Pith. sign in

REVIEW 4 major objections 7 minor 33 references

STAGED: A Multi-Agent Neural Network for Learning Cellular Interaction Dynamics

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read STAGED claims that a multi-agent graph neural ODE with attention can jointly infer intracellular gene regulatory networks and intercellular ligand-receptor communication from spatial transcriptomics data, and that the learned time-varying…

desk verdict A real multi-agent extension of RITINI, but the interaction-inference claim is not demonstrated without a prior-ablation and quantitative edge-recovery metrics. read the letter →

arxiv 2507.11660 v1 pith:7NMAMFKI submitted 2025-07-15 cs.LG cs.MAq-bio.QM

Joao F. Rocha , Ke Xu , Xingzhi Sun , Ananya Krishna , Dhananjay Bhaskar , Blanche Mongeon , Morgan Craig , Mark Gerstein
show 1 more author
Smita Krishnaswamy
This is my paper · ORCID
classification cs.LGcs.MAq-bio.QM
keywords spatialtranscriptomicsgeneregulatorynetworkinferencegraphneuralODEattentionmechanismmulti-agentmodelingcell-cellcommunicationligand-receptorsignalingAlzheimer'sdiseasemicroglia
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

STAGED is a multi-agent neural network that treats each cell as an agent governed by a graph neural ODE: genes are nodes, regulatory influences are edges, and agent-to-agent signaling passes through ligand-receptor pairs. The paper argues that, by training only to predict future gene expression from spatial transcriptomics snapshots, the model's time-varying attention weights recover both the intracellular gene regulatory network and intercellular communication, and that these recovered interactions are biologically meaningful. If that claim holds, spatial transcriptomics becomes a source of mechanistic, testable hypotheses about which genes regulate which, in which cell types, in which spatial neighborhoods, without requiring perturbation experiments up front. The paper supports this with two simulations, one of a tumor microenvironment and one of coupled oscillatory gene circuits, and an application to microglia in an Alzheimer's disease atlas.

What carries the argument

The load-bearing object is the multi-agent graph neural ODE (GDE) with two-level masked attention. Each cell hosts a GDE whose nodes are genes and whose directed edges are regulatory influences; the derivative of each gene is computed from a per-cell-type shared attention embedding plus ligand-receptor-mediated messages from genes in nearby cells, with explicit time lags for gene-gene, gene-ligand, ligand-receptor, and receptor-gene delays. A graph attention layer on the tissue graph, masked to ligand-receptor pairs, decides which neighboring cells' signals matter, while a history-aware attention inside each cell lets regulatory edges strengthen and decay over time. This mechanism is what makes the model able, in principle, to convert spatial expression snapshots into time-varying interaction graphs: the attention weights are the output asked to carry biological meaning, and the same architecture is what the paper uses to make claims about microglia regulatory states.

What would settle it

Run a perturbation experiment in the same cell types as the microglia analysis: knock down one gene that STAGED flags as a high-attention regulator and measure whether its predicted targets shift as the attention weights imply; if they do not, or if attention-high edges show no enrichment over random edges against a genome-wide regulatory assay such as chromatin-accessibility or transcription-factor binding measurements in those cell types, the inference claim collapses. For the synthetic setting, a failure test would be generating a larger ground-truth network with more genes, more cells, and higher noise, and checking whether recovered attention edges match the simulated activators and repressors with correct direction and timing.

Watch

Extended reading notes

Core claim

STAGED's central claim is that a single model can jointly learn both levels of cellular interaction from spatial transcriptomics. Each agent is a cell; inside each cell, a graph attention layer refines gene-gene edge weights over time, and between cells, ligand-receptor-masked attention passes signals from neighboring cells into the derivative predicted by a shared cell-type-specific neural ODE. Because the network is trained end-to-end to minimize prediction error of expression trajectories, the learned attention weights serve as inferred interaction strengths. The paper shows that in a six-gene, seven-cell, two-type synthetic oscillatory system, the attention-inferred gene-gene profiles recapitulate the cyclical activation pattern of the ground-truth network, and that on microglia from a human Alzheimer's atlas, attention maps produce distinct, spatially localized regulatory programs per subpopulation. The paper concludes that STAGED provides interpretable, context-dependent regulatory graphs and positions it as a hypothesis-generation and in silico perturbation tool.

