{"id":"ae8ab61f-0257-4723-84c4-f820bf8f5283","arxiv_id":"2504.15107","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A cytoskeletal-inspired network with mechanosensitive proteins and motors can learn strain responses via temporal contrastive learning, and this learning supports adaptation and homeostasis.","lead":"This paper simulates a spring-and-motor network modeled on the cell cytoskeleton and shows it can be trained, like a neural network, to produce desired mechanical responses. The authors argue that mechanosensitive proteins and molecular motors give cells a general, non-neural way to learn from their environment.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Simulations appear to run outside the stated fast-memory regime: Fig. 2 uses k_u=0.5, giving τ_k=200 s for a 100 s cycle, while the contrastive argument requires τ_k<τ_f<τ_s; no k0_un is specified.","rationale":"The central claim is that mechanosensitive-protein kinetics plus motor activity implement temporal contrastive learning. For that to hold, the motor-density memory must act as a fast derivative estimator of strain so that the asymmetric free/clamped drive produces a clamped-minus-free update; this is the mechanism imported from Ref. [58]. The derivation of the memory kernel (Eqs. 4-5) requires two timescale assumptions: the mechanosensitive protein density must relax fast (Eq. A4), and the motor turnover time τ_k = 1/k_u must be shorter than the drive timescales (τ_k < τ_f < τ_s), as the paper states. Neither assumption is validated. No k0_un is assigned, so the adiabatic elimination of n is untested. Additionally, the reported simulations use rescaled k_u = 0.5; with τ_cyc = 100 s this gives τ_k = 200 s, violating the stated ordering. The simulations therefore demonstrate that some learning occurs, but not that the temporal-contrastive mechanism in the claimed regime is what produces it. Restricting the claim to the simulated slow-memory regime, or re-running in the fast-memory regime, would settle this. I credit the paper for a clean derivation and clear proof-of-principle framing, but the timescale gap is the load-bearing soft spot. Since the resolution is computational and the paper's own stated regime is testable, the CONDITIONAL verdict stands unchanged.","tokens_in":23702,"tokens_out":8878,"duration_ms":81011,"concrete_test":"Repeat the Fig. 2A protocol using the full protein and motor dynamics of Eqs. 2-3 instead of the adiabatic limit Eq. A4, with rescaled k_u chosen so that τ_k < τ_f < τ_s (e.g., k_u = 50 at τ_cyc = 100 s) and with k0_un scanned from 0.1 to 100 times the drive rate. If training error still falls to zero, the fast-memory assumption is not load-bearing; if it fails, the central claim must be restricted to the explicitly demonstrated slow-memory regime or re-derived for the fast regime.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central argument requires the motor-density memory kernel (Eqs. 4-5) to implement temporal contrastive learning in the regime τ_k < τ_f < τ_s, where τ_k = 1/k_u, as stated in Appendix A and Ref. [58]. This kernel is derived by adiabatically eliminating the mechanosensitive-protein density, δn = n0 β δϵ̇ (Eq. A4, 'as the protein dynamics is fast'). The paper never assigns a value to the protein unbinding rate k0_un, so that elimination is unvalidated. More concretely, the simulations appear to run outside the required regime: Fig. 2 uses rescaled k_u = 0.5; with τ_cyc = 100 s this gives τ_k = 200 s, longer than the entire cycle and longer than both τ_f = 20 s and τ_s = 80 s. In that limit δm ≈ β1 δϵ instead of a fast derivative estimate, so the clamped-minus-free contrastive interpretation imported from Ref. [58] is not the mechanism being simulated; the observed training may be a slow-memory averaging effect. The claim that mechanosensitivity plus motor activity yields temporal contrastive learning is thus conditional on a parameter regime stated in the text but not realized in the reported runs.