{"id":"6bcd6b55-d2e0-41bd-a2f2-33ca42bffb00","arxiv_id":"2505.02717","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A new TDA visualization, persistence weighted death simplices, locates topological holes in the original cell images and, with weighted persistence images, separates healthy from diseased tissue samples.","lead":"This paper introduces a way to draw the holes that cell patterns form in tissue images, so topological analysis can be seen directly on the original image. It shows the drawings and related measurements can separate healthy from diseased spleen and lung samples, and they reveal infiltration and structural changes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PWDS localization rests on an unproved assertion that the death simplex lies inside the feature; in non-convex or multi-candidate geometries this may fail, so the biological readings in Figs. 5 and 7 are not yet supported.","rationale":"The reader's weakest assumption is the same one I find most load-bearing: the PWDS visualization is the paper's advertised novelty, and its interpretability rests entirely on the death simplex localizing the feature. If the simplex can sit outside the void or jump between candidate triangles, the qualitative biological claims about red pulp infiltration and lung cavities are not supported by the pipeline. The instability is acknowledged, but instability alone would be tolerable if there were a theorem or a robust empirical check showing that all possible death simplices for a feature lie in the same small region; neither is provided. I also considered the multiple-comparison issue, which is a serious limitation of the application-level claim that '10 cell types ... form significantly different spatial patterns.' However, the methodological PWDS claim is more central to the paper's novelty, and the proposed synthetic test can settle it directly because PH is computable, the ground-truth void is known, and failure is measurable. The reader's conditional verdict remains appropriate; I would not change it, but the condition should be explicit: PWDS needs either a proof, with clear hypotheses on the filtration and geometry, or a benchmark showing that localization error is small before the biological readings in Figures 5 and 7 are used as evidence.","tokens_in":20124,"tokens_out":9534,"duration_ms":121321,"concrete_test":"Build a synthetic benchmark with known 1-cycles: sample points on non-convex closed polygons (e.g., L-shaped or C-shaped annuli) at several densities, with and without internal noise; compute alpha-filtration persistence with gudhi; for each degree-1 feature above a persistence threshold, extract the death simplex and test whether its barycenter and vertices lie inside the ground-truth void (the polygon interior). Repeat under small random perturbations of the points and over many seeds, and repeat for the witness filtration with random landmarks as used in the paper. If a non-negligible fraction (e.g., >5%) of death simplices fall outside the void, or jump to a different void under perturbation, then the claimed 'always roughly inside' property fails and PWDS cannot be used to localize features without an additional localization step.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 'Visualisation of PH' defines PWDS by plotting the death simplex returned by gudhi and asserts: 'the death simplex will always correspond to a simplex within the interior of the topological feature, roughly indicating its position in the original data.' The paper explicitly admits that the death simplex is unstable, because two empty triangles with similar circumradii can swap under perturbation and simultaneous deaths are resolved by arbitrary internal ordering. No proof of the interior-location assertion is given, and it is not a standard theorem of persistent homology. In an alpha filtration, a degree-1 feature dies when a 2-simplex completes a 2-chain homologous to the feature; the simplex that completes the chain need not lie geometrically inside the void when the void is non-convex, when the feature is represented by a non-simple cycle, or when the triangulation contains triangles that cross the boundary. For witness filtrations, landmarks are random subsets, so a paired landmark simplex is even less tied to the Euclidean location of the feature. Since the red/blue triangle readings in Figures 5 and 7 depend on interpreting death simplices as localizing features, the PWDS claim is load-bearing and currently unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a pipeline for topological analysis of multiplexed imaging data. It introduces a visualisation, persistence-weighted death simplices (PWDS), which plots the simplex that kills each persistent homology feature, coloured by birth-value thresholds (proximal/distal) and weighted by persistence. It also studies persistence-image weightings and combines nine vectorisations with k-means clustering to compare healthy and diseased tissue in two datasets (lupus spleen CODEX and COVID-19 lung IMC). The authors report that 10 spleen cell types plus red/white pulp show different spatial patterns between health and disease, and that endothelial-cell descriptors separate COVID-19 disease stages.","tokens_in":20389,"tokens_out":6739,"duration_ms":81269,"significance":"If the central visualisation claim were established, PWDS would be a useful, interpretable and computationally cheap complement to persistence diagrams for spatial biology, localising multiscale features in the original tissue while avoiding the cost of multiparameter persistence. The paper is clearly written, uses standard persistent homology computations, and provides code and data at an accessible repository. The comparison of persistence-image weightings is systematic, and the authors are transparent about the instability of death simplices. However, the method's key assumption that the death simplex lies inside the detected feature is not proved, and the biological conclusions rest on in-sample screening of hundreds of descriptors without correction.","major_comments":[{"comment":"The assertion 'the death simplex will always correspond to a simplex within the interior of the topological feature' is load-bearing for the entire PWDS method and is stated without proof or reference. In an alpha filtration, a degree-1 feature is killed when a 2-simplex completes a 2-chain homologous to the feature; the geometric location of that simplex is not constrained to lie inside the void, and it can fail to do so for non-convex or non-simple cycles. For witness filtrations with random landmarks, the coupling between a landmark simplex and the Euclidean location of the feature is even weaker. Since the biological readings of Figures 5 and 7 (triangles as infiltrating cells or cavities) depend directly on this localisation property, the central methodological claim is currently unsupported. The authors should either prove the interior-location claim for the planar alpha case under suitable genericity/convexity assumptions, or replace the death-simplex visualisation with computed representative or optimal cycles, or validate PWDS on synthetic non-convex examples and state the failure modes.","section":"Visualisation of PH"},{"comment":"The screening procedure selects descriptors that yield correct k-means clusterings from 383 descriptors for the spleen and 548 for the lung, with no correction for multiple comparisons and no hold-out validation. The phrase '10 cell types ... form significantly different spatial patterns' is therefore not supported as a significance claim: under a null model in which descriptors are independent of the disease label, the maximum silhouette or Rand score over hundreds of descriptors will be high by chance. The problem is compounded by the fact that bprox, bdist, p0, and the persistence-image variance and pixel size are estimated from the same samples. The authors should add a permutation null (e.g., shuffling sample labels), nested cross-validation, or an explicit false-discovery control, and should present the distribution of scores across all descriptors rather than only the selected ones.","section":"Results: Topological vectorisations of 10 cell types distinguish lupus samples"},{"comment":"The introductory claim that PWDS 'replicat[es] the main advantage of MPH over PH at no additional computational cost' is not tested anywhere in the paper. Multiparameter persistent homology is never computed on these data or on a synthetic example, so the equivalence is unsupported. The phrase 'features in point clouds at different densities' is also not formalised; the colour thresholds are quantile-based and data-dependent. The authors should either remove the MPH-replication claim or substantiate it with an explicit comparison (for example, a synthetic two-density example where MPH and PWDS are both computed), and they should define 'different densities' precisely.","section":"Introduction"}],"minor_comments":[{"comment":"The displayed definition of blue features reads '{(b,d )∈D| b > bdist⟩}' with a stray angle bracket; also, bprox and bdist are introduced before their quantile-based definitions, which makes the threshold rule harder to follow.","section":"Visualisation of PH"},{"comment":"The description of wflat is clear, but the nearest-neighbour scale s used to set p0 is estimated from the same samples under analysis; the sensitivity of the reported clusterings to s and p0 should be reported.","section":"Vectorisations"},{"comment":"The witness filtration is described without stating the number and selection rule for landmarks beyond 'chosen randomly'; since witness results appear in Tables 2-4, these parameters should be specified.","section":"Methods"},{"comment":"The normalisation by the total number of features is not unit-area normalisation; although the text notes this, the term 'normalised' may still mislead readers and could be renamed.","section":"Normalised Betti curves"},{"comment":"The caption states that thresholds are computed per cell type, while the main text defines them as averages across samples for each point cloud; the dependence of bprox and bdist on the point cloud should be stated consistently.","section":"Figure 5"},{"comment":"Several listed descriptors have very low silhouette scores (e.g., 0.173) and are still presented as 'correct'; a reporting threshold or a clarification of what 'correct' means in these cases would strengthen the table.