Load-bearing premise

The entire biological contribution rests on the assumption that attention weights trained only to minimize gene-expression prediction error reflect true causal regulatory and signaling connections, validated only qualitatively on one small synthetic oscillatory network and then applied to human microglia data without perturbation or external validation.

Editorial extensions

If this is right

  • From gene expression and spatial positions alone, STAGED produces time-resolved gene-gene interaction graphs per cell type, including edges that appear and disappear as dynamics unfold.
  • The model captures cell-cell communication through ligand-receptor pairs, so downstream analysis can read which signaling axes are active in a spatial neighborhood at a given time.
  • Because cell types share parameters but cells have different neighborhoods, the same cell type can exhibit different trajectories and different inferred regulatory programs in different spatial contexts.
  • The training objective is purely predictive, yet the learned attention weights are presented as evidence of regulatory structure; if this holds, STAGED offers a hypothesis-generation tool for spatially resolved disease data.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the attention-weight equivalence holds, then a direct test would be whether attention-high edges predict the outcome of a perturbation, such as a gene knockdown or receptor block, better than random edges.
  • The same architecture could be applied to other spatially resolved readouts, such as protein-level imaging, where ligand-receptor biology is more directly observed and could provide an external check on the inferred intercellular edges.
  • The oscillatory validation uses seven cells and six genes; scaling the ground-truth simulation to larger populations and more genes while measuring recovery of edge direction, sign, and timing under increasing noise would map the reliability boundary of the inference.
  • The microglia result suggests testable biology: if the inferred attention-distinct subpopulations are real, they should respond differently to the same ligand stimulus.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. STAGED is a multi-agent graph neural ODE framework that models each cell as an agent with an internal gene-regulatory graph, extended by intercellular ligand-receptor signaling edges. The model is trained to predict future gene expression from initial expression and spatial coordinates, and the learned attention weights are interpreted as time-varying intra- and inter-cellular interaction strengths. The paper evaluates STAGED on a glioblastoma agent-based simulation, a small synthetic oscillatory GRN (7 cells, 6 genes), and on SEA-AD microglia spatial transcriptomics data, reporting trajectory prediction fits and qualitative attention-derived networks.

Significance. The architecture is original and timely: combining agent-based modeling with graph neural ODEs and attention for spatial transcriptomics is a plausible route toward data-driven inference of multicellular regulatory dynamics. If the attention weights were rigorously shown to recover known regulatory and signaling edges, the tool would be valuable for hypothesis generation. However, the paper's central claim—that STAGED 'infers' intra- and inter-cellular interactions—is currently supported only by qualitative comparisons on a single 6-gene synthetic network and by unvalidated patterns on real data. The manuscript ships a clear algorithmic description (Algorithm 1) and a reproducible simulation environment, but the load-bearing validation is missing. The significance of the contribution is therefore conditional on additional quantitative and ablation-based evidence.

