{"id":"b110b757-ceb8-41d5-b66b-eae130dcd79f","arxiv_id":"2507.11973","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Evolution with simple feedback controllers creates a dominant slow mode that funnels environmental and mutational disturbances, enabling low-dimensional control.","lead":"This paper argues that simple cellular control systems, which monitor only a few summary signals instead of every component, can keep a complex cell stable because evolution creates one special slow direction that most stresses tend to push along. Using math, simulations, and yeast data, the authors show such controllers work by channelling disturbances into that direction, and they predict that losing the controller makes stress responses more stereotyped, not more varied.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The evolutionary claim rests on a fitness function that omits the costs a soft mode should incur; without testing those costs, the reported mode-gap increase may be an artifact.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the evolutionary simulation's fitness function penalizes only the average residual deviation, not slowness, stability margins, or the cost of controller action. This is the most load-bearing point because the paper's headline claim is not merely that a mode gap helps a simple controller, but that selection for homeostasis actively drives the emergence of such a gap. The analytic section is conditional on the gap existing, so the evolutionary simulation is the only direct evidence for the arrow. Without modeling the costs that a soft mode naturally incurs, the reported mode-gap increase could be an artifact of the chosen cost function rather than a general evolutionary outcome. The empirical tests (GxG/GxE correlation and t-SNE clustering) are correlational and do not close this gap; even if they supported dual buffering, they would not show that the fitness function used in the simulations is the right one. A single controlled simulation experiment with an augmented fitness that penalizes small lambda_0 would settle whether the concern lands. If the mode gap still increases under such penalties, the paper's central claim is robust to this objection; if it does not, the claim needs substantial qualification. Since the reader's verdict is already CONDITIONAL and this concern reinforces that conditionality, the recommended verdict is UNCHANGED. No ad hominem is intended; the issue is an incompleteness in the model's fitness function and omitted supplement, not in the authors' reasoning or integrity.","tokens_in":12716,"tokens_out":16717,"duration_ms":217455,"concrete_test":"Rerun the Results A in-silico evolution for k=1 with an augmented fitness F' = <||delta n||> + c / lambda_0, varying c over at least an order of magnitude, and measure the evolved mode gap lambda_1 / lambda_0. If the mode gap no longer increases for c > 0, the central evolutionary claim depends on omitting the cost of slowness/fragility; if it still increases, the concern is resolved. The exact unnormalized fitness should be reported so the original Fig. 2c can be reproduced.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is evolutionary: selection for homeostasis with a low-dimensional controller should itself produce a soft mode. The analytic model (Eqs. 11-14) only shows that, conditional on a pre-existing mode gap, an optimal controller can align with the slow mode and reduce the normalized perturbation effect; it does not show that selection creates the gap. The only direct evidence for the arrow is the in-silico evolution in Results A. But the fitness there is described only as the residual deviation from the fixed point, normalized by stiffness and averaged over environmental fluctuations (Eqs. 4-5). This omits exactly the costs that a soft mode should carry: a small eigenvalue lambda_0 makes the unregulated response slow and hypersensitive (Eq. 8, delta x ~ J^{-1} delta e, with lambda_0 in the denominator), and it reduces the stability margin against parameter or mutational perturbations. If slow recovery, instability, or mutation-induced fragility has any fitness cost, the selection pressure identified in the paper could be offset or reversed. The paper does not analyze this trade-off, and the exact fitness function and evolution protocol are deferred to a supplement that is not provided in this version, so the reported increase in mode gap (Fig. 2c) cannot currently be audited or reproduced.