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a computational model of a disordered spring network inspired by the actomyosin cytoskeleton. Each edge carries two molecular species: a mechanosensitive protein whose unbinding is strain-rate dependent (Eq. 3) and a motor whose binding is promoted by that protein (Eq. 2). Adiabatic elimination of the protein density gives the motor density an implicit memory of local strain, summarized by the kernel in Eq. (5) with timescale τ_k = 1/k_u and strength β_1 = k_1^b β n_0. The authors argue, building on Ref. [58], that driving the network asymmetrically between free and clamped states (fast clamp, slow release) turns this memory into a temporal-contrastive update of the learning degrees of freedom, which in this model are edge rest lengths (Eq. 6) or stiffnesses (Eq. 9). Simulations show decreasing training error for rest-length and stiffness learning, robustness to nonlinearities and edge turnover, classification of strain-gradient inputs, self-organized learning under actomyosin pulsation with length-dependent feedback, and adaptive maintenance of a low-strain 'perinuclear' region. The paper closes with experimentally testable predictions.","tokens_in":23990,"tokens_out":21488,"duration_ms":182418,"significance":"The analytic core is a strength: the memory kernel in Eq. (5) is derived from the model's own equations (Appendices A-B), the identification of τ_k and β_1 with k_u^{-1} and k_1^b β n_0 is checked by direct differentiation, and the learning outcomes are established by independent simulations of the dynamical system rather than assumed. The paper also tests robustness across several axes (activity, turnover, nonlinearities, network size) and offers falsifiable predictions, which is commendable. If the mechanism is as claimed, the work would be a significant minimal biophysical realization of temporal contrastive learning outside the neural context, extending Ref. [58] with thresholded remodeling, active contractility, turnover, and a homeostasis application. The significance is conditional, however, because the reported simulations appear to lie outside the regime in which the temporal-contrastive interpretation is valid; the mechanistic attribution is therefore not yet established by the evidence presented.","major_comments":[{"comment":"Appendix A states the contrastive requirement as τ_k < τ_f < τ_s and asserts that τ_cyc = 100 s is much larger than the motor turnover timescale, with τ_f/τ_s = 1/4. The reported simulations use rescaled motor unbinding rates k_u = 0.4–0.5 (Figs. 2–9, e.g., Fig. 2D), which with τ_cyc = 100 s gives τ_k = 1/k_u = 200–250 s, i.e., 2–2.5 cycles, about ten times τ_f = 20 s. This violates the stated condition and places the runs in the slow-memory regime, in which the kernel of Eq. (5) gives δm ≈ β_1 δϵ(t) rather than the strain-rate estimate δm ≈ β_1 τ_k δϵ̇(t) required for the temporal-contrastive argument of Ref. [58]. The observed learning could therefore be a slow-memory averaging effect rather than contrastive learning, and the substantial 'unlearning' reported in Fig. S5 (rest-length updates reversing sign during the slow release phase) is consistent with a strain-tracking, not a strain-rate-estimating, motor density. Note that if k_u = 0.5 were a real-time rate (0.5 s^{-1}), the rescaled value would be k̃_u = k_u τ_cyc = 50 and the stated regime would hold, so the discrepancy could be a units error in the captions; as written, the text and the simulations are inconsistent. Please clarify the units of k_u, rerun training in the regime τ_k < τ_f < τ_s (e.g., rescaled k_u ≳ 5), or provide per-cycle evidence (δm(t) records and a decomposition of the L0 update into clamp and release contributions) showing that the contrastive component dominates in the regime actually simulated. The Sections III–V attribution to temporal contrastive learning should be restricted to the regime in which it is demonstrated.","section":"Appendix A / Fig. 2"},{"comment":"The memory kernel is derived by adiabatic elimination of the mechanosensitive protein density, δn = n_0 β δϵ̇, justified by the statement 'as the protein dynamics is fast.' No numerical value of the protein unbinding rate k_un^0 is reported for any simulation, so the validity of this elimination on the driving timescales used (τ_f = 20 s for the Fig. 2 parameters, and 2 s for the classification task with τ_cyc = 10 s) is never checked. If the simulations integrate the reduced equations (A7) directly, the fast-protein assumption is imposed rather than tested, and the manuscript should say so explicitly and justify the elimination on timescale grounds; if the full protein dynamics of Eq. (3) is integrated, k_un^0 and k_bn must be given. Please report these values and, ideally, repeat one training run with the full protein dynamics to confirm that Eq. (5) remains the correct memory kernel in the regime of interest.","section":"Appendix A, below