-48aa-b8cb-2177dc3f211d","arxiv_id":"2412.11336","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A mathematical model of enhancer-based gene regulation shows that hierarchical cell identities and progenitor states emerge from competition between enhancers for epigenetic readers, predicting observed blood progenitor states.","lead":"Animals make many cell types through a hierarchy of intermediate progenitor states, but the mechanism behind this hierarchy has been unclear. This paper shows that a mathematical model of enhancer competition, in which regulatory elements compete for shared reader proteins, naturally produces such hierarchies, and it uses the model to predict which blood progenitor states should exist.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Q≈Ξ reduction to symmetric gradient dynamics is not justified even for the paper's own autoregulatory motif: the coarse-grained Q* is the softmax of w, not z, so the persistence-length predictions rest on an unstated parameter assumption.","rationale":"The reader correctly identified the Q≈Ξ reduction as the load-bearing assumption, and my analysis agrees but sharpens it: the paper's own justification in Supplement B.3 does not establish Q≈Ξ even in the idealized autoregulatory motif. The claim that 'Q*_L≈z since it is averaged of all Q_i' is incorrect when Q_i are distinct unit vectors; the average is the softmax of w, which is generally not z. This is not a matter of empirical correlation; it is a structural mismatch unless w_i is chosen to match z_i. Since the whole energy landscape (Eq. 6), the coarse-graining theorem (Eq. 8), and the persistence-length prediction β_L≈μ_L^{-2} are derived only for the symmetric system, the central claim currently lacks a valid derivation for general enhancer networks. A minimal numerical counterexample, as proposed in the concrete test, would settle whether the symmetry reduction is internally consistent. I also note a secondary issue: even granting symmetry, the bound |V_L−V|≤¼βμ_L^2 is pointwise in energy, not a C^1 bound, and is small only for βμ_L^2≪1, well below the claimed destabilization at β_crit≈2|L|μ^{-2}; so the persistence-length prediction for unequal-angle patterns is not rigorously derived. These concerns do not necessarily refute the biological intuition, but they justify keeping the verdict at CONDITIONAL pending a direct test of the symmetric reduction.","tokens_in":42814,"tokens_out":8978,"duration_ms":83121,"concrete_test":"Construct the minimal autoregulatory motif described in Supplement B.3 with two TFs: Ξ_1=Ξ_2=z=(1,0.5), Q_1=(1,0), Q_2=(0,1), w=(0,0). Coarse-graining via Eq. 7 gives Ξ*_L=z and Q*_L=(0.5,0.5), which are not equal. Simulate Eq. 4 with these coarse-grained Q* and Ξ* over β∈[0,20], and compare with the symmetric Eq. 5 using Ξ*=z. If the stable fixed points and bifurcation structure differ, e.g. the progenitor Ξ* is not a fixed point of Eq. 4 at any β, then Q≈Ξ does not follow from the autoregulatory motif and the downstream predictions are unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on replacing Eq. 4 by the symmetric gradient dynamics Eq. 5. The paper's justification (Supplement B.3) is that coarse-graining an autoregulatory motif L (with Ξ_i=z for all i∈L) leaves the dynamics invariant and yields Q*_L≈z, so that Q≈Ξ after coarse-graining. But this is a non-sequitur. In the stated motif each enhancer drives a distinct TF, Q_i=e_i, so Eq. 7 gives Q*_L=Σ_{i∈L} e_i e^{w_i}/Σ e^{w_i} = softmax(w_i), while Ξ*_L=z. For Q*_L≈z one must have softmax(w_i)∝z_i, i.e. the baseline weights w_i must encode the terminal expression program z_i. This is an additional, unstated parameter assumption; nothing in the biology of enhancer competition forces it. If it fails, the coarse-grained dynamics remain non-symmetric, no energy function Eq. 6 exists, and the persistence-length prediction β_L≈μ_L^{-2} (and hence the blood-progenitor ranking) has no derivation. The paper provides no test of this symmetry, and it is not a consequence of the autoregulatory motif itself. A secondary gap: even granting symmetry, the bound |V_L−V|≤¼βμ_L^2 is pointwise in energy and is small only for βμ_L^2≪1, well below the claimed destabilization at β_crit≈2|L|μ^{-2}, so the persistence-length statement for unequal-angle patterns is not rigorously derived.