-
Reproducibility is currently blocked by missing MD provenance and internal inconsistencies in trajectory length, frame counts, and analyzed time windows (Sec. 2.1 vs Sec. 3.1–3.3). The manuscript variously reports $\sim 65,\!000$ frames ($1.3\,\mu\text{s}$ at $20\,\text{ps}$), an analysis slice of frames $5000$–$64999$ ($60,\!000$ frames $\approx 1.2\,\mu\text{s}$), but also $66,\!771$ frames and a $100$–$1435.42\,\text{ns}$ window, and later $66,\!770$ inter-frame transitions. In addition, essentially no MD setup details are given (force field, water/ions, thermostat/barostat, PBC handling, equilibration), preventing assessment or replication.
Recommendation: In Sec. 2.1, provide a single consolidated MD description: peptide termini/protonation, force field (and version), solvent model, ions/concentration, box size/shape, ensemble, $T/P$, thermostat/barostat parameters, constraints, PME/cutoffs, equilibration and production lengths, and confirm periodic boundary conditions were used in contact calculations. Then reconcile the timeline by explicitly tabulating: total simulated time; frame stride ($20\,\text{ps}$) and number of stored frames; analysis slice start/end frame indices (inclusive/exclusive) and corresponding times; and the resulting number of transitions ($=$ analyzed\_frames $- 1$). Update Sec. 3.1–3.3 so all reported statistics, figures, and event counts refer to the same, consistent slice.
-
Core descriptive results about aggregate integrity are internally contradictory, casting doubt on downstream event counts and pre-splitting signatures (Sec. 3.1 vs Sec. 3.2–3.3). Sec. 3.1 claims the system remains a single $30$-peptide aggregate with LCC size $30.0 \pm 0.0$ and aggregate count $1.0 \pm 0.0$, but Sec. 3.2.1 reports average connected components $1.17 \pm 0.40$ and average LCC size $28.89 \pm 2.72$, and Sec. 3.3 reports $1184$ splitting events—these cannot all be simultaneously true under a single graph/contact definition and time window.
Recommendation: Audit and reconcile the definitions and computations used in Sec. 3.1 vs Sec. 3.2–3.3. Explicitly state whether: (i) “aggregate count” uses a different connectivity criterion than “connected components” (e.g., filtering small detachments), (ii) different time ranges were used, or (iii) earlier summary numbers are incorrect. Ensure figures and captions (notably Fig. 1/2) explicitly state whether statistics are computed for the largest aggregate only or for all aggregates per frame, and that the narrative about frequent transient splitting is consistent with the reported LCC/component time series.
-
The operational definition of “splitting events” is under-specified and likely overcounts topological flicker caused by intermittent contacts at 20 ps resolution and a binary edge rule (“any contact $< 4.5\,\text{\AA}$ implies connected”) (Sec. 2.5, Sec. 3.3). With such a criterion, momentary loss of a single marginal contact can disconnect the CG graph, producing many short-lived ‘splits’ that may not correspond to physically meaningful fragmentation. This threatens the central claim that pre-splitting graph signatures are specific and mechanistic rather than tautological consequences of noisy edge disappearance.
Recommendation: In Sec. 2.5, precisely define the production split criteria (minimum parent size, minimum daughter sizes, treatment of single-peptide detachments, handling of $>2$ fragments, and merge/split disambiguation). Add persistence/hysteresis: e.g., require fragments to remain disconnected for $\geq K$ consecutive frames (or $\geq \tau$ time), and/or require CG connectivity via edges with weight $\geq w_{\rm min}$ (minimum contact count), and/or smooth contact weights (EMA/rolling window). In Sec. 3.3 (or Appendix), run a sensitivity analysis varying (i) distance cutoff (e.g., $4.0/4.5/5.0\,\text{\AA}$), (ii) weight threshold $w_{\rm min}$, and (iii) persistence $K$, and report how event counts, event durations, and key pre-splitting trends (Sec. 3.4) change. Also quantify how many splits rejoin within $0.1$–$1\,\text{ns}$ to separate transient flicker from sustained fragmentation.
-
Statistical comparisons between pre-splitting windows and control windows are insufficiently specified and likely invalidate reported extreme $p$-values due to dependence, overlap, and unclear control selection (Sec. 2.6, Sec. 3.4). The manuscript does not clearly define the control-window algorithm, whether windows overlap across events (highly likely with $1184$ events), whether tests are one-/two-sided, how normality was assessed, whether multiple comparisons were corrected, or how temporal autocorrelation and repeated events from the same evolving aggregate are handled.
