Pith. sign in

REVIEW 5 major objections 6 minor 41 references

This paper claims that a hybrid of a GENERIC-based structure-preserving network and an MLP correction network predicts cell migration trajectories from video alone, with velocity accuracies above 85% and enforced thermodynamic consistency.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Augmenting a thermodynamically structured neural network with an environmental MLP predicts cell trajectories and mitotic events from video features at claimed accuracies above 85%.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection The paper's thermodynamic consistency claim collapses under a simple algebraic check: with their fixed L and M, exact degeneracy zeroes out the SPNN update, so the motion and the thermodynamics cannot both be real. the 5 major comments →

arxiv 2509.05388 v1 pith:RAG5JDRI submitted 2025-09-05 cs.CV cs.AI

Augmented Structure Preserving Neural Networks for cell biomechanics

Juan Olalla-Pombo , Alberto Bad\'ias , Miguel \'Angel Sanz-G\'omez , Jos\'e Mar\'ia Ben\'itez , Francisco Javier Mont\'ans This is my paper
classification cs.CV cs.AI
keywords cell migrationstructure-preserving neural networksGENERICtrajectory predictionmitosis predictioncomputer visioncell trackingmachine learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that cell migration can be predicted from video alone by splitting the problem into a thermodynamically constrained mechanical core and a flexible neural correction for environmental cues. The mechanical core is built on the GENERIC framework, which separates reversible energy-conserving dynamics from irreversible entropy-increasing dynamics, and the correction network absorbs factors like cell density gradients and neighbor motion. Trained on two synthetic datasets and one real in vitro dataset of MDCK cells, the combined model reports velocity accuracies above 85% in all coordinates and produces complete 105-frame roll-out trajectories with manageable accumulated error. The same visual features feed a second network that predicts mitotic events with 69% precision on a small real test set. If these results hold, they would make non-invasive, in vivo cell-behavior forecasting practical and provide a way to attribute movement to conserved versus environmental causes.

Core claim

The central claim is that augmenting a structure-preserving neural network with an ordinary multilayer perceptron yields a model that predicts cell trajectories with high accuracy while explicitly enforcing thermodynamic consistency. The SPNN submodel learns two matrices A and B that stand in for the discrete gradients of energy and entropy, updating state as z_{n+1} = z_n + Δt(L A + M B) z_n, with L skew-symmetric and M symmetric positive semi-definite so that energy is conserved and entropy increases by construction. The CoNN submodel takes 23 environmental features (density gradients, neighbor counts, brightness, area change, eccentricity) and produces a velocity correction; a final layer

What carries the argument

The machinery is the GENERIC update with learned gradient matrices. In the discrete form z_{n+1} - z_n = Δt(L A + M B) z_n, the matrices A and B replace the gradients of total energy and entropy, and the fixed matrices L (skew-symmetric) and M (symmetric positive semi-definite) enforce the degeneracy conditions L·Bz=0 and M·Az=0. Training minimizes a combined loss of prediction error plus degeneracy violations. The SPNN learns A and B from position and velocity only, while a separate MLP maps environmental features to a velocity correction; the final velocity is a learned fusion of both outputs and feeds the roll-out for the next frame.

Load-bearing premise

The load-bearing premise is that the gradients of the unknown energy and entropy of a cell system are well approximated by linear functions of the observable state (A z and B z), so the mechanical core reduces to a linear time-invariant update.

What would settle it

Take the real in vitro dataset, train the model on the first half, then roll out predictions over the second half and compare against a baseline MLP that has no GENERIC core but the same environmental features and training budget. If the baseline matches or beats the ASPNN on velocity error over the full horizon, the thermodynamic structure adds no predictive value. Additionally, check whether total energy, as computed from the learned A matrix, drifts systematically on a longer 500-frame video; a systematic drift would indicate the degeneracy conditions are not holding outside the training wi