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes that hierarchical cell identity arises as an emergent property of enhancer-based gene regulation in animals. Starting from a biophysical model of enhancer acetylation and competition for shared epigenetic readers, the authors write the general dynamics in Eq. (4) and then approximate them by the symmetric gradient system Eq. (5). They define a coarse-graining transformation Eq. (7) that maps sets of terminal expression patterns to weighted-average 'progenitor patterns,' and they argue that these progenitor patterns are stable attractors over a persistence length scale set by the maximal angle between patterns, β_L ≈ μ_L^{-2}. The framework is then applied to hematopoiesis, where observed progenitor states are claimed to be those with longest persistence lengths, to erythropoietin-mediated fate bias, to cancer-related transcription factor dysregulation, and to differentiation therapy with HDAC inhibitors. The paper also draws a formal analogy between the symmetric dynamics and dense associative memory networks.","tokens_in":43150,"tokens_out":5071,"duration_ms":49480,"significance":"If the central reduction from Eq. (4) to Eq. (5) can be justified, the paper offers a mechanistically grounded and unusually predictive account of multilineage priming: it yields a quantitative ranking of progenitor states from terminal expression patterns, recovers known blood progenitors including recently identified ones, and makes falsifiable predictions about β modulation by HDAC/HAT perturbations. The mathematical derivations in Supplement F (equal-angle β_crit) and Supplement B (coarse-graining transformation) are self-contained and clean, and the paper ships simulation code and uses external Haemopedia expression data for the blood progenitor ranking. The connection to dense associative memories is original and potentially generative. The main risk is that the entire attractor/persistence-length machinery is derived for Eq. (5), so the validity of the Q ≈ Ξ reduction determines whether the biological conclusions follow.","major_comments":[{"comment":"The reduction from the general dynamics Eq. (4) to the symmetric gradient dynamics Eq. (5) is not actually derived. In the autoregulatory motif specified in Supplement B.3, each enhancer drives a distinct TF, Q_i = e_i, while the binding rows are identical, Ξ_i = z. Applying the coarse-graining transformation Eq. (7) gives Q*_L = Σ_{i∈L} e_i e^{w_i}/Σ_{i∈L} e^{w_i} = softmax(w_i), whereas Ξ*_L = z. For Q*_L ≈ z one needs the baseline weights w_i to encode the terminal expression values z_i; this is an additional parameter restriction that is not implied by enhancer co-binding or autoregulation. Since the energy function Eq. (6), the coarse-graining bound, the persistence-length estimate, and the blood progenitor ranking all depend on the symmetric form Eq. (5), this is a load-bearing gap. I recommend either proving the reduction under explicit conditions, testing Q ≈ Ξ directly on the data used for Ξ and on enhancer-to-gene assignment data, or showing that the main predictions are robust to finite Q − Ξ asymmetry.","section":"Model for enhancer-regulated gene expression, Eq. (4)–(5); Supplement B.3"},{"comment":"The persistence-length prediction β_L ≈ μ_L^{-2} for unequal-angle patterns is not justified by the stated bound. The bound |V_L − V| ≤ ¼ β μ_L^2 is small only when β μ_L^2 ≪ 1, but the equal-angle destabilization derived in Supplement F occurs at β_crit ≈ 2|L| μ^{-2}, i.e. at β μ_L^2 ≈ 2|L|. Thus, at the parameter regime where progenitor destabilization is claimed, the pointwise energy comparison is not small. The ranking of progenitor persistence lengths in Figure 6A therefore rests on an extrapolation beyond the regime in which the coarse-graining approximation is proved. A sharper bound or a separate stability analysis around the coarse-grained state at β ~ μ^{-2} is needed to support the central quantitative prediction.","section":"Results, Eq. (8); Supplement B.4.2 and Supplement F"}],"minor_comments":[{"comment":"The phrase 'Q*_L ≈ z since it is averaged of all Qi' is imprecise: the average of the rows Qi equals softmax(w_i), so the approximation holds only under a specific relation between w and z. State that relation explicitly.","section":"Supplement B.3"},{"comment":"There is a typo in the sentence 'all eigenvalues are negative if and only if Eq. 18 holds' — 'anf only if' should read 'if and only if'.","section":"Supplement F"},{"comment":"The caption contains the typo 'correspondign' instead of 'corresponding'.","section":"Figure 5 caption"},{"comment":"The main-text transformation Eq. (7) uses weights e^{w_i}, while the generalization in Supplement B.5 uses e^{w_i + β u_i}. Clarify in the main text which convention is used for the blood progenitor computations.","section":"Eq. (7) and Supplement B.5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's central modeling assumption Q ≈ Ξ is also central to the authors' previous EnhancerNet framework, and the paper uses that same framework both to construct the model and to initialize Ξ from expression data. An independent test of the symmetry assumption, or at least a clear statement of the new mathematical and biological content beyond the earlier paper, would materially strengthen the novelty claim. The core idea is valuable and the blood progenitor comparison is impressive, but the unproved reduction to Eq. (5) should be resolved before publication in a journal with the standards of this one."