Eq. (A4)"},{"comment":"The paper asserts, rather than derives, that the implicit memory of Eqs. (4)-(5) turns the asymmetric drive into a per-edge learning update proportional to the difference between clamped and free strain. This step is imported from Ref. [58] without adaptation to the present model, which differs materially by the threshold nonlinearity g(x) in the learning rule (Eq. 6) and by the mechanical response of the network to the drive. Because edges with |f_a| < g_c receive no update, the cycle-integrated update cannot be exactly proportional to the free-clamped strain difference even in the correct memory regime. The reduction of the training error is established empirically, but the paper's central mechanistic claim requires either an analytic estimate of the per-cycle update (for example, for a single edge under the sawtooth drive) or a numerical decomposition of the update into clamp-driven and release-driven parts, showing that the contrastive part dominates in the simulated regime.","section":"Section III and Eq. (6)"}],"minor_comments":[{"comment":"Equation (A8) writes the supervised driving force as f = λ(t)∇(λ/2 |ϵ_T − ϵ*_T|²), which differs from Eq. (7) in sign and in the λ factor inside the gradient; this appears to be a typo and should be corrected for consistency.","section":"Appendix A (Supervised learning)"},{"comment":"The in-text citation '[52?]' is unresolved; please complete or remove it.","section":"Section X (Discussion)"},{"comment":"The reference list appears twice with two numbering schemes (Refs. [1]–[46] and Refs. [47]–[92] contain the same entries); please ensure that the manuscript contains a single bibliography whose numbering matches the in-text citations.","section":"References"},{"comment":"The abstract and introduction contain typos ('occuring phenomenolgy', 'homeostatis', 'mechnosensitive', 'analouges'); please copyedit the manuscript.","section":"Abstract"},{"comment":"The phase boundaries in Fig. 8 are described as determined by visual inspection, with boundaries 'drawn as guides to the eyes'; please provide a quantitative criterion (e.g., a threshold on normalized rest-length change or on connectivity change) so that the conserved-geometry versus altered-geometry classification is reproducible.","section":"Fig. 8"},{"comment":"The claim that network size has no significant effect on learning is based on a single trajectory per system size; please add several realizations with error bars, or present the statement as a qualitative observation.","section":"Fig. S4"},{"comment":"The threshold g_c is reported only as a range (10^{-6}–10^{-5}); please state the specific values used in each figure, since the learning dynamics depend on this hyperparameter.","section":"Model, Eq. (6)"},{"comment":"The claim of 'no significant qualitative or quantitative changes' under nonlinear mechanosensitive-protein dynamics is supported by Fig. S2A, which shows only two nonlinearity strengths; reporting final training-error values would make the claim quantitative.","section":"Section V"}],"recommendation":"major_revision","confidential_remarks":"This is a potentially interesting extension of the temporal contrastive learning framework to cytoskeletal networks, and the appendices are solid. The main risk is the parameter-regime mismatch documented in major comment 1, which I find to be genuinely borne out by the manuscript text: the stated condition τ_k < τ_f < τ_s and the reported rescaled k_u = 0.4–0.5 cannot both hold for τ_cyc = 100 s. If the authors can either demonstrate learning in the τ_k < τ_f regime or convincingly show that the slow-memory regime still implements an update dominated by the clamped-minus-free difference, the paper would be suitable for publication here; otherwise the title-level claim of contrastive learning is not supported by the reported runs. I would also ask the editor to have the duplicate reference list and the unresolved '[52?]' citation cleaned up before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. This paper is best read as a proof-of-principle that a disordered spring network with mechanosensitive-protein-gated motor kinetics can remodel itself to achieve desired mechanical responses. That part is plausible and the simulations support it. What the paper does not establish is the headline mechanism: temporal contrastive learning in the parameter regime actually simulated.