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If accurate in longer real videos, the method offers a way to forecast collective cell migration without invasive measurements, using only video frames as input.
  • The split between a physics-constrained core and an environment-correction net gives an interpretable decomposition: the SPNN contribution can be read as the conserved-gradient part of motion, and the MLP contribution as the environmental response.
  • The learned energy and entropy matrices could be used to detect the presence and direction of external energy sources, such as chemical or density gradients, from trajectory data alone.
  • Mitosis prediction at 69% precision, even on a small dataset, suggests that division events are learnable from the same visual dynamics, enabling forecasting of proliferation hotspots.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper's core assumption is that the true energy and entropy gradients of a cell system can be linearly approximated as A z and B z; if that fails on a new tissue or longer timescales, the thermodynamic guarantees become vacuous even though the loss still decreases.
  • The reported energy conservation and entropy increase are enforced by the chosen matrix structure, not measured from biology; the learned potentials have no direct biophysical referent, so a natural extension would be to compare them with known free-energy models of cell mechanics.
  • The roll-out success over 105 frames on a single video leaves open how the model behaves with variable frame rates, cell density changes, or three-dimensional environments; testing on multi-day videos would stress the accumulated-error claim.
  • A concrete next test would be to withhold all environmental features during inference and compare trajectory accuracy against the full model; the gap would quantify how much predictive power actually comes from the CoNN branch versus the GENERIC core.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper presents an Augmented Structure Preserving Neural Network (ASPNN) that combines a GENERIC-based Structure Preserving Neural Network (SPNN) with a multilayer perceptron (CoNN) to predict cell velocities and full roll-out trajectories from video-derived features (positions, velocities, density gradients, surrounding-cell statistics). A separate MLP is introduced to predict mitotic events. The model is evaluated on two synthetic cell-migration datasets and one real in vitro dataset (MDCK cells), with reported velocity accuracies above 85% in all experiments, claims of thermodynamic consistency (energy conservation and entropy growth), and 69% accuracy on mitosis prediction. The mitosis model and the feature-extraction/tracking pipeline (SAM, DeepOCSort) are described in detail.

Significance. If the central claims held, the paper would present a useful combination of structure-preserving dynamical models with learned environmental corrections for cell biomechanics, and a video-only pipeline applicable to in vivo experiments. The explicit goal of enforcing GENERIC degeneracy while learning energy/entropy gradients is conceptually attractive and the mitosis-prediction extension is a valuable direction. However, the thermodynamic mechanism as written contains a basic algebraic incompatibility, the energy and entropy whose evolution is plotted are never defined, and the primary accuracy metric is not robust. These issues are central to the paper's main contribution, so the significance is conditional and the manuscript needs substantial revision before the claims can be accepted.