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: the core mathematical claim, that hierarchical progenitor states arise from symmetric enhancer dynamics, is built on an approximation that the paper does not justify. The coarse-graining transformation itself (Eq. 7) is new and elegant, but the reduction from the general dynamics Eq. 4 to the symmetric gradient system Eq. 5 is where the edifice starts to wobble.\n\nThe paper does well to make the softmax-weighted averaging concrete: it shows that for a block of identical patterns the coarse-grained dynamics are exactly invariant, and for near-identical patterns the energy differs by O(β μ_L^2). That's a real, checkable result. The blood application is a reasonable illustration, but it is not a tight test.\n\nThe soft spot is the one the stress-test flags. In Supplement B.3, the autoregulatory motif is defined by Ξ_i = z for all i in L and Q_i = e_i (each enhancer drives a distinct TF). The paper says \"Q*_L ≈ z since it is averaged of all Qi.\" But Eq. 7 gives Q*_L = softmax(w_i), which equals z only if the baseline weights w encode z. That is an unstated parameter assumption, and without it the coarse-grained dynamics are not symmetric, there is no energy function, and the persistence-length prediction β_L ~ μ_L^{-2} has no derivation. The secondary point is also real: the bound |V_L - V| ≤ (1/4)β μ_L^2 is only small for β μ_L^2 << 1, whereas the claimed destabilization is at β_crit ≈ 2|L| μ^{-2}, so the approximation breaks down before the bifurcation. The persistence-length statement for unequal angles is not rigorously derived.\n\nI'd also push back on the abstract's \"quantitatively predicts lineage relationships.\" The progenitor ranking is one figure, no error bars, no null model. It's suggestive, not quantitative.\n\nThe paper is worth refereeing because the idea is interesting and the math, when pinned to the symmetric case, is clean. But a referee should demand either a proper justification of Q≈Ξ or a clear statement that it's an assumption, and a direct test with statistics. I'd read a revision, I wouldn't cite the current version.","headline":"The central symmetry reduction is unproven, so the persistence-length predictions rest on an unstated parameter assumption; but the coarse-graining idea deserves a referee.","tokens_in":43701,"tokens_out":2765,"would_cite":false,"duration_ms":22773,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["92B05","92C37","37N25"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims cell-identity hierarchies emerge from enhancer competition for shared epigenetic readers, proved via a coarse-graining theorem making weighted averages of terminal expression programs stable progenitor states at…","keywords":["hierarchical cell identity","enhancer competition","epigenetic readers","progenitor states","multilineage priming","coarse-graining transformation","hematopoiesis","differentiation therapy"],"falsifier":"Compute the persistence length $1/\\mu_L^2$ for every subset of the eleven blood lineage programs from an independent, high-resolution expression dataset, then check the ranking claim: every documented progenitor should sit in the top tier of its size class, and every top-ranked subset should be findable by lineage tracing; one observed progenitor with a short persistence length, or one high-ranked subset that exhaustive tracing fails to find, would break the central prediction. A second, independent check targets the $\\beta$ axis: the model predicts that graded HDAC inhibition (raising $\\beta$) should progressively destabilize progenitors and push the population toward terminal states, so a titration experiment in a tissue with known terminal programs that instead freezes or stabilizes an intermediate state would contradict the writer/eraser control claim.","tokens_in":42584,"feed_emoji":"🧬","tokens_out":14274,"duration_ms":119623,"temperature":0.7,"pith_summary":"Hierarchical cell identity — stem, progenitor, terminal — is a signature of animal development, but this paper argues the hierarchy is not wired into the gene-regulatory network; it emerges from how animal enhancers work. The mechanism is enhancer competition: enhancers recruit epigenetic writers that acetylate their histones, and acetylated enhancers then compete for limiting epigenetic reader molecules (such as Brd4) that drive transcription. Because cell-identity enhancers are co-bound by the very genes they activate, the network dynamics reduce to a symmetric 'gradient' form, and the paper proves a coarse-graining theorem: at intermediate competition strength (a parameter $\\beta$ set by the writer/eraser balance), softmax-weighted averages of terminal expression programs, $\\Xi^*_L = \\sum_{i \\in L} \\Xi_i e^{w_i}/\\sum_{i \\in L} e^{w_i}$, become stable attractors — progenitor states — with persistence length of order $\\mu_L^{-2}$, where $\\mu_L$ is the widest angle among the averaged programs. If correct, this one principle