Bottom line · from claim 1 of 6: The paper's own evidence supports the main result. We did not rerun it.

Claims, one verdict each

Each claim is quoted from the paper or its repository. Claim 1 is the main result; the bottom line comes from its verdict only. Open “Evidence” for the numbers, code and log lines behind each verdict.

1. 01○Supported, arithmetic from stated constants; measured ranges not checked: Tether force and diameter

   Checked derivation + Read code

   Paper: First, classical tube mechanics predicts a tether force of 24.3 pN and a diameter of 155 nm without free fitted parameters, matching measurements (10–40 pN; 50–200 nm)

   Abstract; "TNTs are ordinary membrane tubes under tension"

   With the paper's constants, bending rigidity κ = 0.30 pN·µm and tension σ = 25 pN/µm, the standard tube formulas give a force of 24.33 pN and a diameter of 154.9 nm. Both constants are cited standard values, not fitted, and the code uses the same values. The paper itself calls this "a no-free-fit consistency check, not a parameter-free forecast".

   Evidence: the formulas

   force     F  = 2π·√(2σκ) = 2π·√(2 × 25 × 0.30) pN = 2π·√15 pN = 24.33 pN
   diameter  2r = 2·√(κ / 2σ) = 2·√(0.30 / 50) µm = 0.1549 µm = 154.9 nm
   

2. 02○Supported: Median lifetime

   Checked derivation

   Paper: predicting a median lifetime of 60–62 min (95% CI 37–99 min) within the observed 30–120 min

   Abstract; lifetime paragraph

   The closed-form value τ_dep · ln 2, with the paper's τ_dep = 90 min, is 62.4 min. Our reruns of the authors' scripts also printed the result of the Monte Carlo behind the 60.4 min median: 60.4 min (95% CI 36.8–98.9 min) with the authors' seed, and 60.5 min with each of the other two seeds. This claim was outside the claims we fixed for rerunning before the runs, so we set no tolerance for it.

3. 03○Supported, uniform tube only; beaded tubes not checked: Electrotonic length

   Checked derivation

   Paper: the contact network is electrotonically compact (λ ≈ 158 µm ≫ L)

   Abstract; "Electrotonically compact wires"

   With R_m = 10⁴ Ω·cm², R_i = 100 Ω·cm and a 50 nm tube radius, λ = √(R_m·r / 2R_i) = 158.1 µm. The paper uses a 100 nm diameter here, the middle of the observed range, rather than its own predicted 155 nm. With 155 nm, λ is about 197 µm, so the conclusion λ ≫ L holds either way.

4. 04✕Contradicted, for the stated ±0.002 precision: 2D percolation threshold

   Rerun · 3 seeds + Read code

   Paper: finite-size scaling in two and three dimensions constrains percolation thresholds to p_c = 0.257 ± 0.002 (2D)

   Abstract; finite-size scaling section

   We found no code that computes 0.257. The script uses it as a fixed input (pc_inf = 0.257, paper_figs_part2.py, line 136). Fitting the paper's stated finite-size form to the authors' own simulation output gives 0.2525. Fresh simulations with two more seeds give 0.2482 and 0.2370. All three are outside 0.257 ± 0.002, and they differ from each other by up to 0.015. The raw thresholds at the largest size, N = 8000, also spread by about 0.01 across seeds (0.2475 to 0.2573). That spread, rather than any single value, is what a ± 0.002 claim cannot absorb.

   To change this verdict: publish the code and data that produce 0.257 and its ± 0.002.

   Evidence: the three seeded reruns

   Seed	p_c(500)	p_c(2000)	p_c(8000)	Fitted p_c(∞)
   20260907 (the authors')	0.2462	0.2460	0.2514	0.2525
   20260917	0.2446	0.2447	0.2475	0.2482
   20260927	0.2879	0.2617	0.2573	0.2370 (see note)

   The fit is the form stated in the paper, p_c(N) = p_c(∞) − c·N^(−3/8), applied by least squares to the three sizes. Each p_c(N) is the point where P∞ crosses 0.5, as defined in the authors' script. With the authors' seed, the fresh run reproduces the authors' saved percolation curves (paper_fss_curves.json) byte for byte, so the 0.2525 is their own data under their own stated form.

   Note on seed 20260927: here p_c(N) falls as N grows (c < 0), so the fitted line extrapolates downward. Its 0.2370 is an artefact of the form, not a separate estimate of the threshold.

   Evidence: other fit forms (exploratory, after the fact)

   Other reasonable fit choices on the authors' own output give 0.249 to 0.262. So a different fit can give a value near 0.257, but the stated method does not give it with the stated precision of ± 0.002.

5. 05–Not checked: 3D percolation threshold

   Paper: ≈ 0.169 (3D; geometry-dependent)

   Abstract

   Our runs of the authors' scripts printed a 3D estimate of 0.1688. The script computes it from 3D percolation curves saved in the package, which we did not regenerate. This claim was outside the claims we fixed for rerunning before the runs, so we set no tolerance for it.

6. 06✓Reproduced, code matches paper; printed Eq. (5) does not: Optimal connectivity shift

   Rerun · 3 seeds + Read code

   Paper: an optimal connectivity p* = a/2b that falls from 1.0 (healthy) to 0.32 (late pathology)

   Abstract; competing cargo flows section

   All three seeded runs of the authors' code gave p* = 1.0 (healthy), 0.65 (early pathology) and 0.32 (late pathology), matching the paper. But the code does not implement Eq. (5) as printed. The paper calls this series "the full K-cargo numerical solution of Eq. (5)". The interference term of Eq. (5) sums over all pairs of cargo classes (k ≠ j); the code subtracts only the mitochondria × α-synuclein pair. With the code's other inputs, Eq. (5) as printed gives about 0.46 (early) and 0.25 (late).