major comments (4)
  1. [Section IV-B and Algorithm 1] The central claim that STAGED recovers gene regulatory interactions is not established because Algorithm 1 seeds each cell's graph with a prior GRN Gp = {Gk} and then adds ligand/receptor nodes and edges. In the Section IV-B oscillatory experiment, the prior is plausibly the same simulated GRN used to generate the data, so the qualitative agreement in Figure 6B could be inherited from the initial graph rather than learned by the attention mechanism. No ablation removes or randomizes this prior, and no quantitative edge-recovery metric (e.g., AUROC or AUPRC against the known ground-truth edges) is reported. Without such an experiment, the claim of 'inferring' regulatory networks is unverified; please add an ablation comparing attention-derived edges to the prior graph and to the ground truth, and report a standard edge-recovery metric.
  2. [Section IV-B (Figure 5, Figure 6)] The synthetic oscillatory validation is extremely small (7 cells, 6 genes per cell type) and lacks any baseline comparisons. Trajectory prediction accuracy (Figure 5) only shows that the model can fit the training distribution; it does not demonstrate that the attention weights carry information beyond the input structure. Please include comparisons to a non-attention GDE, to RITINI applied cell-wise, and to a simple vector autoregression, and report prediction error on held-out time points or cells. Also report the quantitative edge-recovery performance for both intra- and inter-cellular edges.
  3. [Section IV-C (Figure 7)] The SEA-AD microglia application is used to support biological conclusions ('distinct regulatory programs', 'key genes interact'), but the inferred networks are only inspected qualitatively with no external validation. Because no ground-truth regulatory network exists for these data, the paper should at least provide sanity checks: comparison of inferred edges against known ligand-receptor pairs or published microglia GRNs, stability of attention weights across random initializations, or validation of predicted trajectories on held-out spatial regions or time points. Without such checks, the biological claims go beyond what the evidence supports.
  4. [Abstract and Contributions] The abstract and contribution list state that STAGED 'captures both intercellular and intracellular interactions' and 'infers' them, but the paper does not quantitatively evaluate the intercellular component separately from the intracellular component. For instance, in the Section IV-B experiment, no metric distinguishes how well ligand-receptor edges are recovered versus gene-gene edges. Please disaggregate the evaluation so that the load-bearing claim can be tested independently for each interaction type.
minor comments (7)
  1. [Section III.B, Equation for a_theta,c] The notation in the definition of the attention coefficient β(cs)ij is unclear: the indices i and j are used both to sum over genes and to label the attention coefficient, and the expression 'Σ_{i,j∈G} β(cs)ij x_{c,i,t} x_{s,j,t−δ}' is ambiguous about whether this is a weighted sum over all ordered pairs or a matrix product. Please define the summation range precisely and clarify the tensor dimensions.
  2. [Section III.C] The masked attention constraint (β(cs)ij ← β(cs)ij 1(j∈L and i∈R)) is described for intercellular ligand-receptor edges, but it is not stated whether the same mask is applied to the intracellular gene-gene attention. This distinction is important because the model's interpretability claim depends on what the attention weights are allowed to represent.
  3. [Section IV.A] The glioblastoma simulation validation uses simplified agent states (proliferation, death, total counts) rather than gene expression. This is fine as a sanity check of the multi-agent ODE framework, but the text should explicitly note that this experiment does not exercise the GRN-inference capability.
  4. [Section IV.B, Equation (1)-(2)] The notation Rg(x) is used for the regulatory input function, but later in the same section 'Rg' denotes the set of repressors of gene g. Please rename one of them to avoid confusion. Also, the parameter δg appears both as a decay rate and as part of the time-lag notation (δgg, δgl, δlr, δrg); please introduce distinct symbols.
  5. [Figure 6] Figure 6B shows inferred interactions only for gene 1. For a 6-gene network, please either show all genes or report a summary statistic (e.g., mean AUROC across genes). As presented, the figure may overstate the qualitative match.
  6. [General presentation] There are several typos and style inconsistencies: 'oncolitic' should be 'oncolytic', 'Spatio- Temporal' has an extra space, 'Glioblastoma Tumour Microenvironment' mixes capitalization, and Figure 7 caption refers to 'cell types' when the text says 'subpopulations'. Please proofread throughout.
  7. [Reproducibility] The paper does not mention code availability, hyperparameter settings, or training details (learning rate, number of epochs, ODE solver tolerances, attention heads). Given the emphasis on a new simulation environment, please include a reproducibility statement and specify all training hyperparameters.

Circularity Check

2 steps flagged · score 5.0 of 10

The interaction-inference claim is partially circular: the GAT operates on the ground-truth GRN as an input prior in the synthetic validation, and the real-data pipeline trains on same-lab MIOFlow-inferred trajectories.

