{"id":"4882acd2-3e1e-4cec-b2e7-58f1672c669b","arxiv_id":"2607.19637","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Collective physical instabilities like phase separation can act as near-optimal decoders of molecular distributional codes, reading variance and skewness rather than just the mean.","lead":"This paper argues that collective physical instabilities—phase separation, percolation, and membrane buckling—can act as compact, near-optimal sensors for molecular distributions, while simple binding cannot. It supports this with an information-theoretic matching rule, mean-field theory, and lattice Monte Carlo simulations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Fisher-information derivation drops the ∂λ σ_R² term, which is largest exactly at the phase boundary where the paper claims near-total information capture; until this term is shown negligible, the 'nearly all information' claim is unsupported.","rationale":"The reader's weakest assumption correctly identifies the noise model as load-bearing. I sharpen this: the problem is not only the Gaussian assumption or the specific σ_dec ∝ ⟨R⟩^{1/2} scaling, but the explicit neglect of the ∂λ σ_R² term in the Fisher information. This term is not merely a technical detail; it is largest in exactly the regime where the paper claims maximal information capture. The paper's own SI provides enough information to test whether the omission matters, which is why the concern is concrete and resolvable. I do not think this warrants rejection: the mean-field phase-separation analysis, the percolation and membrane results, and the Monte Carlo demonstrations are largely independent of the information-theoretic optimality framing. However, the strongest claim as worded—that the geometric condition is necessary for efficient decoding and that phase separation captures 'nearly all information'—is not yet supported. The reader's other concerns (MC readout fitting, the kurtosis bound, and the mass-action 'mean-only' statement) are secondary; the variance-derivative omission is the most direct threat to the central claim. Thus the verdict remains conditional, and no adjustment to the reader's CONDITIONAL verdict is needed.","tokens_in":25338,"tokens_out":9024,"duration_ms":96763,"concrete_test":"Recompute Iλ for the phase-separation decoder near the phase boundary (the green line in Fig. 3D) including the full Gaussian Fisher term (∂λ⟨R⟩)²/σ_R² + (∂λ σ_R²)²/(2σ_R⁴), using the SI Section III expressions for ξ^d and droplet-volume fluctuations. If the added term is ≥10% of the reported Iλ at the parameter values used in Fig. 3, the claim that phase separation captures nearly all available information fails and the matching criterion needs revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that collective instabilities are near-optimal decoders rests on the Fisher-information expression Iλ = (∂λ⟨R⟩)²/σ_R² (SI Eq. 1). This expression is obtained from the exact Gaussian Fisher information by dropping the (∂λ σ_R²)²/(2σ_R⁴) term; SI Section I states 'we neglect the final term, assuming that encoder and decoder noise vary slowly with the environment.' That assumption is least secure in the regime the paper advertises: near the phase boundary, where the response R changes sharply and thermodynamic fluctuations are large. The paper's own fluctuation calculation (SI Section III) gives σ_R² ~ ξ^{2d}/Vtot in the dilute phase and droplet-volume fluctuations that diverge near critical points; both are functions of λ. Without evaluating the discarded variance-derivative term, the saturation of Iλ to I_A in Fig. 3D ('nearly all information') is not established. Moreover, the geometric 'contours orthogonal to encoder trajectory' criterion follows from maximizing the numerator at fixed denominator; if σ_R² varies with λ, the optimal decoder direction shifts, so the claimed universality of the matching condition is not a theorem. This does not invalidate the mean-field phase-separation analysis, but it directly undercuts the information-theoretic optimality claim as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that collective physical instabilities—phase separation, percolation, and membrane curvature remodeling—can act as \"natural decoders\" of molecular distributional codes, i.e., of information carried by the full distribution P(n) over a low-dimensional trait n. The core idea is an information-theoretic matching condition: an efficient decoder should have response contours in the (mean, variance) plane orthogonal to the trajectory traced by the encoder family, here taken to be exponential distributions. The authors derive a Gaussian-noise Fisher-information expression, show that simple mass-action binding has contours nearly parallel to the exponential encoder, and that phase separation, percolation, and membrane curvature have nearly orthogonal contours. For phase separation they add mean-field spinodal/critical-point analysis (Eq. 2) and lattice Monte Carlo simulations, claiming robust discrimination of mean and variance, weaker discrimination of skewness, near-total information capture near phase boundaries, and additional discriminatory power from finite valency. The Discussion generalizes these results into a proposed biological design principle.","tokens_in":25737,"tokens_out":6442,"duration_ms":72704,"significance":"If the