major comments (5)
  1. The degeneracy conditions and the chosen matrices are mutually incompatible with any nontrivial SPNN update. L in Eq. (13) is invertible, so L·Bz = 0 implies Bz = 0. M in Eq. (14) is positive definite (block-diagonal with eigenvalues 0.5 and 1.5), so M·Az = 0 implies Az = 0. Then the update in Eq. (7), z_{n+1} - z_n = Δt(L·Az + M·Bz), is identically zero whenever the degeneracy losses (11) are exactly satisfied. Thus the GENERIC submodel cannot simultaneously be thermodynamically consistent in the stated sense and produce nonzero cell velocities. The soft-penalty approach in Eq. (12) only avoids this by leaving the constraints unsatisfied; the reported degeneracy-loss curves (Figures 3, 7, 11) show finite values, not machine zero. The claims of 'compliance with thermodynamic principles' in Figures 5, 9, and 14 are therefore not supported by the architecture.
  2. The paper replaces the discrete gradients of energy and entropy by A z and B z without any justification, and never defines the energy E or entropy S whose gradients are being approximated. As a result, the quantities labelled 'Energy' and 'Entropy' in Figures 5, 9, and 14 are uninterpretable. There is no way to verify that they correspond to a physical energy or entropy, so the thermodynamic-consistency claim is circular: it checks properties of the learned A and B matrices, not compliance with any external thermodynamic benchmark. The manuscript must define E and S (or state explicitly that they are auxiliary learned potentials) and explain how the plotted curves are computed from the model.
  3. The velocity accuracy metric Acc = 100·(1 - |v_pred - v_gt| / |v_gt|) is not a stable evaluation measure. When the ground-truth velocity is near zero, as often happens at turning points or during pauses in cell migration, the metric becomes extremely sensitive or undefined, and a single small-velocity frame can dominate the reported average. The reported accuracies above 85% should be accompanied by robust alternatives (e.g., absolute/relative errors computed with a floor on |v_gt|, trajectory-wise displacement errors) and by error bars over train/test splits or multiple seeds. No standard deviations or confidence intervals are provided anywhere in the results.
  4. The mitosis prediction evaluation is not reported with standard classification statistics. The quantity called 'precision' is defined as correctly predicted positive events out of total positive events, which is recall, not precision. The evaluation uses only 30 trajectories and 12 mitotic events, with a hand-set 6-frame window and a probability threshold of 0.6; no confidence intervals are given. The 0.1% 'false positive' rate is computed over frames rather than events, which is not a standard event-level metric. The authors should report a confusion matrix at the event level, with precision/recall/F1 and confidence bounds, and clearly describe how non-mitotic trajectories are sampled.
  5. The evaluation does not include any comparison with a non-thermodynamically constrained baseline (e.g., a plain MLP or LSTM trained on the same features). Without such a baseline, the paper cannot demonstrate that the GENERIC structure adds predictive value or that the 'correction' MLP is behaving as intended. An ablation isolating the SPNN contribution would also help interpret Figures 6, 10, and 13, where the relative contributions are discussed qualitatively but never quantified in terms of prediction error.
minor comments (6)
  1. The notation 'DE/∂z' and 'DS/∂z' appears to be a typographical error for ∂E/∂z and ∂S/∂z. Please correct.
  2. The caption states 'The horizontal axis represents hundreds of epochs,' but the axis label reads 'Epoch' and the tick labels are 0, 10, 20, 30, 40, 50. Clarify the unit of the x-axis.
  3. The heading reads 'In-silico determistic experiment'; 'determistic' should be 'deterministic'.
  4. The normalization uses max(x) and min(x), presumably over the training set. State explicitly that these statistics are computed only on training data to avoid information leakage into the test set.
  5. No optimizer, batch size, or learning-rate decay details are given beyond the scheduler parameters. The architecture is reproducible only if these choices are specified.
  6. The relationship with prior SPNN work [33] should be clarified: which parts of the GENERIC submodel are adopted from [33] and which are new contributions of this paper? The current text only says the structures are 'based on' the cited work.

Circularity Check

2 steps flagged

Thermodynamic-compliance claim is defined through the fitted A/B matrices and is internally inconsistent with nontrivial SPNN updates; trajectory accuracy is independently benchmarked.

specific steps
  1. fitted input called prediction [Section 2.2.1, Eqs. (5)-(12); results Figures 5, 9, 14]
    "DE/∂z ≃ A z, DS/∂z ≃ B z ... By learning these two matrices, the model is learning the underlying physics of the problem (relations between observable variables and energy and entropy variations) while ensuring compliance with thermodynamical laws. ... ldeg = ||L·B_pred z_pred||^2 + ||M·A_pred z_pred||^2"

    Energy and entropy are never defined independently; their discrete gradients are set equal to the learned matrices A z and B z. The degeneracy loss is part of the training objective, so the reported constant-energy/increasing-entropy behavior is a check of the model's own soft constraints on its own fitted matrices, not an independent physical prediction.

  2. other [Section 2.2.1, Eqs. (7)-(9), (13)-(14)]
    "L = [0 0 1 0; 0 0 0 1; -1 0 0 0; 0 -1 0 0] ... M = [1 -0.5 0 0; -0.5 1 0 0; 0 0 1 -0.5; 0 0 -0.5 1] ... ldeg = ||L·B_pred z_pred||^2 + ||M·A_pred z_pred||^2"

    Since L and M are nonsingular, exact satisfaction of the degeneracy conditions L·Bz = 0 and M·Az = 0 forces Bz = Az = 0 for every z. Substituted into Eq. (7), the SPNN update z_{n+1} - z_n = Δt(L·A + M·B)z becomes identically zero. Thus nontrivial trajectory prediction requires violating the thermodynamic constraints; the claimed simultaneous accurate trajectories and thermodynamic compliance are mutually inconsistent by construction, and any reported compliance is a residual of an unsatisfied soft loss.