  1. fitted input called prediction [Algorithm 1 (Appendix) Steps 1 and 3; Section IV-B, Figures 5-6]
    "Algorithm 1 Inputs: Prior gene regulatory networks for each cell type: Gp = {Gk}... Assign cell-type-specific GRN Gc ← GK(c)... Apply GAT layer on G̃c to get attention weights and embeddings... Figure 6B: The inferred interactions recapitulate the simulated cyclical activation pattern, demonstrating STAGED's ability to capture temporally varying regulatory relationships."

    The GAT computes attention only over edges already present in G̃c, which is initialized from the prior Gp. In the Section IV-B benchmark the prior is the same simulated GRN used to generate the oscillatory expression data, so the ground-truth edges are available to the model as an input. The Figure 6B 'recapitulation' therefore compares attention weights over ground-truth edges to those same edges; no ablation randomizes or removes Gp and no quantitative edge-recovery metric is reported, so the recovery claim is not independent of the model input.

  2. self citation load bearing [Section II (Related Work) and Section IV-C]
    "RITINI pioneered the idea of using attention-based GDEs to learn gene regulatory interactions, using MIOFlow derived gene expression trajectories as training data. STAGED builds upon this, incorporating GDEs within an agent-based framework to simultaneously infer regulatory interactions and spatiotemporal cellular communication."

    The real-data component trains STAGED on continuous trajectories inferred by MIOFlow, a same-laboratory method cited as the source of training targets, and the microglial regulatory programs in Figure 7 are read off attention weights of a model fitted to that same-lab output. The SEA-AD biological-discovery claim is therefore supported by a self-citation chain (MIOFlow/RITINI) rather than by observed or externally validated trajectories, and the paper reports no independent edge-level validation for the inferred microglial networks.

full rationale

The core forecasting equations are not circular: STAGED minimizes an MSE loss between predicted and target expression, and the attention weights are an auxiliary output rather than the training objective. However, the interaction-discovery claim is partially circular. Algorithm 1 initializes each cell graph with a prior GRN Gp and the GAT can only assign attention to edges that already exist in G̃c. In the synthetic oscillatory experiment, that prior is the very GRN used to simulate the data, so Figure 6's 'recapitulation' compares the model output to its own input, with no prior-ablation or edge-recovery metric. The real-data section additionally depends on MIOFlow-inferred trajectories from the same laboratory as training targets, making the SEA-AD regulatory-network inferences a self-referential pipeline. The expression-forecasting result in Figure 5 is a legitimate independent fit, which prevents a higher score; the central inference claim, as validated, is contaminated rather than cleanly derived.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The model introduces no new physical entities. Its learning relies on several domain assumptions: deterministic ODE dynamics, local ligand-receptor-mediated communication, per-cell-type weight sharing, and the interpretability of attention as interaction strength. The last assumption is the least supported, as it is validated only on a small synthetic system. Free parameters include hand-set lags, the neighborhood radius, and unreported architecture choices.