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a theory of low-dimensional integral-feedback control in high-dimensional biological networks, arguing that selection for homeostasis with a simple, low-dimensional controller drives the evolution of a dynamical soft mode. The analytic model (Eqs. 11-14) is intended to show that a large mode gap lets a one-dimensional controller align with the slow mode and reduce the normalized impact of environmental perturbations. In-silico evolution simulations (Fig. 2) are reported to show that selection with a k=1 controller increases the mode gap. The authors then derive two empirical predictions: controllers that buffer environmental perturbations also buffer mutational perturbations (tested with yeast GxG/GxE data), and knocking out such a controller decreases the dimensionality of environmental responses (tested with kinase-inhibition transcriptomics). The central conceptual claim is interesting, but the manuscript currently leaves the simulation protocol, the analytic derivations, and the details of the empirical significance tests in an absent supplement, and the empirical figures lack statistical support.","tokens_in":13029,"tokens_out":7233,"duration_ms":86387,"significance":"If the central claim holds, the paper offers a generic, scale-independent mechanism for low-dimensional structure in biology and provides a new evolutionary explanation for soft modes that does not require assigning a direct function to the slow mode. The theory is not fit to the yeast data, and the predictions are falsifiable; the use of two large public datasets is a strength. However, the current verification is incomplete: the evolutionary result and the analytic formulas cannot be audited without the supplement, and the empirical dual-buffering and dimensionality claims rest on figures without error bars or quantitative tests. The conceptual contribution is substantial enough to justify a major revision rather than rejection.","major_comments":[{"comment":"The central evolutionary claim—that selection for homeostasis with a k=1 controller increases the mode gap—rests entirely on the in-silico evolution, but the fitness function is only described verbally as a normalized, time-averaged residual deviation, and the simulated annealing protocol, parameter ranges, and initial conditions are deferred to a supplement that is not present in this version. Without the explicit fitness function and protocol, the result cannot be audited or reproduced. This is load-bearing because the analytic theory in Section B shows only the benefit of a pre-existing mode gap, not that selection creates it.","section":"Section A, Eqs. (4)-(6), Fig. 2c"},{"comment":"The main analytic results—the solution for δx in Eq. (12), the claim that the optimal sensing and action vectors satisfy α=β=1 in the large-gap limit, and the normalized perturbation effect in Eq. (14)—are stated without derivation, with all details referred to the supplement. Since the supplement is absent, the reader cannot verify the central formula or the assumptions under which the optimality claim holds. Please include the derivations in an appendix or provide the supplement.","section":"Section B, Eqs. (12)-(14)"},{"comment":"The fitness cost only penalizes the residual deviation from the fixed point and does not include costs that a soft mode should naturally incur: slow unregulated recovery, reduced stability margins, hypersensitivity to perturbations (Eq. 8), and potential metabolic or fragility costs. Because a soft mode is by definition slow, a time-averaged deviation cost may or may not favor it depending on the averaging window and the controller dynamics; this trade-off is not analyzed. This is a load-bearing gap, since the reported increase in mode gap could be an artifact of a fitness function that ignores the costs of slowness.","section":"Section A, fitness in Eqs. (4)-(5)"},{"comment":"The dual-buffering empirical claim is supported only by cumulative histograms without error bars, confidence intervals, or significance tests. The criterion for identifying significant negative GxE and GxG interactions is stated to be the same as in the original studies, but the details are deferred to the missing supplement, so the threshold and multiplicity correction cannot be assessed. As presented, the figure does not quantitatively establish that genes buffering 6-7 environmental conditions buffer significantly more mutations than genes buffering 1-2 conditions.","section":"Section C2, Fig. 4d"},{"comment":"The dimensionality-decrease prediction is tested by visual inspection of t-SNE plots and by average distances in t-SNE space. t-SNE is a nonlinear embedding that can produce apparent clustering even in high-dimensional data and is not a measure of intrinsic dimensionality. To support the claim, the authors should report a quantitative dimensionality measure on the gene-expression matrices, such as effective rank or the variance explained by