{"id":"d27efa9c-eb09-4174-9553-247e0873e01d","arxiv_id":"2602.00261","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Bayesian inference over a known class of monotonically saturating concentration profiles gives relative estimation error 1/(a√N), better than the Poisson 1/√N limit when the early rise rate is a nonlinear function of the final concentration.","lead":"This theoretical paper shows that when a cell knows the stereotyped way a chemical concentration rises toward its final level, it can predict that final level from early binding events with error 1/(a√N), beating the standard 1/√N counting limit for a>1. The result could help explain how developing embryos commit to cell fates before morphogen gradients fully form.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Super-Poisson precision in Eq. (26) presupposes exact knowledge of a and k; with realistic parametric uncertainty, c_∞ is unidentifiable and the error floor does not vanish with N.","rationale":"The reader's verdict is CONDITIONAL, with the weakest assumption being the sensor's exact knowledge of the profile class (including a and k), the prior, and the start time. My stress-test sharpens this into a specific identifiability failure: the parameters a and k are not identifiable from binding data alone, so even a perfect start time cannot rescue the super-Poisson prediction if a or k are uncertain. This is the most load-bearing concern because it attacks the core mathematical result (Eq. 26) at its foundation: the claimed 1/(a²N) scaling is not a physical limit but a consequence of assuming a and k are known to infinite precision. The paper's own discussion concedes practical implementation is open, and the biochemical network requires exact rate-constant matching, so the concern is consistent with the paper's honest caveats. It does not reveal a mathematical error in the derivation under the stated assumptions, but it strengthens the conditionality: the super-Poisson advantage is not robust to parametric uncertainty. The proposed concrete test—computing the rank of the Fisher information and the posterior variance under priors on a and k—would definitively show whether the error floor persists. Since the reader already required robustness checks for the assumed profile class, my analysis reinforces the CONDITIONAL verdict rather than changing it; hence UNCHANGED. I selected 'partial' agreement because the reader's weakest assumption mentions a and k but does not identify the singular Fisher information or the non-vanishing error floor that makes the claim fragile even within the assumed model family.","tokens_in":16268,"tokens_out":21544,"duration_ms":237619,"concrete_test":"Compute the 3×3 Fisher information matrix for the model λ(t) = r k c_∞^a t on [0,T] at a fixed c_∞, a, k. Verify it is rank 1 (all entries are proportional to N times outer products of constants), implying the MLE cannot separately identify c_∞, a, k. Then, endow log a and log k with Gaussian priors of small widths σ_a, σ_k and compute the posterior variance of log c_∞ as N→∞ (e.g., via Laplace approximation or MCMC). If the posterior variance converges to a positive constant proportional to σ_a² (or σ_k²) rather than to 1/N, the super-Poisson claim fails under parametric uncertainty. This directly tests whether the ε²/c_∞² ≈ 1/(a²N) result is robust to realistic uncertainty in the assumed profile family.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central super-Poisson claim, ε²/c_∞² ≈ 1/(a²N) (Eq. 26), is derived for a Poisson process with rate λ(t) = r k c_∞^a t under a one-parameter model where a and k are known constants. However, the likelihood depends on c_∞, a, and k only through the combination k c_∞^a: the event-time distribution is independent of these parameters given N, and the count N is Poisson with mean r k c_∞^a T²/2. The 3×3 Fisher information matrix for (c_∞, a, k) is therefore rank 1. If a or k are not known exactly—as is inevitable for a biological sensor that has not been perfectly calibrated—then no amount of data can separate c_∞ from the product k c_∞^a. Adding any continuous prior over a or k introduces a non-vanishing contribution to the posterior variance of log c_∞ that persists as N→∞, because the likelihood is flat along directions that change a and k while keeping k c_∞^a fixed. Consequently, the 1/(a²N) scaling is not achieved; the error saturates at a constant determined by the prior widths. The paper's own biochemical network (Eqs. 29–32) requires rate constants satisfying k_Z^+ k_A/(k_Z^- k_X k_Y) = 1/(r k), which presumes the cell has tuned to the exact k and r. This is a more fundamental load-bearing assumption than the start-time issue: even with a perfect t=0, parametric uncertainty caps prediction accuracy. The reader's weakest assumption alludes to this via 'including a and k' but does not emphasize the identifiability singularity, which is the sharpest technical route to showing the claimed precision is unattainable under realistic uncertainty.