free parameters (4)
  • Signaling time lag δ = not specified
    Scalar lag used in the update equation for intercellular signaling (Section III-B). Chosen by hand, not learned; determines how delayed neighbor signals are.
  • Neighborhood radius r = not specified
    Defines spatial neighbors N(c|y_t) as cells within distance r (Section III-B). This hyperparameter controls the tissue-level graph connectivity.
  • Time lags δgg, δgl, δlr, δrg = not specified
    Four separate lags in Algorithm 1 for gene-gene, gene-ligand, ligand-receptor, and receptor-gene delays. Set by the user, affecting the model's ability to capture delayed regulation.
  • Network and training hyperparameters = not specified
    Number of GAT layers, hidden dimensions, ODE solver settings, learning rate, optimizer, and loss weights are not reported. These are free choices that influence the results.
assumptions (5)
  • domain assumption Gene expression evolves deterministically according to an ordinary differential equation
    The model uses a neural ODE to integrate gene expression dynamics (Section III-B, Algorithm 1). Real biological systems are stochastic and noisy, but the paper does not model stochasticity.
  • domain assumption Only cells within a fixed spatial radius r communicate, and only via known ligand-receptor pairs
    The tissue-level graph is constructed from proximity and masked attention to ligand-receptor edges (Section III-C). This excludes long-range or non-canonical signaling mechanisms.
  • domain assumption Cells of the same type share all regulatory parameters
    Weight-sharing per cell type is a central design choice (Section III-C). It assumes that all cells of a type have identical regulatory logic, which may not hold for heterogeneous states.
  • domain assumption Trajectories inferred from static spatial transcriptomics snapshots represent true biological dynamics
    For the SEA-AD application, STAGED is trained on expression trajectories that are not time-series but inferred (likely using MIOFlow, Section II). This treats pseudotime or inferred paths as ground truth.
  • ad hoc to paper Learned attention weights correspond directly to biological interaction strengths
    The paper interprets attention coefficients as regulatory and signaling strengths (Section IV-B, Figure 6). No identifiability analysis or out-of-sample test supports this mapping.

how reviews work

0 comments
Cite this review

Pith. "Pith review of STAGED: A Multi-Agent Neural Network for Learning Cellular Interaction Dynamics." pith.science (2026). https://pith.science/paper/7NMAMFKI

@misc{pith2026250711660,
  author       = {Pith},
  title        = {Pith review of: STAGED: A Multi-Agent Neural Network for Learning Cellular Interaction Dynamics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7NMAMFKI}},
  note         = {Machine review of arXiv:2507.11660}
}
read the original abstract

The advent of single-cell technology has significantly improved our understanding of cellular states and subpopulations in various tissues under normal and diseased conditions by employing data-driven approaches such as clustering and trajectory inference. However, these methods consider cells as independent data points of population distributions. With spatial transcriptomics, we can represent cellular organization, along with dynamic cell-cell interactions that lead to changes in cell state. Still, key computational advances are necessary to enable the data-driven learning of such complex interactive cellular dynamics. While agent-based modeling (ABM) provides a powerful framework, traditional approaches rely on handcrafted rules derived from domain knowledge rather than data-driven approaches. To address this, we introduce Spatio Temporal Agent-Based Graph Evolution Dynamics(STAGED) integrating ABM with deep learning to model intercellular communication, and its effect on the intracellular gene regulatory network. Using graph ODE networks (GDEs) with shared weights per cell type, our approach represents genes as vertices and interactions as directed edges, dynamically learning their strengths through a designed attention mechanism. Trained to match continuous trajectories of simulated as well as inferred trajectories from spatial transcriptomics data, the model captures both intercellular and intracellular interactions, enabling a more adaptive and accurate representation of cellular dynamics.

Figures

Figures reproduced from arXiv: 2507.11660 by the authors.

