{"id":"accbe78a-29f2-43dc-b679-98f582aceab2","arxiv_id":"2608.07975","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Turning-angle distributions, which are invariant to random rotations, reveal anisotropic diffusion in live-cell single-particle trajectories that standard covariance analysis misses.","lead":"Molecules inside living cells sometimes drift more easily along one direction than another, a property called anisotropy that is hard to see when every cell is tilted at a random angle. A team has developed a turning-angle analysis that stays unchanged under rotations, and used it to reveal hidden anisotropies in the motion of quantum dots and membrane proteins from live-cell tracking data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Experimental r estimates are not goodness-of-fit tested: the CvM statistic is used only to pick a minimum, so the central claim of widespread intracellular anisotropy could be an artifact of FBM model misspecification.","rationale":"The central claim is twofold: turning-angle statistics are a valid rotation-invariant probe (well supported by theory and simulations), and the experimental data show widespread anisotropy (dependent on the FBM model). The reader's weakest assumption identifies model dependence; I agree. The most load-bearing operational gap is the absence of any goodness-of-fit assessment of the fitted FBM against the experimental turning-angle distributions. Without a null distribution for the CvM statistic, the curves in Fig. 5 only rank candidate r values; they do not tell us whether the model can reproduce the data at all. This matters because the fitted r is the sole quantitative evidence for the biological conclusion. The bootstrap test proposed above would settle the concern: if the model passes for the states where anisotropy is claimed, the conclusion is strengthened; if it fails, the r estimates are artifacts. I therefore keep the reader's CONDITIONAL verdict unchanged.","tokens_in":22879,"tokens_out":14528,"duration_ms":153087,"concrete_test":"For each of the six experimental states, run a parametric bootstrap goodness-of-fit test: simulate B = 1000 synthetic datasets from the fitted anisotropic 2D-FBM with parameters (H_est, r*) and the same number of turning angles per state, using the same segmentation pipeline (local convex hull) and a localization-noise level estimated from the data. For each synthetic dataset, compute the CvM distance to the same look-up table. Compare the observed CvM distance to the 95th percentile of the bootstrap null distribution. If the observed distance exceeds that percentile for any state, the anisotropic FBM model is rejected for that state, and the reported r must not be interpreted as a physical anisotropy ratio.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The experimental anisotropy ratios r in Fig. 5 are obtained by minimizing the two-sample Cramér–von Mises distance between the empirical turning-angle distribution and look-up tables simulated from anisotropic 2D-FBM at a fixed Hurst exponent H (Supplementary Methods D). No goodness-of-fit step is reported: the CvM statistic is never compared to its null distribution, so we do not know whether any (H, r) in the look-up grid is statistically compatible with the data. The simulation validation (Fig. 6 and Supp. Figs. S3, S5) only demonstrates that the estimator recovers r when data are truly FBM; it cannot certify the experimental datasets. Given the very low H values (e.g., H_Nav,l = 0.06) and the acknowledged static-noise effects, the empirical distributions may be inconsistent with every stationary Gaussian FBM, in which case the reported r ≈ 0.16–0.41 is a best-fit shape parameter rather than evidence of physical anisotropy. The paper therefore does not support the Discussion statement that anisotropic anomalous diffusion is an overlooked feature of intracellular dynamics until the FBM model is tested against the data.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes turning-angle analysis as a rotation-invariant tool for detecting anisotropy in two-dimensional anomalous diffusion. The authors derive the increment cross-covariance of an anisotropic two-dimensional fractional Brownian motion (2D-FBM) with independent components, show that random rotations average the cross-covariance to zero, and prove that the turning-angle distribution depends on the Hurst exponent H and anisotropy ratio r but not on orientation. They validate the estimator on simulated 2D-FBM with and without localization noise, then apply it to segmented high- and low-mobility states of quantum dots in HeLa cytoplasm and Nav1.6/CD4 channels in hippocampal neurons, reporting anisotropy ratios r ≈ 0.16–0.48 for most states. They conclude that anisotropic anomalous diffusion is an overlooked feature of intracellular dynamics.","tokens_in":23088,"tokens_out":6347,"duration_ms":68456,"significance":"The theoretical contribution is solid and well presented: Eqs. (3)–(6) and Supplementary Eqs. (48)–(51) rigorously establish the rotation-invariance argument, and the paper ships reproducible Julia/Python code, look-up table generation, and simulation studies including localization noise (Fig. 6 and Supp. Figs. S3, S5). If the experimental r estimates were validated, the method would be a useful addition to the single-particle tracking toolbox, since it can pool randomly oriented short trajectories where ensemble cross-covariance vanishes. The inference is not circular in the narrow