full rationale

The trajectory-accuracy half of the paper is not circular: predicted roll-outs are compared against ground-truth positions/velocities from simulation and video, and the reported >85% accuracies are an external benchmark. The circularity is confined to the thermodynamic claim. First, the energy and entropy whose conservation/growth is reported are never defined; their discrete gradients are set equal to the learned matrices A and B (Eqs. 5-6), and the degeneracy loss (Eq. 11) is part of the training objective. The plotted 'energy' and 'entropy' are therefore internal quantities generated by the fitted model, so the compliance result is not an independent physical prediction. Second, with the fixed L and M of Eqs. (13)-(14) (both invertible), exact satisfaction of Eqs. (8)-(9) forces A z = B z = 0, making the SPNN update of Eq. (7) vanish. Hence nontrivial trajectory prediction and exact GENERIC degeneracy cannot both hold; the reported thermodynamic compliance is only an approximate soft-loss artifact, not a derived consistency. These issues reduce the thermodynamic claim by construction, but the trajectory prediction retains independent content, giving a partial-circularity score of 6.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 0 invented entities

The paper introduces no new physical entities. The thermodynamic potentials E and S are not defined explicitly; they are implicitly represented by the learned matrices A and B, but this is a modeling construct, not a new entity.

free parameters (7)
  • Off-diagonal entries of M matrix (-0.5) = -0.5
    Chosen by hand in Eq. 14 to model coupled viscous dissipation in position and velocity; no fitting or justification.
  • Loss weight lambda_d = not reported
    In Eq. 12, lambda_d balances data loss and degeneracy loss; its value is never stated for any experiment.
  • Mitosis recurrence windows = 2 frames (area), 3 frames (brightness)
    Set in Section 3.4 based on visual observation that brightness changes precede shape changes.
  • Mitosis evaluation window and threshold = +/-3 frames around event; probability >0.6
    Section 3.4; these choices inflate apparent performance by allowing detection after the event.
  • Trajectory length cutoff = 105 frames
    Section 2.1; only trajectories longer than 105 frames are used, potentially biasing toward longer-lived cells.
  • Gaussian noise amplitude in synthetic case = 10% of current velocity
    Section 3.2; simulation setting, not a fitted parameter but a chosen condition.
  • Surrounding cell radius and grid size = 75 pixels, 820-pixel square
    Section 2.1, Table 3; hand-chosen feature scales for density and neighborhood features.
axioms (6)
  • domain assumption Cell mechanical state evolution obeys the GENERIC structure dz/dt = L dE/dz + M dS/dz with degeneracy conditions (Eqs. 1-3).
    Invoked without justification as the governing equation for cell mechanics.
  • ad hoc to paper Discrete gradients of energy and entropy can be represented as A z and B z (Eqs. 5-6).
    This linear representation is introduced to make the gradients learnable; no derivation or physical justification is given.
  • ad hoc to paper L and M are fixed as in Eqs. 13-14, with L the standard symplectic matrix and M with hand-chosen -0.5 coupling terms.
    These operators determine the reversible and irreversible dynamics of the SPNN and are not learned.
  • ad hoc to paper The degeneracy conditions enforced as soft penalties in the loss function (Eq. 11) are sufficient to guarantee thermodynamic consistency.
    The conditions are not hard constraints, and the energy/entropy are defined via the same learned matrices that are being penalized.
  • domain assumption The environmental MLP and the SPNN contributions combine additively in velocity space.
    The final layer sums SPNN and MLP velocities; the paper does not test interaction terms.
  • ad hoc to paper Relative velocity error (Eq. 16) is a meaningful accuracy measure for trajectory prediction.
    This metric diverges when true velocity is near zero and is never validated against distance-based errors.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Augmented Structure Preserving Neural Networks for cell biomechanics." pith.science (2026). https://pith.science/paper/RAG5JDRI

@misc{pith2026250905388,
  author       = {Pith},
  title        = {Pith review of: Augmented Structure Preserving Neural Networks for cell biomechanics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RAG5JDRI}},
  note         = {Machine review of arXiv:2509.05388}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Cell biomechanics involve a great number of complex phenomena that are fundamental to the evolution of life itself and other associated processes, ranging from the very early stages of embryo-genesis to the maintenance of damaged structures or the growth of tumors. Given the importance of such phenomena, increasing research has been dedicated to their understanding, but the many interactions between them and their influence on the decisions of cells as a collective network or cluster remain unclear. We present a new approach that combines Structure Preserving Neural Networks, which study cell movements as a purely mechanical system, with other Machine Learning tools (Artificial Neural Networks), which allow taking into consideration environmental factors that can be directly deduced from an experiment with Computer Vision techniques. This new model, tested on simulated and real cell migration cases, predicts complete cell trajectories following a roll-out policy with a high level of accuracy. This work also includes a mitosis event prediction model based on Neural Networks architectures which makes use of the same observed features.