Figure 1
Figure 1. STAGED infers regulatory and cell–cell interaction networks using spatial transcriptomics data. A. Spatial transcriptomics data from Xenium [26]. Cells in regions A and B express similar genes but occupy distinct spatial neighborhoods. Traditional single-cell methods, which lack spatial context, treat such cells as equivalent, ignoring important differences in their local signaling landscapes. B. STAGED models each … view at source ↗
Figure 2
Figure 2. STAGED model: Each agent in STAGED is a cell-type–specific model composed of a graph attention network (GAT) and a neural ODE. The GAT aggregates information from spatially proximal cells via lig￾and–receptor–mediated interactions, producing a context-aware embedding of each gene. This embedding is passed through a multi-layer perceptron (MLP) to predict the gene’s time derivative, which the ODE solver then integrat… view at source ↗
Figure 3
Figure 3. Agent-based simulation of glioblastoma A. Agent-based simulation model for a glioblastoma in an oncolitic virus environment from [18]. B. Graph is used to maintain information on the overall state of the agent. C. STAGED was able to fully recover population dynamics. B. STAGED infers regulatory networks that exhibit complex oscillatory dynamics In the second setting, we tested STAGED’s ability to infer complex gene … view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Spatially resolved stochastic simulation of gene expression dynamics in a randomly constructed gene regulatory network (GRN) using GillesPy2. A. Ground truth GRN where nodes represent genes, directed red edges indicate repression and directed green edges indicate activ…
Figure 5
Figure 5. Figure 5: STAGED accurately predicts complex gene expression dynamics. We simulated a population of 7 cells of 2 different cell types, each consisting of a small regulatory network of 6 genes that exhibited oscillatory dynamics. Cells were coupled by spatial proximity and Gaussi…
Figure 6
Figure 6. Figure 6: Temporally varying regulatory interactions inferred using the attention mechanism in STAGED. A. Simulated gene regulatory network (GRN) exhibiting oscillatory dynamics, where genes activate sequentially in a circular pattern. B. Time-resolved gene-gene interaction prof…
Figure 7
Figure 7. Figure 7: STAGED identifies cell-type-specific gene dynamics and regulatory networks in microglia subpopula￾tions. A. Top expressed genes across three distinct microglia subpopulations. STAGED recovers the gene dynamics and focus in different genes for each of the cell types. B.…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 26 canonical work pages

  1. [1]

    Fast unfolding of communities in large networks

    Vincent D Blondel, Jean-Loup Guillaume, Renaud Lambiotte, and Etienne Lefebvre. Fast unfolding of communities in large networks. Journal of Statistical Mechanics: Theory and Experiment , 2008(10):P10008, October 2008

    2008

  2. [2]

    A smart local moving algorithm for large-scale modularity-based community detection

    Ludo Waltman and Nees Jan van Eck. A smart local moving algorithm for large-scale modularity-based community detection. The European Physical Journal B , 86(11), November 2013

    work page 2013

  3. [3]

    Visualizing data using t-sne

    Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9(86):2579–2605, 2008

    2008

  4. [4]

    Umap: Uniform manifold approximation and projection

    Leland McInnes, John Healy, Nathaniel Saul, and Lukas Großberger. Umap: Uniform manifold approximation and projection. Journal of Open Source Software , 3(29):861, September 2018

    work page 2018

  5. [5]

    Moon, David van Dijk, Zheng Wang, Scott Gigante, Daniel B

    Kevin R. Moon, David van Dijk, Zheng Wang, Scott Gigante, Daniel B. Burkhardt, William S. Chen, Kristina Yim, Antonia van den Elzen, Matthew J. Hirn, Ronald R. Coifman, Natalia B. Ivanova, Guy Wolf, and Smita Krishnaswamy. Visualizing structure and transitions in high-dimensional biological data.Nature Biotechnology, 37(12):1482–1492, December 2019

    work page 2019

  6. [6]

    Manifold interpolating optimal-transport flows for trajectory inference

    Guillaume Huguet, Daniel Sumner Magruder, Alexander Tong, Oluwadamilola Fasina, Manik Kuchroo, Guy Wolf, and Smita Krishnaswamy. Manifold interpolating optimal-transport flows for trajectory inference. Advances in neural information processing systems , 35:29705–29718, 2022

    work page 2022

  7. [7]

    Improving and generalizing flow-based generative models with minibatch optimal transport, 2023

    Alexander Tong, Kilian Fatras, Nikolay Malkin, Guillaume Huguet, Yanlei Zhang, Jarrid Rector-Brooks, Guy Wolf, and Yoshua Bengio. Improving and generalizing flow-based generative models with minibatch optimal transport, 2023

    work page 2023

  8. [8]

    Trajectorynet: A dynamic optimal transport network for modeling cellular dynamics, 2020

    Alexander Tong, Jessie Huang, Guy Wolf, David van Dijk, and Smita Krishnaswamy. Trajectorynet: A dynamic optimal transport network for modeling cellular dynamics, 2020

    work page 2020

Show all 33 references
  1. [9]