sense that the turning-angle distribution is generated from the model covariance matrix and compared with data; however, the central experimental claim currently rests on fitting the empirical turning-angle distributions to a single parametric model without any goodness-of-fit test, so the reported anisotropies are not yet established as physical properties of the cells.","major_comments":[{"comment":"The Cramér–von Mises statistic is used only to select the minimizing r and to define the uncertainty region through the rule dCvM(r) ≤ 2 dCvM(r*); its null distribution is never computed, so the procedure contains no goodness-of-fit test. Consequently, Fig. 5 and the Discussion statement that anisotropic anomalous diffusion is widespread are not supported: for the very low estimated Hurst exponents (e.g., HNav,l = 0.06), the empirical turning-angle distributions may be inconsistent with every stationary Gaussian 2D-FBM in the look-up table, in which case the reported r values are best-fit shape parameters rather than physical anisotropies. Please add a parametric bootstrap or permutation test that compares the observed minimum CvM distance to the distribution obtained under the fitted model with the same trajectory lengths, segmentation, and noise, and report the resulting p-values or rejection regions.","section":"Supplementary Methods D, Eqs. (55)–(58)"},{"comment":"The experimental anisotropy estimates are obtained by comparing noisy measured turning-angle distributions to look-up tables generated from noiseless 2D-FBM (Methods E), while Fig. 6 evaluates the effect of noise by comparing noisy simulations to the same noiseless tables. The paper reports that localization noise broadens the uncertainty ranges, but it does not report whether the minimum of the CvM curve shifts with noise level; if it shifts, the point estimates rQD,h = 0.17, rNav,h = 0.16, etc. are biased. Please construct look-up tables with localization noise matched to each dataset, or demonstrate explicitly that the bias is negligible for the estimated H and signal-to-noise ratios.","section":"Section II D and Fig. 6"},{"comment":"The estimation protocol assumes the same Hurst exponent H in both intrinsic directions (H1 = H2 = H) and independent intrinsic components (Eq. 11), and the look-up tables are generated under exactly that model. Any anisotropy in the scaling exponents, or any correlation in the intrinsic frame, would be absorbed into the fitted r rather than detected by the procedure; the paper does not test these model assumptions. Please add a model-checking step, for example by examining Gaussianity of the increments, fitting a model with H1 ≠ H2, or comparing the observed turning-angle distribution to the fitted model with a residual-based diagnostic, before interpreting r as a physical anisotropy ratio.","section":"Supplementary Methods A and D"},{"comment":"The uncertainty intervals are defined by the ad hoc rule dCvM(r) ≤ 2 dCvM(r*) and are not calibrated confidence sets; statements such as the high-mobility CD4 state being inconclusive for anisotropy are based on this uncalibrated region rather than on a statistical test. Please calibrate the interval by simulation or replace it with a proper confidence set, such as a bootstrap or likelihood-based confidence region.","section":"Fig. 5 and Supplementary Eq. (58)"}],"minor_comments":[{"comment":"The segmentation threshold η = µ + 0.5σ is chosen without a sensitivity analysis; please report how the estimated r values change with η, since segmentation errors propagate directly into the turning-angle distributions.","section":"Methods C"},{"comment":"The cross-covariance panels (e–j) are described as using normalized increments, but the normalization procedure is not given in the caption or the Methods; please state explicitly how the normalization was performed.","section":"Fig. 3"},{"comment":"The turning angle uses the principal value of cos⁻¹, so the distribution lives on [0, π] and discards the sign of the turn; the text should note that left/right (chiral) anisotropy is not captured by this definition.","section":"Eq. (8)"},{"comment":"The alternative-distance robustness check is shown only for Nav1.6; please add analogous panels for the quantum dot and CD4 datasets to support the claim that the inferred r is not specific to the CvM distance.","section":"Supplementary Fig. S4"},{"comment":"The code is publicly available, but the experimental datasets are only available 'on reasonable request'; depositing the processed trajectories would improve reproducibility.","section":"Data and Code Availability"},{"comment":"The spelling 'Nav 1.6' in the abstract and Introduction is inconsistent with 'Nav1.6' used elsewhere; please standardize.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The theoretical and simulation parts are publishable, but the experimental claim currently lacks a goodness-of-fit test, which is essential for a paper whose headline result is that anisotropy is widespread in live cells. I would encourage the editor to request the calibration analyses described in the major comments before further consideration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the turning-angle estimator for anisotropy in randomly oriented 2D-FBM is a genuine and well-tested contribution, but the experimental half of the paper overreaches because the FBM assumption is never tested against the data. The reported r values could be best-fit shape parameters rather than physical anisotropies.