central claims hold, the paper offers a genuinely new and compact principle for biological sensing: a single collective instability can integrate information from an entire molecular distribution without a large array of tailored sensors. The manuscript contains substantial technical assets: the mean-field spinodal condition (Eq. 2) is derived in closed form, the critical-point conditions are explicit, the binodal construction via exponential tilting is elegant, and the lattice Monte Carlo scheme is described with enough detail to be reproducible. The paper also makes falsifiable predictions, such as the n≥3 bridging threshold creating a discrimination axis absent from mean-field theory. These strengths are real. However, the information-theoretic optimality claims are not yet fully supported: a term is dropped from the Fisher-information expression in precisely the regime where the paper claims saturation, and the Monte Carlo readout is a fitted classifier, so part of the \"natural decoding\" claim is built in by construction. The manuscript is therefore valuable and promising, but needs substantial revision before the central optimality statements can be accepted.","major_comments":[{"comment":"The exact Gaussian Fisher information contains a term (∂λ σ_R²)²/(2 σ_R⁴). The paper neglects it on the stated assumption that \"encoder and decoder noise vary slowly with the environment\". But in the regime advertised as the main result—near the phase boundary—the paper's own fluctuation calculations (SI Section III) give σ_R² ~ ξ^{2d}/V_tot in the dilute phase and droplet-volume fluctuations that diverge near critical points; these are strongly λ-dependent. Without quantifying the dropped term, the saturation of Iλ to I_A in Fig. 3D and the geometric orthogonality criterion are not established. This is a load-bearing assumption for the paper's central optimality claim.","section":"SI Section I, Eq. (1); main text Fig. 3D"},{"comment":"The statement that \"mass-action binding detects only the mean\" is contradicted by the paper's own formula R_binding = 1 − Σ_n P(n)(1−p_bind)^n. At finite p_bind this is a nonlinear functional of P(n), not just of the mean; two distributions with identical mean and different variance will generally give different R_binding because (1−p)^n is convex. The claim is only true in the small-p_bind linear-response limit. Since this contrast anchors the paper's narrative, it should be explicitly qualified to that limit, or the stronger claim should be removed.","section":"Abstract, Fig. 1C, SI Section II"},{"comment":"The Monte Carlo readout R_MC is defined as a linear classifier on binned cluster-size distributions, with weights fit to separate the two target distributions. Any true difference in the cluster-size distributions then yields a nonzero discriminative readout by construction. The claim that finite valency enhances discrimination beyond mean-field theory therefore needs out-of-sample evaluation (e.g., cross-validation) or, ideally, a fixed physical readout such as largest-cluster size or a prescribed functional of the cluster distribution. As written, the broad discriminable regions in Fig. 4C–D may substantially overstate the natural, un-tuned decoding capacity of the phase-separating system.","section":"Fig. 4, SI Section IV"},{"comment":"The paper infers that distributions differing only in kurtosis have identical spinodal and critical-point loci and therefore a \"vanishing bound on robustness\" for kurtosis discrimination. This inference is not proven. The binodal away from the critical point is governed by the exponential-tilting relation Z(γ); as the SI itself notes, \"away from the critical point all γ contribute to the shape of the binodal, albeit at higher and higher orders.\" Identical spinodal/critical-point structure therefore does not imply vanishing binodal-region sensitivity. Either prove the stated bound or soften the conclusion; this is important because the Monte Carlo section is framed as lifting this mean-field bound.","section":"Main text after Eq. (2); SI Section III"}],"minor_comments":[{"comment":"The intermediate expression for the Gaussian Fisher information contains a factor 1/2 in the first term that is not carried into the final Eq. (1). Please clarify whether this is a typo or an intended approximation, and define σ_R² consistently in the main text.","section":"SI Section I"},{"comment":"The robustness metric r = Perfect/(Perfect + Imperfect) excludes the entire \"outside both\" region where weak dilute-phase amplification is available. Since the text says this region can still support discrimination near critical points, please state explicitly why excluding it does not bias the robustness comparison.","section":"Fig. 3G"},{"comment":"The cluster-size bins are described as \"logarithmically spaced\" but the listed edges [1,5,10,50,100,500,1000] are not logarithmically spaced. Correct the description or the bin edges.","section":"SI Section IV"},{"comment":"In the membrane-curvature free energy, the notation σ_mem is used both for surface tension and as a coefficient multiplying |∇h|²; the displayed equation is missing the factor 1/2 in the tension term. Please make the notation consistent.","section":"Main text Eq. (1)"}],"recommendation":"major_revision","confidential_remarks":"The paper is well-matched to physics.bio-ph