Figures

Figures reproduced from arXiv: 2509.05388 by Alberto Bad\'ias, Francisco Javier Mont\'ans, Jos\'e Mar\'ia Ben\'itez, Juan Olalla-Pombo, Miguel \'Angel Sanz-G\'omez.

Figure 1
Figure 1. Figure 1: Segmented image with masks represented as colored layers over cells in the [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Complete process pipeline. Note that position and velocities inputs to the SPNN [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Evolution of loss functions for data (left) and degeneracy conditions (right) for [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of two predicted and real trajectories for the first [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Evolution of thermodynamical variables for a complete trajectory. [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Contribution of each part of the model to the final output velocities for the [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Evolution of loss functions for data (left) and degeneracy conditions (right) for [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of two predicted and real trajectories for the second artificial case. [PITH_FULL_IMAGE:figures/full_fig_p019_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Evolution of thermodynamical variables for a complete trajectory from the [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Contribution of each part of the model to the final output velocities for the [PITH_FULL_IMAGE:figures/full_fig_p021_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Evolution of loss functions for data (left) and degeneracy conditions (right) for [PITH_FULL_IMAGE:figures/full_fig_p022_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Comparison of four predicted and true trajectories for the real experiment. [PITH_FULL_IMAGE:figures/full_fig_p023_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Contribution of each part of the model to the final output velocities for the [PITH_FULL_IMAGE:figures/full_fig_p024_13.png] view at source ↗
Figure 15
Figure 15. Figure 15: Evolution of loss function for mitosis prediction model. [PITH_FULL_IMAGE:figures/full_fig_p025_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Comparison of model outputs and ground truth data for two trajectories with [PITH_FULL_IMAGE:figures/full_fig_p026_16.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 37 canonical work pages · 1 internal anchor

  1. [1]

    Horwitz, D

    R. Horwitz, D. Webb, Cell migration, Current Biology 13 (19) (2003) R756–R759

    work page 2003

  2. [2]

    Kurosaka, A

    S. Kurosaka, A. Kashina, Cell biology of embryonic migration, Birth Defects Research Part C: Embryo Today: Reviews 84 (2) (2008) 102– 122

    work page 2008

  3. [3]

    Friedl, D

    P. Friedl, D. Gilmour, Collective cell migration in morphogenesis, regen- eration and cancer, Nature reviews Molecular cell biology 10 (7) (2009) 445–457

    work page 2009

  4. [4]

    Torres, M

    P. Torres, M. Castro, M. Reyes, V. Torres, Histatins, wound healing, and cell migration, Oral diseases 24 (7) (2018) 1150–1160

    work page 2018

  5. [5]

    W.J.Polacheck, I.K.Zervantonakis, R.D.Kamm, Tumorcellmigration in complex microenvironments, Cellular and Molecular Life Sciences 70 (2013) 1335–1356

    work page 2013

  6. [6]

    J. Wu, Z. Mao, H. Tan, L. Han, T. Ren, C. Gao, Gradient biomaterials and their influences on cell migration, Interface focus 2 (3) (2012) 337– 355

    work page 2012

  7. [7]

    Dormann, C

    D. Dormann, C. J. Weijer, Chemotactic cell movement during develop- ment, Current opinion in genetics & development 13 (4) (2003) 358–364

    work page 2003

  8. [8]

    P. A. Iglesias, P. N. Devreotes, Navigating through models of chemo- taxis, Current opinion in cell biology 20 (1) (2008) 35–40

    work page 2008

  9. [9]

    C.-L. Hu, F. S. Barnes, A theory of necrotaxis, Biophysical Journal 10 (10) (1970) 958–969

    work page 1970

  10. [10]

    Trepat, J

    X. Trepat, J. J. Fredberg, Plithotaxis and emergent dynamics in collec- tive cellular migration, Trends in cell biology 21 (11) (2011) 638–646

    work page 2011

  11. [11]