\n\nWhat's new: the rotation-invariant anisotropy estimator built on turning-angle distributions. The theory is clean: Eqs. (3)-(6) derive the cross-covariance, and Eqs. (48)-(51) show rotation invariance. Simulation validation with localization noise is thorough, and the estimator recovers r well when the data are truly FBM. The GitHub code is a plus. I also give credit for checking multiple distance measures (CvM, KS, Wasserstein, L1, L2, L∞) and finding robust minima.\n\nThe soft spot is the missing goodness-of-fit step. The experimental r values come from minimizing the CvM distance against look-up tables generated under anisotropic 2D-FBM at a fixed H. The CvM statistic is never compared to its null distribution, so we don't know whether any FBM with any (H,r) is statistically compatible with the data. The stress-test note gets this right. Given H values as low as 0.06 and known static noise, the empirical distributions might not be FBM at all. In that case r≈0.16-0.41 is just a shape parameter, not evidence of cytoskeletal tracks or membrane domains. The Discussion's claim that anisotropic anomalous diffusion is widespread is not supported until the model is tested.\n\nMinor concerns: H is estimated from the first five TA-MSD points with no uncertainty propagation into r; the segmentation threshold eta is heuristic; data are 'available on request' rather than deposited. These are minor relative to the goodness-of-fit gap.\n\nBottom line: the method is real and worth a proper referee. The paper should be revised to include a parametric bootstrap or other null-distribution test for the CvM statistic, and ideally apply it to each experimental state. If the FBM model is rejected, the anisotropy claims need to be reframed.\n\nRecommendation: send to peer review. It deserves serious referee time, with the expectation of a major revision focused on model validation.","headline":"A genuinely useful rotation-invariant anisotropy estimator for 2D-FBM, but the experimental anisotropy claims need a goodness-of-fit test before they convince.","tokens_in":23647,"tokens_out":2120,"would_cite":true,"duration_ms":20990,"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":"Turning-angle distributions carry a rotation-invariant signature of spatial anisotropy, revealing anisotropic motion in live cells that covariance analysis averages away.","keywords":["anomalous diffusion","fractional Brownian motion","turning angles","anisotropy","single-particle tracking","live-cell imaging","quantum dots","neuronal membrane"],"falsifier":"Simulate an isotropic-but-non-FBM process with the same anomalous exponent—for instance a continuous-time random walk with heavy-tailed waiting times or a heterogeneous diffusion process—and feed its turning angles through the paper's look-up-table estimator. If the estimator returns $r < 1$ with a narrow uncertainty interval, the anisotropy finding would be a modeling artifact; a goodness-of-fit test of the experimental turning angles against the best anisotropic FBM distribution would settle the same question on the live-cell data.","tokens_in":22680,"feed_emoji":"📐","tokens_out":8609,"duration_ms":88631,"temperature":0.7,"pith_summary":"This paper argues that the distribution of turning angles—the angle a trajectory makes between consecutive displacement steps—is a rotation-invariant fingerprint of spatial anisotropy in anomalous diffusion. For a two-dimensional fractional Brownian motion built from independent components with diffusivities $D_1 \\ge D_2$, the turning-angle distribution depends on the Hurst exponent $H$ and the ratio $r = D_2/D_1$, but not on the orientation of the trajectory. The authors exploit this invariance to estimate $r$ from experimental single-particle tracks, obtaining clear anisotropies for quantum dots in the cytoplasm ($r \\approx 0.17$–$0.38$) and Nav1.6 channels on neuronal membranes ($r \\approx 0.16$–$0.41$), even though the ensemble cross-covariance of the increments is zero in every dataset. If the argument is right, anisotropic anomalous diffusion is a common, hidden feature of intracellular motion, and isotropic models should not be assumed by default.","feed_headline":"Turning angles expose anisotropies invisible to covariance analysis","feed_subtitle":"It estimates how much faster molecules move along one axis even from short, randomly oriented tracks.","key_machinery":"The central object is the turning angle $\\theta$ between two consecutive displacement vectors $U$ and $V$, with distribution $f(\\theta)$ derived from the joint Gaussian law of the increments, whose covariance matrix is $\\Sigma(r)$ with blocks $\\Lambda_r = 2\\operatorname{diag}(1, r)$ and off-diagonal $\\gamma_H \\Lambda_r$, where $\\gamma_H = 2^{2H-1} - 1$. The key identity is rotation invariance: $\\theta(A_\\varphi U, A_\\varphi V) = \\theta(U,V)$, so the distribution depends on $H$ and $r$ but not on the random orientation of a trajectory. This invariance lets the analyst pool randomly oriented trajectories into one histogram and estimate $r$ by comparing that histogram with look-up tables of turning-angle distributions for a fixed $H$, using the Cramér–von Mises distance.","core_discovery":"The paper's central claim is that turning-angle statistics can detect anisotropy where covariance-based statistics cannot. Writing the anisotropic process in its principal frame with independent components and diffusion constants $D_1 \\ge D_2$, the increment cross-covariance of a randomly rotated