    Spatially resolved, highly multiplexed rna profiling in single cells

    Kok Hao Chen, Alistair N Boettiger, Jeffrey R Moffitt, Siyuan Wang, and Xiaowei Zhuang. Spatially resolved, highly multiplexed rna profiling in single cells. Science, 348(6233):aaa6090, 2015

    2015

  2. [10]

    High resolution mapping of the tumor microenvironment using integrated single-cell, spatial and in situ analysis

    Amanda Janesick, Robert Shelansky, Andrew D Gottscho, Florian Wagner, Stephen R Williams, Morgane Rouault, Ghezal Beliakoff, Carolyn A Morrison, Michelli F Oliveira, Jordan T Sicherman, et al. High resolution mapping of the tumor microenvironment using integrated single-cell, ...

    2023

  3. [11]

    Bridging genomics and tissue pathology: 10x genomics explores new frontiers with the visium spatial gene expression solution

    Nikhil Rao, Sheila Clark, and Olivia Habern. Bridging genomics and tissue pathology: 10x genomics explores new frontiers with the visium spatial gene expression solution. Genetic Engineering & Biotechnology News , 40(2):50–51, 2020

    2020

  4. [12]

    Museum of spatial transcriptomics

    Lambda Moses and Lior Pachter. Museum of spatial transcriptomics. Nature methods, 19(5):534–546, 2022

    2022

  5. [13]

    Exploring tissue architecture using spatial transcriptomics

    Anjali Rao, Dalia Barkley, Gustavo S Franc ¸a, and Itai Yanai. Exploring tissue architecture using spatial transcriptomics. Nature, 596(7871):211–220, 2021

    2021

  6. [14]

    An introduction to spatial transcriptomics for biomedical research

    Cameron G Williams, Hyun Jae Lee, Takahiro Asatsuma, Roser Vento-Tormo, and Ashraful Haque. An introduction to spatial transcriptomics for biomedical research. Genome medicine, 14(1):68, 2022

    2022

  7. [15]

    The expanding vistas of spatial transcriptomics

    Luyi Tian, Fei Chen, and Evan Z Macosko. The expanding vistas of spatial transcriptomics. Nature Biotechnology, 41(6):773–782, 2023

    2023

  8. [16]

    Hyperedge representations with hypergraph wavelets: applications to spatial transcriptomics

    Xingzhi Sun, Charles Xu, Jo ˜ao F Rocha, Chen Liu, Benjamin Hollander-Bodie, Laney Goldman, Marcello DiStasio, Michael Perlmutter, and Smita Krishnaswamy. Hyperedge representations with hypergraph wavelets: applications to spatial transcriptomics. ArXiv, pages arXiv–2409, 2024

    2024

  9. [17]

    On learning agent- based models from data

    Corrado Monti, Marco Pangallo, Gianmarco De Francisci Morales, and Francesco Bonchi. On learning agent- based models from data. Scientific Reports, 13(1), June 2023

    2023

  10. [18]

    Spatial computational modelling illuminates the role of the tumour microenvironment for treating glioblastoma with immunotherapies

    Blanche Mongeon, Julien H ´ebert-Doutreloux, Anudeep Surendran, Elham Karimi, Benoit Fiset, Daniela F Quail, Logan A Walsh, Adrianne L Jenner, and Morgan Craig. Spatial computational modelling illuminates the role of the tumour microenvironment for treating glioblastoma with i...

    2024

  11. [19]

    Neural ordinary differential equations

    Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. Advances in neural information processing systems , 31, 2018. 9

    2018

  12. [20]

    Graph neural ordinary differential equations

    Michael Poli, Stefano Massaroli, Junyoung Park, Atsushi Yamashita, Hajime Asama, and Jinkyoo Park. Graph neural ordinary differential equations. arXiv preprint arXiv:1911.07532 , 2019

    1911 arXiv

  13. [21]