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the precision with which a cell can estimate the future steady-state concentration (or its spatial position in an embryo) from the stochastic binding times of ligand molecules to a receptor, when the concentration profile is known to belong to a parametric family. The authors formulate the problem as Bayesian MAP estimation for an inhomogeneous Poisson process. For a piecewise-linear saturating profile c(t) = k c_∞^a t for t < t1 and c∞ thereafter, they derive ε²/c∞² ≈ 1/(a²N) when the measurement is made during the growth phase, where N is the total number of binding events. Since a≈1.8 for a simple Bicoid model, this beats the Poisson counting limit 1/N. They further show that a simple biochemical network can approximately implement the MAP estimator. The constant-concentration limit recovers the standard 1/N result, and simulations in Fig. 3(b) support the error scaling.","tokens_in":16683,"tokens_out":8008,"duration_ms":94766,"significance":"If correct, the result is significant because it identifies a mechanism—using prior knowledge of concentration-profile structure—by which cells could extract more than Poisson-level information from a fixed number of binding events and predict future concentrations before steady state is reached. This could contribute to explaining the speed-precision paradox in developmental decisions. The Fisher-information calculation is transparent and internally consistent; the authors are explicit that the piecewise-linear profile is an analytically tractable representative rather than a full biophysical model. The paper also provides a concrete biochemical network implementation, which is a strength. The main weakness is that the advertised super-Poisson scaling requires the sensor to know the exact profile family, including parameters a and k and the start time, and the paper does not analyze the consequences of relaxing these assumptions.","major_comments":[{"comment":"The likelihood for T ≤ t1 depends on (c∞, a, k) only through μ = r k c∞^a. The 3×3 Fisher information matrix for (c∞, a, k) is thereby rank 1, so if a or k are not known exactly, the posterior variance of c∞ does not vanish as N→∞; the ε²/c∞² ≈ 1/(a²N) scaling is a property of the perfectly calibrated model. The manuscript does not state this assumption explicitly or discuss its implications. Please add an explicit statement that a and k are assumed known, and analyze the robustness of the result, e.g., by placing priors on log a and log k and showing how the error floor scales with their widths. The biochemical network in Eqs. (29)–(33) similarly requires rate constants tuned to 1/(rk), which is a strong requirement.","section":"Section III, Eq. (26)"},{"comment":"The derivation sets t=0 at the arrival of the first molecules, but the sensor cannot observe this time until the first binding event occurs. The paper acknowledges that 'a sensor cannot know that development has already started' but then proceeds as if the start time were known. An unknown start time t0 adds a parameter; even if it is identifiable in principle, it will increase the estimation error and may weaken the improvement over the Poisson limit. Please analyze this case or add a caveat to the claims of prediction accuracy.","section":"Section III, before Eq. (23)"},{"comment":"The Discussion and abstract present the super-Poisson result as a general property of structured concentration profiles, but the main derivation is for the specific piecewise-linear family in Eq. (23). The only link to actual Bicoid profiles is the scaling of the maximum rate of change, v ~ c∞^1.8 (Fig. 1(c)). The paper should state more explicitly that the ε²/c∞² ≈ 1/(a²N) result is proved only for this representative family, and that the Fisher information for realistic saturating profiles may differ.","section":"Section V"}],"minor_comments":[{"comment":"Typo: 'renrmalize' should be 'renormalize'.","section":"Section II, after Eq. (3)"},{"comment":"The text refers to 'D ss', but no variable D has been defined; this should likely read 'Z ss' or the appropriate readout variable.","section":"Section IV, after Eq. (33)"},{"comment":"The statement 'Numerical averages over n=100 repetitions produced estimator means and variances consistent with our calculations here in the N≫1 limit (not shown)' is unsupported because no results are shown. Please either include the data or cite the simulation in Fig. 3(b) as the verification.","section":"Section II, after Eq. (19)"},{"comment":"The estimator is implicit because t1 depends on c∞. The paper notes this, but a short demonstration of how the self-consistency is resolved in practice would improve readability.","section":"Section III, Eq. (24)"}],"recommendation":"major_revision","confidential_remarks":"The paper makes a clean, internally consistent Fisher-information calculation for an idealized sensing scenario, and the constant-concentration limit correctly recovers known results. The central concern is that the super-Poisson claim is sensitive to exact knowledge of the profile parameters a and k, and to the start time t=0, neither of which is biologically guaranteed. These are not internal inconsistencies, but they are load-bearing for the advertised biological significance. A revision that explicitly states these assumptions and quantifies the