Recommendation: Expand Sec. 2.6 into a fully specified protocol: define event windows (length, alignment time, overlap rules) and control windows (how ‘stable’ is defined quantitatively, required split-free duration, size matching, time matching, and exclusion buffers around events). Use dependence-aware inference: e.g., event-level aggregation (one value per event), enforce a refractory period between events, and/or use block bootstrap / permutation respecting temporal correlation. Report effect sizes (e.g., Cohen’s $d$ / Cliff’s delta) and confidence intervals alongside $p$-values, and apply multiple-hypothesis correction (FDR/Bonferroni) across metrics. In Sec. 3.4, report sample sizes (number of events retained after de-overlap, number of control windows) and robustness to window length (e.g., $0.5/1/2\,\text{ns}$).
-
Several key metrics are ambiguously defined or mathematically inconsistent with reported values, undermining interpretability and reproducibility—most notably weighted density exceeding $1$ (Sec. 2.4.1 vs Sec. 3.2.1) and unclear conventions about undirected symmetry and double-counting of weights (Secs. 2.2–2.4). Weighted density is described as a normalized quantity bounded by $1$, yet CG weighted density is reported around $2.65 \pm 0.32$. Additionally, shortest-path-based centralities in weighted graphs depend on whether weights are treated as ‘strengths’ or ‘costs’, which is not clarified (Sec. 2.4).
Recommendation: Provide the exact formulas used in code for: (i) weighted density (including whether sums are over $i<j$ edges or over all $i \neq j$ adjacency entries, and what is used as the denominator), and (ii) any centrality that depends on shortest paths (betweenness/closeness): clarify whether weights are inverted to convert strengths into distances/costs. If the computed quantity is not bounded by $1$, rename it (e.g., mean weight per possible edge) or correct the normalization and recompute affected results/figures (Sec. 3.2–3.4). State explicitly that graphs are undirected with symmetric adjacency and zero diagonals, and document how isolated nodes/components are handled in each metric.
-
The FG analyses central to the mechanistic claim are under-specified: it is unclear which FG node/edge set is used for FG Fiedler calculations and how ‘bridging strength’ is defined and tracked backward in time (Sec. 2.3.2, Sec. 2.4.2, Sec. 2.7, Sec. 3.4.2–3.5). This ambiguity also affects interpretation of the reported ‘counterintuitive’ increase of FG Fiedler value before splitting, which could be a selection effect (e.g., LCC becomes smaller/denser when peripheral residues disconnect).
Recommendation: In Sec. 2.3.2 and Sec. 2.4.2, explicitly define the FG graph object used in each analysis: whole-system FG graph vs FG restricted to residues belonging to the parent CG aggregate; whether metrics are computed on the FG largest connected component (and how it is identified) or the full FG graph; and whether the combinatorial or normalized Laplacian is used (with explicit formula and isolated-node convention). In Sec. 2.7, formalize bridging strength: define how the two ‘future fragments’ are identified at the split frame (e.g., two largest CG components), how each residue is assigned to a future fragment at earlier times, and exactly which FG edges are summed (inter-fragment only; inter-peptide only; weight aggregation). In Sec. 3.4.2, report FG LCC size/weight changes alongside FG Fiedler aligned to splits to distinguish genuine compaction from component-selection artefacts; include at least one illustrative example graph/snapshot.
-
The manuscript demonstrates ‘predictive signatures’ mainly via distribution shifts and $p$-values, but does not quantify practical predictive value or compare against simpler baselines, leaving the bigger-picture utility unclear (Sec. 3.4). Relatedly, bridging-contact decline may be partially tautological if it is too directly coupled to the split definition (connectivity/contact loss).
Recommendation: Augment Sec. 3.4 with an explicit prediction assessment: classify windows as ‘split within next $1\,\text{ns}$’ vs ‘no split’ using thresholds or a simple model (logistic regression / survival model), and report ROC-AUC / PR-AUC, precision/recall at meaningful operating points, and lead-time distributions. Compare graph metrics (CG $\lambda_2$, density, bridging strength) to baseline structural descriptors (e.g., total inter-peptide contacts, number of CG edges, aggregate $R_g$, SASA, mean peptide degree/contact number). Use event-level cross-validation or blocked time splits to avoid leakage. Explicitly discuss to what extent each metric provides information beyond ‘contacts are decreasing’ and where the graph framework adds mechanistic value.
-
Claims about robustness and generality are overstated given evidence from a single peptide sequence and a single simulation setup, with limited discussion of force-field/contact-definition dependence and finite sampling (Sec. 3.6, Sec. 4).
Recommendation: Tone down general statements in the Abstract, Sec. 3.6, and Sec. 4 to clearly separate (i) what is shown for KYFIL under the specific MD/contact definition from (ii) what is proposed as a general framework. Add a Limitations paragraph noting dependence on: force field/solvent model, concentration/box conditions, finite $1.3\,\mu\text{s}$ sampling, the $4.5\,\text{\AA}$ cutoff and time resolution, and the use of undirected distance-based contacts (no orientation/energetics). Outline concrete next validations (other sequences, variants, conditions, force fields; experimental or replicate simulations).