    Gjorevski, A

    N. Gjorevski, A. S. Piotrowski, V. D. Varner, C. M. Nelson, Dynamic tensile forces drive collective cell migration through three-dimensional extracellular matrices, Scientific reports 5 (1) (2015) 11458. 28

    work page 2015

  12. [12]

    M. R. Amieva, H. Furthmayr, Subcellular localization of moesin in dy- namic filopodia, retraction fibers, and other structures involved in sub- strate exploration, attachment, and cell-cell contacts, Experimental Cell Research 219 (1) (1995) 180–196

    work page 1995

  13. [13]

    Saucedo-Mora, M

    L. Saucedo-Mora, M. Á. Sanz, F. J. Montáns, J. M. Benítez, A sim- ple agent-based hybrid model to simulate the biophysics of glioblastoma multiforme cells and the concomitant evolution of the oxygen field, Com- puter Methods and Programs in Biomedicine 246 (2024) 108046

    work page 2024

  14. [14]

    E. M. Benetti, M. K. Gunnewiek, C. A. van Blitterswijk, G. J. Vancso, L. Moroni, Mimicking natural cell environments: design, fabrication and application of bio-chemical gradients on polymeric biomaterial sub- strates, Journal of Materials Chemistry B 4 (24) (2016) 4244–4257

    work page 2016

  15. [15]

    Malijauskaite, S

    S. Malijauskaite, S. Connolly, D. Newport, K. McGourty, Gradients in the in vivo intestinal stem cell compartment and their in vitro recapit- ulation in mimetic platforms, Cytokine & Growth Factor Reviews 60 (2021) 76–88

    work page 2021

  16. [16]

    T. M. Keenan, A. Folch, Biomolecular gradients in cell culture systems, Lab on a Chip 8 (1) (2008) 34–57

    work page 2008

  17. [17]

    S.Rafieyan, E.Vasheghani-Farahani, N.Baheiraei, H.Keshavarz, Mlate: Machine learning for predicting cell behavior on cardiac tissue engineer- ing scaffolds, Computers in Biology and Medicine 158 (2023) 106804

    work page 2023

  18. [18]

    F. M. Garcia-Moreno, J. Ruiz-Espigares, M. A. Gutiérrez-Naranjo, J. A. Marchal, Using deep learning for predicting the dynamic evolution of breast cancer migration, Computers in Biology and Medicine 180 (2024) 108890

    work page 2024

  19. [19]

    Colen, M

    J. Colen, M. Han, R. Zhang, S. A. Redford, L. M. Lemma, L. Morgan, P. V. Ruijgrok, R. Adkins, Z. Bryant, Z. Dogic, et al., Machine learning active-nematic hydrodynamics, Proceedings of the National Academy of Sciences 118 (10) (2021) e2016708118

    work page 2021

  20. [20]

    M. Mak, C. A. Reinhart-King, D. Erickson, Microfabricated physical spatial gradients for investigating cell migration and invasion dynamics, PLoS one 6 (6) (2011) e20825. 29

    work page 2011

  21. [21]

    Schlie, M

    S. Schlie, M. Gruene, H. Dittmar, B. N. Chichkov, Dynamics of cell at- tachment: adhesion time and force, Tissue Engineering Part C: Methods 18 (9) (2012) 688–696

    work page 2012

  22. [22]

    K. Webb, V. Hlady, P. A. Tresco, Relationships among cell attachment, spreading, cytoskeletal organization, and migration rate for anchorage- dependent cells on model surfaces, Journal of Biomedical Materials Re- search: An Official Journal of The Society for Biomaterials and The Japanese Society for Biomaterials 49 (3) (2000) 362–368

    work page 2000

  23. [23]

    Y. Mao, L. Han, Z. Yin, Cell mitosis event analysis in phase contrast microscopyimagesusingdeeplearning, Medicalimageanalysis57(2019) 32–43

    work page 2019

  24. [24]

    Delgado-Rodriguez, R

    P. Delgado-Rodriguez, R. M. Sánchez, E. Rouméas-Noël, F. Paris, A. Munoz-Barrutia, Automatic classification of normal and abnormal cell division using deep learning, Scientific Reports 14 (1) (2024) 14241

    work page 2024

  25. [25]