\n\nThe new content is real: the specific molecular realization—LIM-domain-like proteins promoting motor binding, motor density carrying implicit strain memory, rest length or stiffness as the learning degree of freedom—plus the range of tasks (single input-output, classification, turnover, self-organized pulses, adaptation). Appendices A and B are clean; the memory kernel derivation is correct, and the turnover robustness checks are worth having. The paper also deserves credit for being explicit that this is a proof of principle.\n\nThe soft spot is a parameter-regime mismatch, not a style issue. The memory-kernel argument requires τ_k < τ_f < τ_s, and the paper even says so. But with τ_cyc=100 s, τ_f/τ_s=1/4, and the rescaled k_u=0.5, τ_k = 1/k_u = 200 s, which is longer than the entire cycle. So the simulated runs are outside the regime where δm approximates the strain-rate derivative. The paper doesn't give k0_un, so the adiabatic elimination of the mechanosensitive protein density (Appendix A, 'as the protein dynamics is fast') is unvalidated. I think the reader under-weighted this: the observed learning may be a slow-memory averaging effect, not temporal contrastive learning. Fixing the paper means either rerunning with k_u large enough to satisfy τ_k < τ_f, or revising the mechanistic claim to match the slow-memory regime and checking what the training actually implements.\n\nMinor issues, worth a referee report but not central: single realizations, no error bars, phase boundaries drawn by eye, no code or data. None of these change the verdict by themselves.\n\nFinal call: yes, send this to peer review. The idea is important enough and the demonstrations are broad enough that a serious referee should engage. My own use of the paper: I would not cite it as evidence for contrastive learning in cytoskeletal networks until the regime issue is fixed, but I would bring it to a reading group and ask the authors for the missing parameter values. It's a solid paper with one load-bearing mismatch.","headline":"A useful proof-of-principle for learning in cytoskeletal networks, but the reported simulations run outside the stated temporal-contrastive regime, so the mechanistic claim is not yet demonstrated.","tokens_in":24519,"tokens_out":3327,"would_cite":false,"duration_ms":31898,"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":"A disordered Hookean spring network carrying mechanosensitive proteins and motors can learn a desired strain response by remodeling rest lengths, through a contrastive-learning rule powered by strain memory.","keywords":["physical learning","contrastive learning","mechanosensitive proteins","molecular motors","cytoskeletal networks","actomyosin","adaptation","homeostasis"],"falsifier":"Run the same training protocol with the mechanosensitive protein unbinding rate set comparable to or slower than the driving timescale (for example, $k_{un} \\lesssim 1/\\tau_f$): if training error still decreases as in the fast-protein case, the protein-relaxation assumption is not load-bearing, and if learning fails, the assumption is confirmed as necessary. A complementary experiment would measure whether adaptation to repeated stretch is impaired when mechanosensitive protein recruitment is blocked in an actomyosin-like system.","tokens_in":23449,"feed_emoji":"🧬","tokens_out":7846,"duration_ms":72927,"temperature":0.7,"pith_summary":"This paper argues that a disordered network of Hookean springs, standing in for the cytoskeleton, can learn a desired mechanical response if its edges carry two generic biomolecular components: force-generating motors and mechanosensitive proteins that bind faster as strain rate rises. The central claim is a proof of principle: strain-history-dependent motor dynamics give each edge an implicit memory of strain, so driving the network between an unconstrained free state and a clamped state that enforces the desired output implements contrastive learning by remodeling edge rest lengths or stiffnesses. If correct, this means mechanosensitivity and motor activity are sufficient ingredients for non-neural learning in cells, and the same machinery can produce adaptation-like behavior such as keeping a perinuclear region at low strain under changing external forces.","feed_headline":"Mechanosensitive proteins let spring networks learn strain responses","feed_subtitle":"A cytoskeleton-inspired model tunes its rest lengths by contrast, with no neurons in sight.","key_machinery":"The load-bearing object is the strain memory kernel in