reconstructs the observed blood hierarchy: ranking all subsets of the eleven mouse blood lineage programs by persistence length puts the known progenitors on top, explains multilineage priming and signaling biases such as erythropoietin's, and predicts why HDAC inhibitors (which raise $\\beta$) drive cancer cells to differentiate.","feed_headline":"One mechanism predicts every blood progenitor state","feed_subtitle":"Enhancer competition plus shared gene programs makes cell hierarchy emerge — no pre-wired network needed.","key_machinery":"The load-bearing object is the coarse-graining transformation of Eq. 7: for a subset $L$ of enhancer types with weights $w_i$ and patterns $\\Xi_i$, the transformed pattern $\\Xi^*_L = \\sum_{i\\in L} \\Xi_i e^{w_i} / \\sum_{i\\in L} e^{w_i}$ (a softmax-weighted average) together with $w^*_L = \\log \\sum_{k \\in L} e^{w_k}$ rewrites the network so that its dynamics match the original fine-grained dynamics to $O(\\beta \\mu_L^2)$. This transformation is exact when the averaged patterns are identical — the autoregulatory-motif case that justifies the symmetric gradient form $\\dot{x} = \\Xi^T \\mathrm{softmax}(\\beta \\Xi x + w) - x = -\\nabla V$ with $V = -\\beta^{-1}\\log Z + \\tfrac{1}{2}\\|x\\|^2$ — and it is commutative and self-similar under repeated application. It converts the biology of enhancer competition and cooperation into a quantitative statement: a subset of expression programs becomes a stable progenitor exactly when $\\beta$ drops below a threshold of order $\\mu_L^{-2}$, which is the 'persistence length' used to predict which progenitors exist.","core_discovery":"The central claim, stated on the paper's own terms, is that the enhancer-regulated dynamics $\\dot{x} = Q^T \\mathrm{softmax}(\\beta \\Xi x + w) - x$ are self-similar under a specific coarse-graining. For any subset $L$ of enhancer types, replacing the patterns $\\Xi_i$, $i \\in L$, by the weighted average $\\Xi^*_L = \\sum_{i\\in L} \\Xi_i e^{w_i} / \\sum_{i\\in L} e^{w_i}$ and the weights by $w^*_L = \\log \\sum_{i \\in L} e^{w_i}$ reproduces the original dynamics up to a correction of order $\\beta \\mu_L^2$, where $\\mu_L$ is the maximal angle between patterns in $L$; the transformed potential satisfies $V_L \\approx V + O(\\beta \\mu_L^2)$. Hence $\\Xi^*_L$, the 'progenitor pattern', is a stable attractor of the fine-grained network at intermediate $\\beta$, with persistence length $\\sim \\mu_L^{-2}$; for $|L|$ isolated equal-angle patterns the destabilization threshold is $\\beta_{\\mathrm{crit}} \\approx 2|L| \\mu^{-2}$. The authors derive this for the symmetric (gradient) form obtained when autoregulatory enhancer motifs make $Q \\approx \\Xi$, and they show the resulting hierarchy is jointly controlled by $\\beta$ and by the activation vector $w$: raising $w_i$ enlarges pattern $i$'s basin and pulls all its associated progenitors toward it, modularly. Applied to hematopoiesis, persistence lengths computed from the eleven mouse blood lineage programs rank the observed progenitor states above competing subsets, including recently identified progenitors, and annealing simulations with noise and production feedback reproduce balanced blood output that can be biased, as erythropoietin biases toward erythrocytes.","pith_inferences":["The persistence-length rule is derived for the dynamics in general, not for blood specifically, so the same ranking calculation should transfer to other well-mapped hierarchies such as neural crest, thymic epithelium, or intestinal lineages; testing it there is a direct way to probe the claim's generality beyond hematopoiesis.","Because the progenitor pattern is a softmax-weighted average with weights $e^{w_i}$, progenitor identity is graded in the signaling level $w_i$ rather than discrete; a 'bipotential' state is a continuum of weighted mixtures, which may explain why lineage-tracing studies disagree about which progenitors genuinely exist — a reading the paper gestures at but does not develop.","The same machinery could be inverted as a discovery tool: terminal expression programs could be treated as stored patterns and the coarse-graining transformation used to generate candidate progenitor profiles for single-cell validation in tissues where progenitors are not yet catalogued.","The symmetry assumption $Q \\approx \\Xi$ is presented as an approximation; if it fails in real networks, the persistence-length ordering might persist with rescaled effective $\\beta$, but the paper does not test this robustness, so whether the hierarchy prediction survives asymmetric enhancer coupling remains open."],"forward_implications":["In blood formation only the subsets of the eleven lineage programs with the longest persistence lengths should appear as stable progenitors; the paper verifies this ranking against the known catalog of mouse blood progenitors, including the recently identified basophil/megakaryocyte/erythrocyte and