    Szabó, R

    A. Szabó, R. M. Merks, Cellular potts modeling of tumor growth, tumor invasion, and tumor evolution, Frontiers in oncology 3 (2013) 87

    work page 2013

  26. [26]

    Kirillov, E

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y. Lo, P. Dollár, R. Girshick, Segment anything, arXiv:2304.02643 (2023)

    Pith/arXiv arXiv 2023

  27. [27]

    Bradski, The OpenCV Library, Dr

    G. Bradski, The OpenCV Library, Dr. Dobb’s Journal of Software Tools (2000)

    2000

  28. [28]

    van der Walt, J

    S. van der Walt, J. L. Schönberger, J. Nunez-Iglesias, F. Boulogne, J. D. Warner, N. Yager, E. Gouillart, T. Yu, the scikit-image contrib- utors, scikit-image: Image processing in python, PeerJ 2 (2014) e453. doi:10.7717/peerj.453

    doi:10.7717/peerj.453 2014

  29. [29]

    Deep OC-SORT: Multi-Pedestrian Tracking by Adaptive Re-Identification

    G. Maggiolino, A. Ahmad, J. Cao, K. Kitani, Deep oc-sort: Multi- pedestrian tracking by adaptive re-identification, arXiv preprint arXiv:2302.11813 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  30. [30]

    J. Cao, X. Weng, R. Khirodkar, J. Pang, K. Kitani, Observation-centric sort: Rethinking sort for robust multi-object tracking, arXiv preprint arXiv:2203.14360 (2022). 30

    Pith/arXiv arXiv 2022

  31. [31]

    H. C. Öttinger, M. Grmela, Dynamics and thermodynamics of complex fluids. ii. illustrations of a general formalism, Physical Review E 56 (6) (1997) 6633

    work page 1997

  32. [32]

    I. Romero, Thermodynamically consistent time-stepping algorithms for non-linear thermomechanical systems, International journal for numer- ical methods in engineering 79 (6) (2009) 706–732

    work page 2009

  33. [33]

    Hernández, A

    Q. Hernández, A. Badías, D. González, F. Chinesta, E. Cueto, Structure-preserving neural networks, Journal of Computational Physics 426 (2021) 109950

    work page 2021

  34. [34]

    Falkner, A

    S. Falkner, A. Klein, F. Hutter, Bohb: Robust and efficient hyperpa- rameter optimization at scale, in: International conference on machine learning, PMLR, 2018, pp. 1437–1446

    work page 2018

  35. [35]

    Thery, M

    M. Thery, M. Bornens, Cell shape and cell division, Current opinion in cell biology 18 (6) (2006) 648–657

    work page 2006

  36. [36]

    Ito, Role of histone modification in chromatin dynamics, The journal of biochemistry 141 (5) (2007) 609–614

    T. Ito, Role of histone modification in chromatin dynamics, The journal of biochemistry 141 (5) (2007) 609–614

    work page 2007

  37. [37]

    A. S. Belmont, S. Dietzel, A. C. Nye, Y. G. Strukov, T. Tumbar, Large- scale chromatin structure and function, Current opinion in cell biology 11 (3) (1999) 307–311

    work page 1999

  38. [38]

    Giancotti, S

    V. Giancotti, S. Brovedan, M. Fonda, R. Grandinetti, E. Russo, Intrinsic and extrinsic fluorescence of histones h2a and h2b: a conformational study, Biophysical Chemistry 7 (4) (1978) 269–277

    work page 1978

  39. [39]

    U. Ruby, V. Yendapalli, Binary cross entropy with deep learning tech- nique for image classification, Int. J. Adv. Trends Comput. Sci. Eng 9 (10) (2020)

    work page 2020

  40. [40]

    Marel, M

    A.-K. Marel, M. Zorn, C. Klingner, R. Wedlich-Söldner, E. Frey, J. O. Rädler, Flow and diffusion in channel-guided cell migration, Biophysical journal 107 (5) (2014) 1054–1064

    work page 2014

  41. [41]

    A. B. Georgi, P. T. Stukenberg, M. W. Kirschner, Timing of events in mitosis, Current biology 12 (2) (2002) 105–114. 31

    work page 2002

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.