When a clot grows, the blood vessel network itself changes shape. Neutrophil extracellular traps (NETs) — webs of DNA thrown out by immune cells — snag platelets and red cells, raise flow resistance along specific vessel paths, and redirect perfusion elsewhere; the resulting stagnant, shear-altered flow then recruits and activates still more neutrophils in a self-amplifying loop. The molecular wiring of this loop is mapped in remarkable detail, from platelet microvesicles carrying HMGB1 into neutrophils to PAD4-mediated chromatin unspooling (Neutrophil extracellular traps mediated by platelet microvesicles promote thrombosis and brain injury in acute ischemic stroke, Neutrophil extracellular traps mediate deep vein thrombosis: from mechanism to therapy). What does not exist is any computational model of the loop as a changing graph: every existing tool either solves flow over a frozen vessel network or generates graphs with no biophysics attached.
This matters because the two things are individually stuck. On the machine-learning side, deep graph generators that predict a full adjacency matrix cost time quadratic (or worse) in node count, which has capped them at graphs of a few hundred to a thousand nodes; a recent line of work — writing a graph as a sequence of edges so that cost grows with the number of edges $M$, not $N^2$ — has pushed deep generation to 10,000–100,000 nodes. On the biology side, no framework can simulate a microvascular bed of realistic size (tens of thousands of segments) being progressively occluded by cellular deposits, let alone do it fast enough to run thousands of virtual patients or search over interventions. I propose to close both gaps with one system: a generative model, trained inside a physics loop, that emits occlusion events on a vascular graph in linear time, and a functional evaluation metric that judges generated networks by what they do to perfusion, not just how they look.
Core Claim
The claim is that NET-driven microvascular occlusion can be modeled as conditionally generated graph edits — each event either deposits a NET mesh on an edge (driving its conductance toward zero), removes an obstructed segment, or adds a collateral bypass — and that an edge-sequence generative model with a continuous physics head can amortize this process at $O(M)$ per event, where dense generators and numerical solvers cannot go. Two ingredients are new. First, the generator is hemodynamically conditioned: its input is the current flow state on the graph (pressures, shear, stagnation flags from a fast solver in the spirit of Graflow: Neural Blood Flow Solver for Vascular Graph), and its output is the next batch of occlusion edits, closing the feedback loop that the biology demands. Second, the model explicitly encodes the dual role of NETs as a conditioning variable with two timescales: rapid vital NETosis (5–60 minutes, non-lytic, cell keeps crawling) versus slow suicidal NETosis (2–4 hours, lytic, terminal web), as characterized by New Insights into Neutrophil Extracellular Traps: Mechanisms of Formation and Role in Inflammation and Neutrophil extracellular traps mediate deep vein thrombosis: from mechanism to therapy. Protective trapping versus pathological occlusion becomes a controllable dial, not a footnote.
The design deliberately extends the sequence serialization of AutoGraph (Flatten Graphs as Sequences: Transformers are Scalable Graph Generators), whose SENT format visits every edge exactly once so that any sequence prefix is a valid subgraph and sequence length is proportional to edge count. AutoGraph's stated limitation is that it handles only categorical tokens, no continuous attributes; NET-Graph adds a regression head over edge-local physical quantities. The single physical relation we need is Hagen–Poiseuille conductance per vessel segment,
$$Q_e = c_e\,(p_u - p_v), \qquad c_e \propto r_e^4 / \ell_e,$$
where $Q_e$ is flow on edge $e$ between nodes $u,v$ with pressures $p$, radius $r_e$, and length $\ell_e$. A NET deposit multiplies $c_e$ by a learned factor in $[0,1]$; the solver then recomputes pressures under flow conservation, and the new flow field conditions the next generation step. One display equation carries the whole physics coupling; everything else is sequence modeling.