and the core mean-field and simulation work is solid. I agree with the conditional verdict: the idea is attractive but the information-theoretic optimality and \"simple binding detects only the mean\" statements are overclaimed. These are fixable with additional calculations and qualifications, so I recommend major revision rather than rejection. The reviewer's stress-test concern about the dropped variance-derivative term lands and should be addressed directly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper argues that collective instabilities—phase separation, percolation, membrane curvature—can decode the shape of a molecular distribution, not just its mean, and that they approach an information-theoretic optimum. The core mean-field analysis is sound: the spinodal condition depends on variance, and the critical point adds skewness. The observation that polydispersity shifts phase boundaries, and that finite valency adds a threshold, is real and connects naturally to a broad biological literature. The information-theoretic framing is new and worth taking seriously.\n\nBut there are soft spots. The most important concerns the optimality claim. The Fisher-information expression in the main text drops the (∂λσ²_R)² term. The stress-test worried that this term is largest near the phase boundary, but it is positive—dropping it underestimates Fisher information. So the near-saturation claim is conservative, not inflated. The load-bearing assumption is the Gaussian noise model itself; near a phase boundary the response distribution can be bimodal or skewed, and the paper does not show how that would change the geometric criterion. Also, the variance-derivative term can shift the optimal decoder direction, so the universality of \"orthogonal contours\" is not a theorem. That claim should be softened.\n\nSecond, the \"mass-action binding detects only the mean\" statement is contradicted by their own formula: R_binding = 1 − Σ P(n)(1−p)^n, which depends on the full distribution. The contours run nearly parallel to the exponential family, but that is about this encoder family, not about simple binding in general. The overstatement should be corrected.\n\nThird, the Monte-Carlo readout is a linear classifier fit to the two distributions it then claims to discriminate. There is no cross-validation or error bar. This is a real circularity, and the claim that finite valency lifts the mean-field bound on skewness discrimination needs stronger support.\n\nThe kurtosis \"vanishing bound\" is asserted without derivation. The spinodal and critical-point analysis go up to moments 2 and 3, but the binodal relation involves a moment-generating function that contains all moments. If the bound holds, it needs a proof; if not, it should be retracted.\n\nWho is this for? Researchers in biological physics, cellular sensing, and synthetic biology will get a useful conceptual framework and some solid mean-field results. The paper deserves a serious referee, but the central optimality claim is not established as stated. Send it to peer review, and expect major revision.","headline":"Useful conceptual framing, but the optimality claim needs work: the Fisher-information approximation drops a positive term, and the Monte-Carlo readout is fit to the task.","tokens_in":26152,"tokens_out":3737,"would_cite":true,"duration_ms":39825,"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":"This paper claims that collective physical instabilities—phase separation, percolation, and membrane curvature remodeling—can act as near-optimal decoders of molecular distributional codes, while simple one-site mass-action binding cannot.","keywords":["distributional codes","phase separation","Fisher information","collective instabilities","percolation","membrane curvature","multivalency","encoder-decoder matching"],"falsifier":"Measure the noise statistics and Fisher information of a phase-separating readout in a reconstituted system as λ is varied. If the response–noise relation near the phase boundary does not follow σ_dec ∝ ⟨R⟩^{1/2}, or if a decoder whose contours are not orthogonal to the exponential encoder curve achieves higher Fisher information than an orthogonal one, the paper's central claim fails.","tokens_in":25243,"feed_emoji":"🔬","tokens_out":3692,"duration_ms":35111,"temperature":0.7,"pith_summary":"The paper tries to establish that cells can read information encoded in the shape of a molecular distribution—such as the number of phosphorylations or ubiquitin chain lengths—without building one sensor per variant. It derives an information-theoretic matching condition: an efficient decoder must have response contours that cross the encoder's trajectory orthogonally in the mean-variance plane. Phase separation, percolation, and membrane curvature instabilities satisfy this condition for biologically natural distributions, while simple mass-action binding does not. Using mean-field theory and lattice Monte Carlo simulations, the paper shows that phase separation reads beyond the mean—capturing variance and, more weakly, skewness—and that near phase boundaries it extracts nearly all the information in the molecular population. If true, this means a single collective instability can serve as a compact yet near-optimal sensor for distributional codes.","feed_headline":"A single phase transition can read