Eq. (5): $K(t-t') = \\beta_1[\\delta(t-t') - (1/\\tau_k)e^{-(t-t')/\\tau_k}]$, obtained by linearizing fast mechanosensitive-protein kinetics and coupling them to motor binding. It converts local strain history into motor density, hence active force, and makes each edge's learning rule $\\dot{L}^0 = \\alpha g(f^a)$ respond to the difference between clamped and free strain rather than to instantaneous strain alone. The learning degree of freedom (LDOF) is the edge rest length (or stiffness), whose thresholded, active-force-dependent remodeling carries the contrastive update.","core_discovery":"The paper's central claim is that a disordered elastic network whose learning degrees of freedom, the rest length or stiffness of each edge, are updated only when the local active motor force exceeds a threshold can learn a prescribed strain at a target edge in response to a strain applied at a source edge. The mechanism runs through a memory kernel: because mechanosensitive proteins react to strain rate and promote motor binding, motor density obeys $\\delta m = \\int_{-\\infty}^{t} K(t-t')\\,\\delta\\epsilon(t')\\,dt'$ with $K(t-t') = \\beta_1[\\delta(t-t') - \\tau_k^{-1}e^{-(t-t')/\\tau_k}]$, so the active force at each edge carries a filtered estimate of strain history. When training alternates sharply between a free state and a clamped state, this implicit memory makes the rest-length update $\\dot{L}^0_{jk} = \\alpha g(f^a_{jk})$ behave as a contrastive learning rule, reducing training error toward zero. The authors show the learning survives nonlinear protein kinetics and nonlinear elasticity, persists under edge turnover when severing is slow enough, extends to classifying strain-gradient signs, and can be driven by self-organized actomyosin-like pulses rather than an external supervisor; in the pulse-driven case the same rule yields adaptation of a low-strain region.","pith_inferences":["If this mechanism is general, any cellular machinery with a fast strain-rate-sensitive binding step feeding a slower active process could serve as a physical learning module; the proteins modeled here are one instance.","The memory-kernel form suggests a quantitative design rule: the contrastive signal works best when the motor turnover timescale $\\tau_k = k_u^{-1}$ is short compared with the driving cycle, so tuning motor unbinding rate should tune learning speed, a dependence the simulations use but do not systematically sweep.","The adaptation result implies that continuous learning and homeostasis may be the same process; a testable extension is to ask whether repeated, changing perturbations preserve the learned low-strain state better than a single static perturbation.","Because the learning rule uses only local active force, it should transfer to vertex or tissue-scale models, where junction remodeling under myosin pulses could implement the same contrastive update."],"forward_implications":["Training error at the target edge falls toward zero over repeated free-clamped cycles, so the network stores the source-to-target strain relation in its rest lengths.","Learning is not tied to one geometry: it works for random source-target pairs, across tested network sizes, and for stiffness remodeling as well as rest-length remodeling.","Nonlinearities in protein kinetics and nonlinear elastic strain stiffening do not break learning; moderate nonlinearity can even lower training error.","Network turnover is tolerable: edges can sever and reconnect, and as long as severing is slow relative to reconnection, learning continues; increased contractility can restore learning after turnover-induced loss.","The same mechanism supports classification of strain-gradient signs and, with length-dependent pulsation amplitude, self-organized learning without an external supervisor; it also drives adaptation to keep a designated region near zero strain."],"supporting_citations":[{"why":"Supplies the temporal contrastive learning protocol through implicit non-equilibrium memory that the paper adapts to cytoskeletal networks.","marker":"[58]"},{"why":"Introduces the allostery-inspired response-learning task in mechanical networks that defines the source-to-target strain problem.","marker":"[55]"},{"why":"Provides the supervised physical-learning framework with learning degrees of freedom updated by clamped-versus-free contrast.","marker":"[57]"},{"why":"Frames the definition of learning without neurons in physical systems that the paper