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The main risk is that PWDS becomes a visualisation with an unproven localisation property, while the biological conclusions are selection artefacts from screening many descriptors without correction. I would encourage a revision that proves or constrains death-simplex localisation in the planar alpha case and adds a null-model or cross-validated analysis. If the authors cannot supply these, the claims should be scaled back to exploratory findings."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know about this paper: it gives you a new way to map persistent homology features back onto tissue images, and it tests several persistence-image weightings that emphasize different scales. The core trick is simple—plot the death simplex, color by birth threshold, weight by persistence—and that is genuinely useful for exploratory spatial biology.\n\nWhat is actually new: PWDS as a combination is new, and the weightings w3 = bp and w4 = bp^2 are specific variants I have not seen in the vectorisation literature. The paper applies these to two public multiplexed imaging datasets and shows that even simple descriptors (Betti curves, percentile summaries) separate healthy from diseased samples. Code and data are public, and the pipeline is described in enough detail to reproduce. The CD4+ T cell example, where density is unchanged but topology differs, is a nice illustration of why PH can add something.\n\nThe soft spots are real, and they are in two places. First, the load-bearing assertion that the death simplex lies inside the topological feature is asserted, not proved. The stress-test note is right: in a non-convex void, or when a feature is represented by a non-simple cycle, the simplex that completes the homology chain need not sit geometrically inside the void. The paper explicitly admits the death simplex is unstable, but then relies on its location for the red/blue triangle readings in Figures 5 and 7. That part is currently a heuristic, and it should be presented as one. Second, the statistical screening is over-claimed. Hundreds of descriptors are tested on 9 spleen or 30 lung samples, with thresholds and persistence-image parameters chosen on the same data. No multiple-comparison correction, no held-out validation. The phrase \"significantly different spatial patterns\" is not supported as written.\n\nNeither flaw kills the paper. The vectorisation results do not depend on death simplex localization, and the Betti curve analyses stand on their own. But the biological interpretations in the PWDS figures need either a proof of the interior-location claim (which I doubt exists in general) or an explicit re-framing as an exploratory heuristic. The statistical claims need toning down and, ideally, a validation cohort.\n\nWho this is for: TDA practitioners in spatial biology who want an interpretable visualisation and a few more vectorisation options. It deserves a serious referee, not a desk reject, but the revision should require the localization claim to be demoted to a heuristic and the statistical language to match what was actually done.","headline":"PWDS is a cheap, interpretable visualization and the persistence-image weightings are worth having, but the death-simplex localization claim is unproved and the biological claims are oversold; still, it deserves a serious referee.","tokens_in":20900,"tokens_out":1881,"would_cite":true,"duration_ms":22622,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["55N31","62R40"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper's central claim is that plotting the death simplex back onto the tissue image, coloured by birth and weighted by persistence, visually and statistically separates healthy from diseased cell patterns.","keywords":["persistent homology","multiplexed imaging","spatial biology","persistence weighted death simplices","persistence images","Betti curves","lupus spleen","COVID-19 lungs"],"falsifier":"Take a synthetic loop with several nearly empty triangles of similar circumradius inside it and perturb the points slightly so the death simplex jumps between triangles; if the chosen triangle can land far from the feature's interior, the localisation claim fails. A direct check is to compute exact feature boundaries on synthetic loops and measure the Hausdorff distance between those boundaries and the PWDS triangles; large distances would falsify the claim that the triangles indicate feature