effects of relaxing them would make the contribution much stronger. The paper's scope is appropriate for physics.bio-ph, and I do not see grounds for rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline: this paper adds something real to the biochemical sensing literature. It shows that if a cell has a prior over a known class of monotonically saturating concentration profiles, MAP estimation can predict the eventual steady-state concentration with error ε²/c∞² ≈ 1/(a²N) — better than the Poisson counting 1/N when the profile exponent a>1. The Fisher-information calculation is straightforward and clean, the constant-concentration limit recovers the known Endres–Wingreen/ML result, and the piecewise-linear model is explicitly presented as an analytically tractable surrogate, not as a literal morphogen profile. The biochemical reaction-network sketch is a plausible existence proof, and the numerics line up.\n\nThe soft spots are real. The most important is the parameter-identifiability issue: in the pre-saturation regime the likelihood depends on c∞, a, and k only through k c∞^a. With a and k unknown, the Fisher information matrix for the triple is rank 1, so no amount of data separates c∞ from a or k. Any continuous prior on a or k leaves a posterior width in log c∞ that does not vanish as N→∞. So the 1/(a²N) scaling is only as good as the sensor's exact knowledge of a and k. The reader's note flagged this indirectly; the stress-test makes it precise. The paper should own this limitation and either argue that a and k are effectively hardwired constants the cell knows, or analyze the error with priors over them. The start-time assumption is similarly idealized: time measured from arrival of first molecules may be physically accessible, but the paper gives it only a passing mention.\n\nThese are caveats, not fatal flaws. The core mathematical claim holds as a conditional limit: given the profile family and parameters, the bound is exactly what the paper says. The paper is honest about the idealized setup, and the Drosophila discussion is appropriately speculative. I'd take it as a solid theoretical contribution.\n\nRecommendation: send to peer review. The authors need to add a section on robustness to unknown profile parameters and start-time uncertainty; otherwise a sharp referee will raise the identifiability point, and it should be addressed before publication. I would cite it.","headline":"A clean, genuinely new result for sensing with known structured profiles; the super-Poisson scaling depends on exact knowledge of a and k, and the paper needs to face that limitation before publication.","tokens_in":17151,"tokens_out":2957,"would_cite":true,"duration_ms":36395,"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 cell that knows the shape of a signal can read its final strength early, with accuracy better than simple molecule counting.","keywords":["concentration sensing","Bayesian inference","maximum a posteriori estimation","Berg-Purcell limit","shot noise","morphogen gradient","positional information","developmental biology"],"falsifier":"Fit a measured mutant morphogen profile where the early rise and the eventual plateau are decoupled (so the assumed c(t)=k c_∞^a t form fails). If the precision of downstream boundary placement does not degrade from the predicted 1/(a√N) scaling, or if a sensor using only post-saturation events matches the full-trajectory predictor, the central claim would be contradicted. A cleaner test: in an embryo with a manipulated early profile (e.g., shifted production onset), measure whether the MAP-predicted boundary position — computed from binding events in the first minutes — anticipates the eventu","tokens_in":16141,"feed_emoji":"🧬","tokens_out":4232,"duration_ms":45184,"temperature":0.7,"pith_summary":"The paper argues that when a cell knows the family of possible concentration profiles (as in embryogenesis), it can treat concentration sensing as Bayesian inference over that family instead of merely counting molecules. For profiles whose early rate of rise correlates strongly with the eventual plateau level — c(t) ≈ k c_∞^a t before saturation — the MAP estimator achieves fractional error 1/(a√N) for N binding events. For a > 1 this beats the standard Poisson counting limit 1/√N, so the cell can estimate the future steady-state concentration (and hence its position in the embryo) before the gradient has settled. The authors show the required computations can be implemented by simple biochemical reaction networks, and connect the result to the speed–precision puzzle in Drosophila development.","feed_headline":"Cells can predict a signal's final strength before it arrives","feed_subtitle":"Knowing the shape of a morphogen profile lets embryos commit to fates before the gradient stabilizes","key_machinery":"The key object is the maximum a posteriori (MAP) estimator over a parameterized family of concentration profiles c_θ(t), where each binding time is drawn from an inhomogeneous Poisson process with rate r c_θ(t). The workhorse identity is the Cramér–Rao bound applied to the log-posterior: ε² ≈ [∫₀ᵀ (r/c̄)(∂c̄/∂θ)² dt]⁻¹. For the piecewise-linear saturating profiles, this