ensemble averages to zero; the turning-angle distribution, by contrast, is unchanged by rotation because $\\theta(A_\\varphi U, A_\\varphi V) = \\theta(U,V)$. The empirical distribution of turning angles can therefore be pooled over arbitrarily oriented short trajectory segments and matched, via the Cramér–von Mises distance, to simulated look-up tables for candidate $r$ values. Applied to quantum dots in the cytoplasm and to Nav1.6 and CD4 proteins on hippocampal neurons, this procedure yields $r$ estimates near 0.2–0.4 for the anisotropic states, with the CD4 high-mobility state consistent with isotropy. The discovery is not just a new estimator: it is evidence that anisotropic anomalous diffusion occurs in live cells.","pith_inferences":["A direct test of model dependence: simulate isotropic continuous-time random walks or heterogeneous diffusion processes with matching anomalous exponent and run the same look-up-table estimator; any spurious $r < 1$ would show the anisotropy estimate is not uniquely diagnostic of FBM.","Because the estimator only needs a simulated turning-angle table, it can be extended to any anisotropic random-walk model, making the method a general anisotropy assay rather than an FBM-specific one.","A systematic survey across many membrane proteins, cytoplasmic tracers, and cell types would test the paper's suggestion that anisotropy is widespread; the current evidence rests on three molecular systems.","If anisotropies of order $r \\approx 0.2$ are typical, then directionally averaged descriptions of cytoplasm and membrane transport may need to be replaced by tensor-valued effective transport parameters."],"forward_implications":["Ensemble cross-covariance functions that vanish after pooling cannot be taken as evidence of isotropy; a randomly oriented anisotropic ensemble also produces zero cross-covariance.","Anisotropy can be resolved state by state in two-state trajectories, so short segments classified by local convex hull remain analyzable.","Localization noise widens the uncertainty of the anisotropy estimate, so reporting noise levels alongside $r$ values is necessary for interpreting single-particle data.","Anomalous diffusion characterization in cells should not assume isotropy a priori; anisotropic 2D-FBM fits may be required for correct parameter estimation."],"supporting_citations":[{"why":"Supplies the correlated-component 2D-FBM construction that the principal-axis parametrization used here is shown to be equivalent to.","marker":"[39]"},{"why":"Establishes the turning-angle (directional change) distribution as a signature of complex dynamics and provides the definition of the turning angle used in the analysis.","marker":"[44,45]"},{"why":"Provides the quantum-dot cytoplasmic trajectories, the two-state segmentation, and the earlier finding that their motion follows switching FBM.","marker":"[30]"},{"why":"Contributes the local convex hull method used to segment trajectories into low- and high-mobility states.","marker":"[62]"},{"why":"Characterized the Nav1.6 and CD4 membrane protein dynamics and their two diffusive states, which the current analysis re-examines for anisotropy.","marker":"[53]"},{"why":"Supplies the Cramér–von Mises two-sample criterion used to match empirical turning-angle distributions to look-up tables.","marker":"[63]"},{"why":"Provides the circulant embedding method used to simulate realizations of anisotropic 2D-FBM for the look-up tables.","marker":"[70]"}],"fun_headline_variants":["Turning angles detect anisotropy invisible to covariance analysis","Short tracks show anisotropy when analyzed via turning angles","Turning-angle statistics uncover hidden anisotropy in live cell motion","Anisotropic diffusion in cells exposed by turning-angle analysis"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The estimate of $r$ is only physically meaningful if the true motion is fractional Brownian motion with one Hurst exponent in both directions and independent components along the intrinsic axes; under a different underlying process, the best-fitting $r$ may not correspond to any real directional bias in the cell.","fun_headline_variants_meta":{"raw":{"variants":["Turning angles detect anisotropy invisible to covariance analysis","Short tracks show anisotropy when analyzed via turning angles","Turning-angle statistics uncover hidden anisotropy in live cell motion","Anisotropic diffusion in cells exposed by turning-angle analysis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000743,"raw_usage":{"total_tokens":3292,"prompt_tokens":899,"completion_tokens":2393,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":515,"completion_tokens_details":{"reasoning_tokens":2332}},"tokens_in":515,"tokens_out":2393,"duration_ms":22102,"temperature":1.0,"reasoning_tokens":2332,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T00:35:50.801145+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate an isotropic-but-non-FBM process with the same anomalous exponent—for instance a continuous-time random walk with heavy-tailed waiting times or a heterogeneous diffusion process—and feed its turning angles through the paper's look-up-table estimator. If the estimator returns $r < 1$ with a narrow uncertainty interval, the anisotropy finding would be a modeling artifact; a goodness-of-fit test of the experimental turning angles against the best anisotropic FBM distribution would settle the same question on the live-cell data.","supporting_citations":[],"review_version":1}