principal components, together with uncertainty estimates and a statistical test.","section":"Section D, Fig. 5d-f"}],"minor_comments":[{"comment":"The text says selection 'leads to a reduction in the system's mode gap—the ratio of the second and first eigenmodes λ1/λ0,' but Fig. 2c shows an increase in mode gap, and the later theory consistently uses a large λ/λ0 as the beneficial regime. Please correct this apparent contradiction.","section":"Section A, first paragraph"},{"comment":"In the sentence 'in the limit of a large mode gap λ/λ >>1,' the ratio should be λ/λ0 >> 1.","section":"Section B"},{"comment":"The slow mode is denoted λ0 in Section B but λ1 in Eqs. (15)-(16), and the condition 'λ1 >> λ_i>1' is undefined. Please use consistent indexing and explicit inequalities.","section":"Section C1, Eqs. (15)-(16)"},{"comment":"'Rapmycin' should be 'Rapamycin,' and the sentence 'mTOR is buffering over the effect of over 400 knockouts' is ungrammatical and should be rewritten.","section":"Section C2"},{"comment":"The caption does not state whether the curves are cumulative distribution functions or what the sample sizes are for each bin; please add a legend, sample sizes, and error bars or at least counts.","section":"Fig. 4d caption"},{"comment":"The caption uses 'm nodes' and 'k random and independent combinations' with variables that are not defined in the main text; please align the caption notation with Eqs. (4)-(6).","section":"Fig. 2 caption"},{"comment":"The sensing and action vectors are restricted to the span of the slow mode and the mean perturbation direction; the text should state explicitly that the optimality claim for α=β=1 is within this restricted family of controllers.","section":"Section B, parameterization of s and a"}],"recommendation":"major_revision","confidential_remarks":"This is a promising theory paper, but the current version reads like an extended abstract: the central evolutionary simulation, the analytic derivations, and the empirical significance-test details all live in an absent supplement, and the empirical figures lack statistical quantification. I would encourage the editor to request a full revision with the supplement and additional analyses rather than reject, because the core idea—selection for simple controllers generating soft modes—is novel and testable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis one is worth your time, but read it with the supplement missing in mind. The new idea is genuinely new: if a controller is constrained to sense and act on a low-dimensional projection, selection for homeostasis can itself favor a soft (slow) mode, because the controller then cancels most of the perturbation by sitting on that mode. That is distinct from earlier soft-mode explanations that require the mode to have a direct function like allostery or internal modeling.\n\nThe analytic model is the strongest part. Equation 14 is a clean, closed-form expression showing how a mode gap reduces the normalized perturbation effect, and the benefit grows with system dimensionality. The dual-buffering prediction—genes that buffer environmental perturbations should also buffer mutations—is crisp, and the check against the yeast GxG/GxE datasets is a genuine use of public data with no obvious parameter fitting. The Tpk123 result is suggestive but qualitative.\n\nNow the soft spots. The biggest one is the evolutionary claim. The only direct evidence that selection creates the mode gap is the in-silico evolution in Section A, and the fitness function there penalizes only the time-averaged residual deviation from the fixed point. A soft mode makes the unregulated response slow and hypersensitive; any fitness cost of slowness, instability, or fragility would work against the selection pressure you identify. The paper does not analyze this trade-off. That does not kill the idea, but it leaves the central arrow from 'selection for homeostasis' to 'emergent soft mode' undefended. Second, large parts of the derivations, the simulation protocol, and the statistical tests are in a supplement that is not present in the arXiv version. That blocks auditing and reproduction. Third, the dual-buffering plot (Fig. 4d) has no error bars and is correlational; minor but should be fixed.