adopts.","marker":"[54]"},{"why":"Documents myosin II-dependent recruitment of mechanosensitive LIM-domain proteins, grounding the motor-protein coupling in the model.","marker":"[68]"},{"why":"Shows actin filaments act as tension sensors through tension-dependent cofilin binding, motivating the catch-bond-like severing rule used in the turnover study.","marker":"[78]"},{"why":"Documents pulsed actomyosin contractions during morphogenesis, which motivates the self-organized pulse driving used for unsupervised learning.","marker":"[80]"}],"fun_headline_variants":["Mechanosensitive networks learn strain patterns without neurons","Cytoskeletal model learns via mechanosensitive motors, no synapses","Forces teach spring networks: contrastive learning in cytoskeleton","Biological learning without neurons: mechanosensitive motors do it"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole argument rests on the mechanosensitive proteins reacting to each stretch instantly: in Appendix A the protein density is linearized to its instantaneous steady state, so the only memory in the motor dynamics comes from the simple kernel of Eq. (5); if those proteins bind and unbind too slowly, the strain memory the learning needs gets smeared out.","fun_headline_variants_meta":{"raw":{"variants":["Mechanosensitive networks learn strain patterns without neurons","Cytoskeletal model learns via mechanosensitive motors, no synapses","Forces teach spring networks: contrastive learning in cytoskeleton","Biological learning without neurons: mechanosensitive motors do it"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000179,"raw_usage":{"total_tokens":1269,"prompt_tokens":883,"completion_tokens":386,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":499,"completion_tokens_details":{"reasoning_tokens":315}},"tokens_in":499,"tokens_out":386,"duration_ms":4224,"temperature":1.0,"reasoning_tokens":315,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:33:26.843513+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same training protocol with the mechanosensitive protein unbinding rate set comparable to or slower than the driving timescale (for example, $k_{un} \\lesssim 1/\\tau_f$): if training error still decreases as in the fast-protein case, the protein-relaxation assumption is not load-bearing, and if learning fails, the assumption is confirmed as necessary. A complementary experiment would measure whether adaptation to repeated stretch is impaired when mechanosensitive protein recruitment is blocked in an actomyosin-like system.","supporting_citations":[{"cited_title":"Temporal contrastive learning through implicit non-equilibrium memory","cited_arxiv_id":null,"evidence_quote":"Supplies the temporal contrastive learning protocol through implicit non-equilibrium memory that the paper adapts to cytoskeletal networks."},{"cited_title":"Designing allostery-inspired response in mechanical networks","cited_arxiv_id":null,"evidence_quote":"Introduces the allostery-inspired response-learning task in mechanical networks that defines the source-to-target strain problem."},{"cited_title":"Supervised learning in physical networks: From machine learning to learning machines","cited_arxiv_id":null,"evidence_quote":"Provides the supervised physical-learning framework with learning degrees of freedom updated by clamped-versus-free contrast."},{"cited_title":"Learning without neurons in physical systems","cited_arxiv_id":null,"evidence_quote":"Frames the definition of learning without neurons in physical systems that the paper adopts."},{"cited_title":"Quantitative proteomics of the integrin adhesome show a myosin ii-dependent recruitment of lim domain proteins","cited_arxiv_id":null,"evidence_quote":"Documents myosin II-dependent recruitment of mechanosensitive LIM-domain proteins, grounding the motor-protein coupling in the model."},{"cited_title":"Actin filaments function as a tension sensor by tension-dependent binding of cofilin to the filament","cited_arxiv_id":null,"evidence_quote":"Shows actin filaments act as tension sensors through tension-dependent cofilin binding, motivating the catch-bond-like severing rule used in the turnover study."},{"cited_title":"Pulsed contractions of an actin–myosin network drive apical constriction","cited_arxiv_id":null,"evidence_quote":"Documents pulsed actomyosin contractions during morphogenesis, which motivates the self-organized pulse driving used for unsupervised learning."}],"review_version":1}