    Multivariate time series forecasting with dynamic graph neural odes

    Ming Jin, Yu Zheng, Yuan-Fang Li, Siheng Chen, Bin Yang, and Shirui Pan. Multivariate time series forecasting with dynamic graph neural odes. IEEE Transactions on Knowledge and Data Engineering , 35(9):9168–9180, 2022

    2022

  14. [22]

    Inferring dynamic regulatory interaction graphs from time series data with perturbations

    Dhananjay Bhaskar, Daniel Sumner Magruder, Matheo Morales, Edward De Brouwer, Aarthi Venkat, Frederik Wenkel, Guy Wolf, and Smita Krishnaswamy. Inferring dynamic regulatory interaction graphs from time series data with perturbations. In Learning on Graphs Conference , pages 22...

    2024

  15. [23]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems , 30, 2017

    2017

  16. [24]

    Graph attention networks

    Petar Veli ˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. arXiv preprint arXiv:1710.10903 , 2017

    2017 arXiv

  17. [25]

    Kaplan, Kyle J

    Michael Hawrylycz, Eitan S. Kaplan, Kyle J. Travaglini, Mariano I. Gabitto, Jeremy A. Miller, Lydia Ng, Jennie L. Close, Rebecca D. Hodge, Brian Long, Tyler Mollenkopf, Shoaib Mufti, Nicole M. Gatto, Eric B. Larson, Paul K. Crane, Thomas J. Grabowski, C. Dirk Keene, and Ed S. ...

    2024

  18. [26]

    10x Genomics Xenium Analyzer

    Office of Science and Technology Resources. 10x Genomics Xenium Analyzer. Accessed: Mar. 04, 2025

    2025

  19. [27]

    Imageflownet: Forecasting multiscale image-level trajectories of disease progression with irregularly-sampled longitudinal medical images

    Chen Liu, Ke Xu, Liangbo L Shen, Guillaume Huguet, Zilong Wang, Alexander Tong, Danilo Bzdok, Jay Stewart, Jay C Wang, Lucian V Del Priore, et al. Imageflownet: Forecasting multiscale image-level trajectories of disease progression with irregularly-sampled longitudinal medical...

    2024 arXiv

  20. [28]

    Deep learning unlocks the true potential of organ donation after circulatory death with accurate prediction of time-to-death

    Xingzhi Sun, Edward De Brouwer, Chen Liu, Smita Krishnaswamy, and Ramesh Batra. Deep learning unlocks the true potential of organ donation after circulatory death with accurate prediction of time-to-death. medRxiv, pages 2024–11, 2024

    2024

  21. [29]

    chronode: A framework to integrate time-series multi-omics data based on ordinary differential equations combined with machine learning

    Beatrice Borsari, Mor Frank, Eve S Wattenberg, Ke Xu, Susanna X Liu, Xuezhu Yu, and Mark Gerstein. chronode: A framework to integrate time-series multi-omics data based on ordinary differential equations combined with machine learning. bioRxiv, pages 2023–12, 2023

    2023

  22. [30]

    Optimal transport: old and new , volume 338

    C ´edric Villani et al. Optimal transport: old and new , volume 338. Springer, 2008

    2008

  23. [31]

    Optimal transport for applied mathematicians , volume 87

    Filippo Santambrogio. Optimal transport for applied mathematicians , volume 87. Springer, 2015

    2015

  24. [32]

    Trajectorynet: A dynamic optimal transport network for modeling cellular dynamics

    Alexander Tong, Jessie Huang, Guy Wolf, David Van Dijk, and Smita Krishnaswamy. Trajectorynet: A dynamic optimal transport network for modeling cellular dynamics. In International conference on machine learning , pages 9526–9536. PMLR, 2020

    2020

  25. [33]

    Matthew, F

    S. Matthew, F. Carter, J. Cooper, et al. Gillespy2: A biochemical modeling framework for simulation driven biological discovery. Letters in Biomathematics , 10(1):87–103, 2023. 10 APPENDIX Algorithm 1 STAGED: Spatio-Temporal Agent-Based Graph Evolution Dynamics Inputs: • Spati...

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.