\n\nOverall: the analytic mechanism is solid as a conditional statement, and the framing is worth having in the literature. But the paper overstates what the simulations establish. I'd send it to peer review only after the supplement is provided and the fitness-cost objection is addressed. If you read it, start with Section B and the yeast analysis, and take the t-SNE with a grain of salt.","headline":"New idea about why simple controllers work—soft modes as a byproduct of selection for homeostasis—but the evolutionary simulations rest on a fitness function that may not penalize the costs of slowness.","tokens_in":13539,"tokens_out":3670,"would_cite":true,"duration_ms":43365,"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":"Selection for homeostasis in a high-dimensional network can create a soft mode that a low-dimensional controller then senses and acts on, which is why simple controllers like cAMP or (p)ppGpp can work.","keywords":["integral feedback control","soft modes","dimensionality reduction","homeostasis","evolutionary dynamics","mutational robustness","yeast knockout collection","stress response"],"falsifier":"Measure the relaxation time of a network after a perturbation in systems evolved under a simple controller: if imposing a fitness penalty proportional to recovery time prevents the mode gap from evolving, or if a measured soft mode carries a clear fitness cost in the organism's natural environment, the claim that homeostasis alone selects for soft modes would be falsified.","tokens_in":1555,"feed_emoji":"🧬","tokens_out":2309,"duration_ms":80316,"temperature":0.7,"pith_summary":"Simple biological controllers — low-dimensional integral feedback loops such as cAMP and (p)ppGpp — appear to regulate networks with thousands of components. This paper argues that they can do so because selection for homeostasis itself drives the controlled network to evolve a soft mode: one direction in state space that relaxes far more slowly than all others. Once such a mode gap exists, any environmental perturbation pushes the system mostly along that soft mode, so a controller that senses and acts on that single direction can restore homeostasis. The theory predicts that such controllers should buffer mutational perturbations as well as environmental ones, and that knocking the controller out should make the response to diverse environments lower-dimensional rather than higher. Both predictions are checked against yeast fitness and transcriptomic data.","feed_headline":"Simple controllers work because evolution builds a slow mode","feed_subtitle":"Homeostasis selection alone builds a direction that channels perturbations, letting simple controllers buffer both stress and mutations.","key_machinery":"The mode gap $\\lambda/\\lambda_0$ between the slowest mode and all other modes is the object that carries the argument, together with the effective rank of the ensemble of perturbation effects. A large mode gap lowers the effective dimensionality of how environmental perturbations move the system, so a low-dimensional integral feedback controller that projects the state onto one vector and acts along one vector can cancel the dominant slow mode. The paper's analytic result is that optimal sensing and action vectors align with the soft mode, reducing the residual perturbation to equation 14.","core_discovery":"The central claim is that a mode gap is not a fixed property of the network but an evolved consequence of having to be regulated by a controller with fewer measurements than the system has dimensions. In the linearized model, a system with a single slow mode $\\lambda_0$ and faster modes $\\lambda$ is controlled by sensing along one vector $\\vec{s}$ and acting along one vector $\\vec{a}$. In the limit of a large mode gap $\\lambda/\\lambda_0$, the optimal sensing and action vectors align with the slowest eigenvector, and the normalized effect of a perturbation reduces to $\\sqrt{\\sum_{i\\neq 0}(\\mu_i^2+\\sigma_i^2)} \\big/ \\sqrt{\\sum_{i\\neq 0}(\\mu_i^2+\\sigma_i^2)+(\\lambda/\\lambda_0)^2(\\mu_0^2+\\sigma_0^2)}$, meaning the soft mode is cancelled and only the fast modes contribute. The same soft mode channels environmental and mutational perturbations, so a controller selected for environmental robustness also buffers mutations. In simulations, evolution with low controller complexity produces a mode gap; evolution with high complexity does not.","pith_inferences":["Beyond the paper: the same mechanism could explain why 'capacitor' proteins like Hsp90 buffer both environmental and genetic perturbations, if they sit on the soft mode of the proteostasis network.","Beyond the paper: a direct experimental test would measure the relaxation spectrum of a metabolic or signaling network before and after selection under a simple controller, looking for the slowest eigenvalue to split from the rest only when the controller is low-dimensional.","Beyond the paper: the framework suggests that deliberately engineering a large mode gap in a synthetic host network could make a simple controller effective, reducing the need for complex multi-input regulatory circuits."],"forward_implications":["If selection for homeostasis