integral evaluates to a²N, directly yielding the super-Poisson error scaling. The analysis also derives a stochastic differential equation (Eq. 17) for how the estimate updates between and at binding events, and shows that for a=2 the update rule reduces to operations (ratios of clock-measured time squared and binding count)","core_discovery":"The central claim is that prior knowledge of a restricted class of spatiotemporal concentration profiles converts the sensing problem from estimating an unknown concentration to identifying which member of a known family produced the observed binding times. In the piecewise-linear model c(c∞, t) = k c_∞^a t for t < c_∞^{1-a}/k and c∞ afterward, the MAP estimator has variance ε²/c∞² ≈ 1/(a²N) during the growth phase. Since N is the number of binding events, a > 1 means the relative error is smaller than 1/√N, the Poisson limit that bounds classical maximum-likelihood and Berg–Purcell sensing. The same scaling holds for positional inference, with the improvement mirrored through the profile's","pith_inferences":["A direct implication the authors leave implicit: the improvement is equivalent to an effective count of a²N, so a downstream readout that behaves like a counter would need to count correlated events to mimic the Bayesian estimator.","One could test the theory experimentally by perturbing the early phase of a morphogen profile (e.g., shifting production rate or start time) and measuring whether downstream gene expression boundaries track the MAP prediction rather than the instantaneous concentration.","The requirement that the sensor know t=0 (arrival of first molecules) suggests a testable prediction: if a cell cannot set time origin reliably, the predicted 1/(a√N) precision should degrade; this could be probed by artificially varying the delay between fertilization and the first detectable binding.","The biochemical implementation shown for a=2 suggests a design principle: cells may use constitutive clocks (protein accumulation) to normalize event counts by time factors, which could be sought in real regulatory networks."],"forward_implications":["A nucleus in a developing embryo can infer its position before the morphogen gradient reaches steady state, using binding events from the very beginning of the profile.","The effective sensing period extends back to the arrival of the first molecules, so N counts events from the start of the process, not just from stabilization — multiplying accuracy gains.","For profiles with a≈1.8 (as fitted to a diffusion–degradation model of Bicoid), the error prefactor is roughly 3.2 times smaller (1/a² ≈ 0.31), a substantial improvement over Poisson counting.","The super-Poisson advantage disappears (a=1) or reverses (a<1) when early dynamics do not amplify differences in the eventual saturation value, so the effect is specific to strongly coupled profiles.","The same MAP formalism applies to any reproducible, pre-programmed spatiotemporal signal, not just development, with the appropriate prior over profiles."],"fun_headline_variants":["Embryos predict future morphogen levels before they arrive","Structure-aware sensing beats Poisson shot noise","MAP estimation with known profiles exceeds classical precision","Predicting concentration futures outperforms standard limits","Leveraging profile shape predicts signals ahead of time"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The sensor must know the exact parametric family of profiles (including the exponent a), the prior over the parameter, and the time origin measured from the arrival of the first molecules; if the real profile deviates from the assumed family, or the cell cannot establish t=0 and the prior, the 1/(a√N) accuracy is not guaranteed.","fun_headline_variants_meta":{"raw":{"variants":["Embryos predict future morphogen levels before they arrive","Structure-aware sensing beats Poisson shot noise","MAP estimation with known profiles exceeds classical precision","Predicting concentration futures outperforms standard limits","Leveraging profile shape predicts signals ahead of time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00025,"raw_usage":{"total_tokens":1398,"prompt_tokens":761,"completion_tokens":637,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":505,"completion_tokens_details":{"reasoning_tokens":567}},"tokens_in":505,"tokens_out":637,"duration_ms":7645,"temperature":1.0,"reasoning_tokens":567,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T06:05:36.844821+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fit a measured mutant morphogen profile where the early rise and the eventual plateau are decoupled (so the assumed c(t)=k c_∞^a t form fails). If the precision of downstream boundary placement does not degrade from the predicted 1/(a√N) scaling, or if a sensor using only post-saturation events matches the full-trajectory predictor, the central claim would be contradicted. A cleaner test: in an embryo with a manipulated early profile (e.g., shifted production onset), measure whether the MAP-predicted boundary position — computed from binding events in the first minutes — anticipates the eventu","supporting_citations":[],"review_version":1}