positions.","tokens_in":19920,"feed_emoji":"🔬","tokens_out":11747,"duration_ms":122908,"temperature":0.7,"pith_summary":"Persistent homology summarises the multiscale shape of a cell cloud, but its output does not say where in the tissue a detected loop or void sits. This paper argues that the simplex which kills a feature—its death simplex—can be plotted back onto the original image, coloured by feature birth and weighted by persistence, to give a visualisation (PWDS) that locates biological structures and separates low-density from high-density features at no extra computational cost. The paper also tunes persistence-image weightings so that vectorised topological summaries emphasise either small-scale dense-tissue patterns or large-scale noisy loops. Applied to a lupus mouse spleen CODEX dataset and a COVID-19 human lung IMC dataset, the pipeline finds that 10 spleen cell types plus red and white pulp cells form spatial patterns that differ between healthy and diseased samples, and that endothelial-cell topology separates COVID-19 disease stages. If the claims hold, the payoff is an interpretable and computationally cheap route from multiplexed tissue images to multiscale spatial biomarkers.","feed_headline":"Coloured triangles expose disease-linked cell patterns in tissue","feed_subtitle":"Plotting the death simplex back onto the image separates healthy from diseased samples in lung and spleen","key_machinery":"The death simplex is the simplex that kills a homological feature when it enters the filtration (for a void in degree 1, the last empty triangle that fills it in). It is unstable under small perturbations of the points, but the paper assumes it lies inside the feature, so it gives an approximate location. PWDS draws these simplices in the original coordinates and assigns colour by birth thresholds $b_{\\mathrm{prox}}=\\langle P_{90}\\rangle$, $b_{\\mathrm{dist}}=\\langle P_{98}\\rangle$ and intensity by persistence $d-b$. The other machinery is the persistence-image vectorisation: a persistence diagram is converted to birth–persistence coordinates, convolved with Gaussians weighted by $w_{\\mathrm{flat}}$, $w_1=p$, $w_2=p^2$, $w_3=bp$, $w_4=bp^2$, and integrated on a grid; together with normalised Betti curves and percentile statistics these feed k-means clustering.","core_discovery":"The central discovery is that the death simplex—the simplex whose addition to the filtration destroys a persistent feature—can serve as a proxy for the feature's location in the original point cloud. By plotting these simplices on the tissue image, coloured red for features with small birth values (formed by proximal cells), blue for features with large birth values (formed by distal cells), and purple for intermediate births, with colour intensity proportional to persistence, PWDS turns a persistence diagram into a spatial map. The paper claims this reproduces the main advantage of multi-parameter persistent homology—distinguishing features at different densities—while staying within ordinary one-parameter persistent homology and hence at no additional computational cost. On the vectorisation side, the paper proposes persistence-image weightings $w_1=p$, $w_2=p^2$, $w_3=bp$, $w_4=bp^2$ alongside a flat weighting, so that summaries can emphasise large-scale structure or noisy loops; for dense tissues like the red pulp a stronger weight such as $w_2$ is needed before large-scale features become visible. The application claims are that PWDS and normalised Betti curves reveal red-pulp infiltration into white pulp in lupus and the loss of alveolar cavities with dense cellular infiltrate in COVID-19 lungs, and that topological descriptors separate healthy from diseased samples for 10 spleen cell types plus the two pulp groups, with endothelial cells giving the best separation of COVID-19 disease stages.","pith_inferences":["The paper does not test whether the PWDS red/blue counts exceed what a null model with randomly relabelled cell types would produce; such a permutation test would strengthen the claim that the observed infiltration signal is spatial rather than compositional.","Because death simplices are unstable, quantitative use of PWDS may need aggregation over many features or a smoothed version; the paper presents PWDS as a visualisation, so an aggregate PWDS statistic is a natural next step.","The birth-value thresholds are set by averaging the 90th and 98th percentiles across samples; an adaptive per-sample or scale-free threshold could remove this heuristic when PWDS is applied to new tissues.","The 'no additional computational cost' claim could be tested directly by running PWDS and a multi-parameter persistent-homology pipeline on the same dataset and comparing runtime and density-distinguishing power."],"forward_implications":["PWDS gives a spatial readout of persistent homology for any 2D point cloud, so cell-type-specific tissue architecture can be inspected directly instead of only through summary diagrams.","The birth-threshold colouring distinguishes features formed by