a molecular distribution","feed_subtitle":"Phase separation, percolation, and membrane buckling decode variance and skewness; simple binding only reads the mean.","key_machinery":"The central object is the Fisher-information matching condition Iλ = (δ⟨R⟩/δP(n) · ∂P(n)/∂λ)^2 / (σ²_dec + σ²_enc), which factorizes decoder and encoder contributions and yields the geometric criterion that constant-response contours should be orthogonal to the encoder's curve in distribution space. The paper applies this to three collective processes—phase separation, percolation, and membrane curvature—using the exponential family P(n) ∝ e^{-λn} as the reference encoder. The load-bearing identity is that each collective instability has a convex response kernel in the trait n (e.g., the spinodal depends on μ and σ², percolation on ⟨n(n−1)⟩, membrane instability on ⟨n²⟩), whereas the mass-ac","core_discovery":"The central discovery is a geometric matching principle for decoding molecular distributions. Information theory shows that the Fisher information of a response R about an environmental parameter λ factorizes into a decoder sensitivity and an encoder direction, implying that optimal decoders have response contours orthogonal to the curve traced by P(n;λ) as λ varies. For the exponential encoder family P(n) ∝ e^{-λn}, phase separation, percolation, and membrane curvature remodeling all produce response contours that cut across the encoder trajectory, and their response kernels are convex in the trait n, amplifying high-n tails. Simple mass-action binding, by contrast, is concave and its conto","pith_inferences":["Our inference: the geometric orthogonality criterion could be used experimentally as a diagnostic—measure a candidate decoder's response contours in the mean-variance plane and check whether they cross the encoder family's curve.","Our inference: the framework suggests a general evolutionary rationale for why cells often encode information in distributions over related variants rather than distinct species: both encoding and decoding use physically natural, relatively non-specific processes.","Our inference: because the authors note their models are equilibrium, active or driven versions of these instabilities might show different information scaling or even sharper discrimination, an untested extension.","Our inference: a synthetic reconstitution experiment with a multivalent binder and controlled substrate valency distribution could directly test the predicted superlinear information gain near the phase boundary."],"forward_implications":["If the claim holds, a single phase-separating system could replace many tailored sensors for reading molecular distributions.","Near phase boundaries, the Fisher information extracted by phase separation scales superlinearly with system size and approaches the information limit set by counting individual molecules.","Phase separation can discriminate variance and skewness of input distributions, not just the mean, expanding what a cell can learn from a distribution.","Finite valency—through the threshold for network formation and monovalent sequestration—creates new discrimination axes that mean-field theory misses.","The encoder–decoder matching criterion provides a principled way to evaluate other collective readouts, such as cooperative binding, linker flexibility, or multi-component condensates."],"fun_headline_variants":["Phase separation reads molecular distributions beyond the mean","Collective instabilities decode molecular variance and skewness","Instabilities beat binding at reading molecular distribution codes","Phase transitions as compact sensors for molecular populations","Cells exploit phase separation to sense distribution codes"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The derivation assumes the readout noise is Gaussian, independent of λ, and scales as the square root of the mean response; if real cellular readouts are strongly non-Gaussian or if noise near a phase boundary scales differently, the geometric orthogonality criterion and the near-optimality conclusion do not follow.","fun_headline_variants_meta":{"raw":{"variants":["Phase separation reads molecular distributions beyond the mean","Collective instabilities decode molecular variance and skewness","Instabilities beat binding at reading molecular distribution codes","Phase transitions as compact sensors for molecular populations","Cells exploit phase separation to sense distribution codes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000136,"raw_usage":{"total_tokens":979,"prompt_tokens":737,"completion_tokens":242,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":481,"completion_tokens_details":{"reasoning_tokens":173}},"tokens_in":481,"tokens_out":242,"duration_ms":3001,"temperature":1.0,"reasoning_tokens":173,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T12:09:58.107577+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the noise statistics and Fisher information of a phase-separating readout in a reconstituted system as λ is varied. If the response–noise relation near the phase boundary does not follow σ_dec ∝ ⟨R⟩^{1/2}, or if a decoder whose contours are not orthogonal to the exponential encoder curve achieves higher Fisher information than an orthogonal one, the paper's central claim fails.","supporting_citations":[],"review_version":1}