   To change this verdict: publish code that solves Eq. (5) as printed and gives 0.65 and 0.32, or correct Eq. (5) or the text to match the code.

   Evidence: Eq. (5) as printed, next to the code

   The code's objective (paper_figs_part4.py, lines 24–25; paper_figs_part2b.py, line 111, which computes the series, subtracts the same single pair):

       return terms["mitochondria"] + terms["Ca"] + terms["alpha-syn"] + terms["prions"] \
              - lam*np.abs(terms["mitochondria"])*np.abs(terms["alpha-syn"])
   

   Eq. (5) as printed is W(p) = Σ_k w_k α_k f_k(p) − λ Σ_{k≠j} |α_k α_j| f_k f_j. Its interference term carries no fitness weights w_k. Our own calculation keeps the code's four cargo classes, their weights, rates and forms, the three load scenarios and λ = 0.3, and evaluates the interference term as printed. The code applies each load scenario to every term of the class; we read it as scaling the loading rate α_k. Under that reading, the code's single pair equals the printed term for that pair. p* is the p in [0, 1] that maximises W, on the 400-point grid that paper_figs_part2b.py uses.

   Interference term	Healthy	Early	Late
   The code: mitochondria × α-synuclein only	1.0	0.649	0.321
   Eq. (5), each unordered pair k ≠ j once	1.0	0.461	0.253
   Eq. (5), each ordered pair (k, j) and (j, k)	0.815	0.381	0.218

   The 4,001-point grid of paper_figs_part4.py moves each value by about 0.001. Under either reading of the sum, Eq. (5) as printed misses the paper's 0.65 and 0.32 by more than our tolerance of 0.01 for p*.

What we did not check

These limits carry the same weight as the findings above.

• The 3D threshold, the lifetime simulation and the five proposed experiments. None of these was among the claims we fixed for rerunning before the runs. The scripts printed values for the first two (see those claims); we did not regenerate the saved 3D curves.
• Whether the cited constants and measured ranges are right. We checked that the results follow from the stated inputs, not the inputs themselves.
• The scripts needed small edits to run. We used numpy 1.26.4, the version pinned in the package's own code/requirements.txt. Three of the six scripts stopped with errors. We ran them after three small edits, which are published as a diff. One edit replaces np.trapezoid, a name that numpy 1.26.4 does not have, with np.trapz (parts 2b and 4). The other two change function calls in part 2. The package also ships saved simulation results, which the part 2 script reads in place of a new simulation. We removed them so that the simulation ran fresh. With the authors' seed, 17 of the 19 saved files are byte-identical to the fresh run, and one differs only by rounding (at most 6 × 10⁻¹⁶).
• Why the saved calibration file differs from the script. The saved paper_beta_corners.json lists nine parameter corners with P∞ from 0.305 to 0.964, close to the paper's stated range of 0.31 to 0.96. The shipped script writes four corners with P∞ from 0.022 to 0.045. One visible difference: the script simulates T = 3.0 (paper_figs_part4.py, line 134), while the paper states T = 180 min. We did not test whether this explains the gap. No claim we checked uses this file.
• More seeds. Three seeds show the spread; they do not pin down its size.
• The rest of the paper. This report checks the claims listed above, looks up every entry in the reference list and reruns code only where it says so. It is not a full review of the paper. In the same sections as the claims above, this report does not assess the tension range the paper derives from the observed diameters, the beaded-tube correction or the sensitivity scan of the optimal connectivity (Fig. 16). Elsewhere it does not check the rupture-length model, the first-passage times, the hysteresis and memory results, parallel-tube loading or the comparison with mycorrhizal networks (Table 3). On 2 Nov 2026 we also ran the paper through the free tier of Google's Deep Review. It raised further comments, which it rated critical or major, on parts this report does not assess. For comments on the whole paper, you can use a general AI review tool yourself; Deep Review (https://research.google/deepreview/) offers free weekly reviews with a Google account.

How we checked

Repository

https://github.com/zhangsongbo2026/Tunneling-nanotubes

Commit

5fd7953d1a204b15cd3c32e91033c81172141cff (the code is in a zip inside the repository)

Environment

Python with numpy 1.26.4, as pinned in the package's code/requirements.txt; contained runner with no network.

Commands

the six figure scripts, paper_figs_part1, part2, part2b, part3, part4 and part5, run in order. Edits: np.trapezoid to np.trapz (parts 2b and 4); W_func(p_grid, **s) to *s (part 2); f_form(form, p) to f_form(form, p_grid) (part 2, figure 8b).

Seeds

20260907 (the authors'), 20260917, 20260927.

Runtime

about 10 minutes per seed (615 s for the authors' seed).

Time spent

writing this draft from the finished rerun outputs took about 1 minute. This excludes the reruns (see Runtime) and the claim list, which were prepared separately.

Outputs and logs

the per-seed outputs, the logs, the edit diffs and the fit script are published with this report: outputs.

References

Found

34

Matched automatically

20

Not matched

14

Checked by hand

14, and all 14 lead to real papers. 7 are the cited paper with a shortened or reworded title. For 6, the journal, volume and pages cited hold a paper with a different title. Three of these are on a different subject: Desir 2016 (Oncotarget 7, 43150) is on hypoxia in ovarian cancer cells, not stroke; Karlsson 2001 (Nature 409, 150) is on networks of lipid nanotubes, not dendritic cells; Lou 2012 (PLoS ONE 7, e33093) is on mesothelioma, not glioma. Three are on a related subject, with a different title: Gousset 2009, Jana 2022 and Zamberlan 2025. For 1, Victoria and Zurzolo 2017, the cited title matches the same authors' paper in a different journal (J. Cell Biol. 216, 2633–2644).

Could not find

0

Author reply

Changes to this report