proximal versus distal cells, so infiltration or exclusion phenomena become visible as blue triangles appearing inside large red voids.","Weighted persistence images can be tuned to the scale of interest; the paper shows that for dense populations a strong weight like $p^2$ is needed to expose large-scale features that a flat or linear weight masks.","Normalised Betti curves detect changes in the density of dense cell regions even when overall cell density does not change significantly, as the paper demonstrates for CD4+ T cells in lupus spleen (t-test p-value ≈ 0.16).","The same pipeline applied to COVID-19 lungs identifies endothelial-cell topology as the descriptor that best separates alveolitis from later disease stages, and this separation is not simply density thresholding."],"supporting_citations":[{"why":"Supplies the stability theorem for persistence diagrams, the guarantee that small data perturbations give small diagram perturbations.","marker":"[6]"},{"why":"Extends stability to the algebraic stability of persistence barcodes, underpinning the reliability of PH summaries.","marker":"[7]"},{"why":"Introduces multidimensional persistence, the framework whose main advantage PWDS claims to replicate at one-parameter cost.","marker":"[17]"},{"why":"Provides the lupus murine spleen CODEX dataset, the cell-type annotations, and the red/white pulp compartment classification used in the application.","marker":"[20]"},{"why":"Provides the COVID-19 human lung IMC dataset with disease-stage annotations and the prior local analysis that motivates the topological study.","marker":"[21]"},{"why":"Defines the alpha shape construction used to build filtered simplicial complexes from cell centroids.","marker":"[22]"},{"why":"Gives the union-of-balls dual-shape foundation for the alpha filtration.","marker":"[23]"},{"why":"Supplies the weak characterisation of Delaunay triangulation that underlies the witness complex construction.","marker":"[24]"},{"why":"Introduces witness complexes, used for the larger multi-cell-type point clouds.","marker":"[25]"},{"why":"Introduces persistence images, the stable vectorisation whose weightings the paper tunes.","marker":"[27]"}],"fun_headline_variants":["Death simplex paints tissue disease patterns","From persistence diagrams to spatial disease maps","Topology locates disease in lung and spleen","Red and blue death simplices separate healthy and sick","Spatial topology decodes tissue disease states"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the death simplex always lies inside the topological feature it kills, so the coloured triangle on the tissue image really marks where the loop or void is; the paper asserts this but does not prove it, and the death simplex itself is unstable under small data perturbations.","fun_headline_variants_meta":{"raw":{"variants":["Death simplex paints tissue disease patterns","From persistence diagrams to spatial disease maps","Topology locates disease in lung and spleen","Red and blue death simplices separate healthy and sick","Spatial topology decodes tissue disease states"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000226,"raw_usage":{"total_tokens":1470,"prompt_tokens":952,"completion_tokens":518,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":451}},"tokens_in":568,"tokens_out":518,"duration_ms":6994,"temperature":1.0,"reasoning_tokens":451,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:42:56.157180+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a synthetic loop with several nearly empty triangles of similar circumradius inside it and perturb the points slightly so the death simplex jumps between triangles; if the chosen triangle can land far from the feature's interior, the localisation claim fails. A direct check is to compute exact feature boundaries on synthetic loops and measure the Hausdorff distance between those boundaries and the PWDS triangles; large distances would falsify the claim that the triangles indicate feature positions.","supporting_citations":[{"cited_title":"‘Persistence Images: A Stable Vector Representation of Persistent Homology’","cited_arxiv_id":null,"evidence_quote":"Introduces persistence images, the stable vectorisation whose weightings the paper tunes."},{"cited_title":"‘Single cell spatial analysis reveals inflammatory foci of immature neutrophil and CD8 T cells in COVID-19 lungs’","cited_arxiv_id":null,"evidence_quote":"Provides the COVID-19 human lung IMC dataset with disease-stage annotations and the prior local analysis that motivates the topological study."},{"cited_title":"‘A weak characterisation of the Delaunay triangulation’","cited_arxiv_id":null,"evidence_quote":"Supplies the weak characterisation of Delaunay triangulation that underlies the witness complex construction."},{"cited_title":"‘Topological estimation using witness complexes’","cited_arxiv_id":null,"evidence_quote":"Introduces witness complexes, used for the larger multi-cell-type point clouds."}],"review_version":1}