creates soft modes, then low-dimensional controllers in biology are not a design puzzle: they are a signature of evolutionary pressure on the controlled network.","Environmental and mutational robustness should be correlated: genes that buffer many environmental conditions should buffer many mutations, as observed in the yeast knockout data.","Knocking out a controller that listens to a soft mode should expose that mode, making the transcriptomic response to diverse stresses more stereotyped and lower-dimensional.","Soft modes can evolve purely for dimensionality reduction, without any direct function such as allostery, giving an evolutionary origin for low-dimensional structure observed across biological scales.","The theory connects to cryptic genetic variation and global epistasis: mutations buffered along a soft mode hide their effects until the controller is compromised."],"supporting_citations":[{"why":"Provides the yeast GxG double-knockout fitness data used to test mutational buffering.","marker":"[13]"},{"why":"Provides the yeast GxE knockout-environment fitness data used to test environmental buffering.","marker":"[14]"},{"why":"Supplies the kinase-inhibition transcriptomics data that the dimensionality-reduction prediction is checked against.","marker":"[30]"},{"why":"Establishes integral feedback control as the biological mechanism that the model generalizes.","marker":"[22]"},{"why":"Supplies the soft-modes framework that the paper builds on for low-dimensional biological systems.","marker":"[11]"},{"why":"Links soft modes to epistasis, supporting the mutational-buffering part of the argument.","marker":"[12]"},{"why":"Provides the Hsp90 capacitor phenomenon as an empirical precedent for dual environmental and mutational buffering.","marker":"[25]"}],"fun_headline_variants":["Simple controllers drive evolution of soft modes","Low-dimensional control evolves a soft mode","Homeostasis selection generates soft modes","Soft modes evolve to enable simple controllers"],"cache_read_input_tokens":15616,"weakest_assumption_plain":"The evolutionary simulations score fitness only by how close the system stays to its fixed point on average, ignoring any cost of slow recovery, instability margins, or controller effort; if slow relaxation itself is costly, selection may not produce the large mode gap on which the predictions depend.","fun_headline_variants_meta":{"raw":{"variants":["Simple controllers drive evolution of soft modes","Low-dimensional control evolves a soft mode","Homeostasis selection generates soft modes","Soft modes evolve to enable simple controllers"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000298,"raw_usage":{"total_tokens":1741,"prompt_tokens":978,"completion_tokens":763,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":594,"completion_tokens_details":{"reasoning_tokens":713}},"tokens_in":594,"tokens_out":763,"duration_ms":9718,"temperature":1.0,"reasoning_tokens":713,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:57:50.991715+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the relaxation time of a network after a perturbation in systems evolved under a simple controller: if imposing a fitness penalty proportional to recovery time prevents the mode gap from evolving, or if a measured soft mode carries a clear fitness cost in the organism's natural environment, the claim that homeostasis alone selects for soft modes would be falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the yeast GxG double-knockout fitness data used to test mutational buffering."},{"cited_title":"Husain and A","cited_arxiv_id":null,"evidence_quote":"Provides the yeast GxE knockout-environment fitness data used to test environmental buffering."},{"cited_title":"Costanzo, A","cited_arxiv_id":null,"evidence_quote":"Supplies the kinase-inhibition transcriptomics data that the dimensionality-reduction prediction is checked against."},{"cited_title":"Friedlander, A","cited_arxiv_id":null,"evidence_quote":"Establishes integral feedback control as the biological mechanism that the model generalizes."},{"cited_title":"Bahar, On the functional significance of soft modes pre- dicted by coarse-grained models for membrane proteins, Journal of General Physiology135, 563 (2010)","cited_arxiv_id":null,"evidence_quote":"Supplies the soft-modes framework that the paper builds on for low-dimensional biological systems."},{"cited_title":"Leo-Macias, P","cited_arxiv_id":null,"evidence_quote":"Links soft modes to epistasis, supporting the mutational-buffering part of the argument."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Hsp90 capacitor phenomenon as an empirical precedent for dual environmental and mutational buffering."}],"review_version":1}