eosinophil/neutrophil states.","Enhancer competition becomes a doseable control axis: lowering $\\beta$ stabilizes progenitor states while raising $\\beta$ drives differentiation, which is the paper's mechanism for why HDAC inhibitors act as differentiation therapy and why mutations that blunt $\\beta$ or widen program overlap cause blocked maturation.","Transcription-factor mutations that make two lineage programs more similar shrink $\\mu_L$ and lengthen the persistence of the shared progenitor, explaining how factors such as Lmo2, Tal1, and Erg stabilize mixed stem-differentiated leukemia states.","Signaling inputs $w_i$ bias differentiation modularly: raising one terminal program's weight enlarges its basin and shifts every progenitor containing it toward that fate, without touching unrelated progenitors — the quantitative form of erythropoietin's effect on erythroid output.","Annealing protocols (lower then slowly raise $\\beta$), combined with noise and feedback on $w$, yield controlled transitions between identity states and balanced production across fates, giving a general recipe for steering differentiation in any tissue with known terminal programs."],"supporting_citations":[{"why":"Prior enhancer-selection model of cell-identity dynamics; supplies the dynamical equations, the initialization of the binding matrix from expression profiles, and the reprogramming validations that the present work extends.","marker":"[49]"},{"why":"Mouse blood-lineage gene-expression data used to build the transcription-factor binding matrix and to evaluate persistence-length predictions across all lineage subsets.","marker":"[95]"},{"why":"Reference review of hematopoiesis models providing the catalog of observed progenitor states and overlapping fate potentials against which the predictions are compared.","marker":"[4]"},{"why":"Experimental characterization of p300/CBP-driven enhancer activation and transcriptional pause release that anchors the model's acetylation-eraser-reader mechanism and timescales.","marker":"[48]"},{"why":"Recently identified basophil- and megakaryocyte/erythrocyte-linked progenitor populations used as prospective tests of the predicted progenitor identities.","marker":"[85, 86]"}],"fun_headline_variants":["Enhancer competition alone predicts every blood progenitor","Epigenetic tug-of-war yields cell hierarchy","Math model: enhancer interplay spawns all blood progenitors","Self-similarity in enhancers explains lineage tree","One mechanism: enhancer competition builds cell hierarchy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The assumption the whole argument rests on is that, once the enhancers that are always activated together are merged, the network is effectively symmetric — that is, the binding matrix that says which transcription factors switch on an enhancer is the same as the matrix that says how strongly that enhancer drives each factor's expression, so the dynamics can be written as downhill motion on a single energy landscape.","fun_headline_variants_meta":{"raw":{"variants":["Enhancer competition alone predicts every blood progenitor","Epigenetic tug-of-war yields cell hierarchy","Math model: enhancer interplay spawns all blood progenitors","Self-similarity in enhancers explains lineage tree","One mechanism: enhancer competition builds cell hierarchy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001128,"raw_usage":{"total_tokens":4799,"prompt_tokens":1161,"completion_tokens":3638,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":777,"completion_tokens_details":{"reasoning_tokens":3563}},"tokens_in":777,"tokens_out":3638,"duration_ms":27087,"temperature":1.0,"reasoning_tokens":3563,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:03:02.111346+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the persistence length $1/\\mu_L^2$ for every subset of the eleven blood lineage programs from an independent, high-resolution expression dataset, then check the ranking claim: every documented progenitor should sit in the top tier of its size class, and every top-ranked subset should be findable by lineage tracing; one observed progenitor with a short persistence length, or one high-ranked subset that exhaustive tracing fails to find, would break the central prediction. A second, independent check targets the $\\beta$ axis: the model predicts that graded HDAC inhibition (raising $\\beta$) should progressively destabilize progenitors and push the population toward terminal states, so a titration experiment in a tissue with known terminal programs that instead freezes or stabilizes an intermediate state would contradict the writer/eraser control claim.","supporting_citations":[{"cited_title":"Demircigil , author J","cited_arxiv_id":null,"evidence_quote":"Reference review of hematopoiesis models providing the catalog of observed progenitor states and overlapping fate potentials against which the predictions are compared."}],"review_version":1}