Research Program
The program has four components, and I want to be honest about their maturity. The first is a mechanistic reference simulator, built from the biology in the file: vessel trees generated by branching rules, an L-BFGS or Graflow-style pressure solve per timestep, and rule-based NETosis — platelet activation at stagnant, low-shear bifurcations (per Bridging inflammation and venous thrombosis: the NLRP3 inflammasome connection and the PMV–HMGB1 pathway of Gao et al. above), NET deposition raising resistance, thrombin-driven platelet recruitment closing the loop. This simulator is the slow-but-trustworthy oracle: based on Graflow's numbers, an L-BFGS constrained solve takes roughly 12 minutes per network of about 1,700 nodes, so a thousand-step occlusion trajectory on a 20,000-segment bed is far out of reach of direct simulation — which is precisely the gap the learned model fills. The second component is the generative surrogate: a SENT-style decoder-only Transformer whose tokens alternate between structural edits (categorical, exactly as in AutoGraph) and physics tokens (continuous conductance factors, shear, concentration proxies of HMGB1/PF4/Cit-H3), trained on paired simulator trajectories. The third is the feedback harness, alternating generation and solving, with interventions injected at the conditioning level: DNase I removes deposited NET edges, lactadherin clears PMVs upstream, anti-TF blunts the procoagulant input — mirroring the interventions reported in the acute-lung-injury work of Neutrophil extracellular traps are indirectly triggered by lipopolysaccharide and contribute to acute lung injury and the stroke work of Gao et al. The fourth is evaluation: an extension of PolyGraph Discrepancy (PolyGraph Discrepancy: a classifier-based metric for graph generation) from purely topological descriptors to hemodynamic descriptors — perfused-edge fraction, pressure drop across the bed, occlusion-cluster size distribution, time-to-flow-cease. Topological twins with different perfusion should score as different; the existing PGD cannot see that, and this "Hemo-PGD" fills the functional-fidelity gap the metric paper leaves open.
A genuine risk, stated plainly: there is no dataset of real NET-occluded microvascular graphs, so the surrogate is trained and validated against a synthetic oracle. The proposal therefore leans on two anchors to keep it from being self-referential. One is structural realism: training vasculature will match topology statistics extracted by imaging pipelines such as SLIME: robust, high-speed 3D microvascular mapping and 4D microscopy segmentation Microvascular Dynamics from 4D Microscopy Using Temporal Segmentation, with graph-distribution checks via PGD. The other is biological behavior: the model must reproduce, qualitatively and in the correct rank order, the published effects of interventions — for instance, that combined DNase I and anti-Tissue-Factor antibody reduced thrombin generation by approximately 69% in the ex vivo assay of the AIS study, and that lactadherin reduced infarct volume in MCAO mice while DNase I improved lung wet/dry ratios and cytokine levels. Getting these directions and orderings right from mechanism, without fitting to them, is the strongest available test of scientific content.
Baselines
The do-nothing reference is a static network with no occlusion dynamics: a vessel graph with a single Graflow-style pressure solve, matching the setting of Graflow: Neural Blood Flow Solver for Vascular Graph, which itself achieved only about a 15% time reduction relative to a ~12-minute L-BFGS solve on ~1,700-node synthetic trees. This reference defines the floor: any dynamic model must at least reproduce the well-known observation that perfusion degrades over time under inflammatory stimulus, which a static solve cannot represent at all. The strongest existing competitor is the mechanistic oracle simulator itself — rule-based NETosis coupled to the L-BFGS/Graflow solver — because it contains all the biology we encode and requires no learning; the surrogate must beat it not on fidelity (it is the ground truth) but on cost, matching its trajectory distributions on Hemo-PGD and topological PGD while running orders of magnitude faster, since each oracle step embeds a minutes-long solve. A second learned competitor is an unconditioned topology generator in the AutoGraph style, run with the official codebase BorgwardtLab/AutoGraph and post-hoc physics (generate edits, then solve): this isolates whether hemodynamic conditioning — the feedback of flow state into generation — is the key ingredient or whether any linear-time edit generator suffices. All baselines run on identical vessel-tree populations, the same compute budget, and the same hyperparameter search allowance; beating the oracle means Hemo-PGD within its sampling-error band at a wall-clock target of at least 100× faster per trajectory, a target consistent with AutoGraph's reported ~100× inference speedup over DiGress and GEEL's 0.12 s versus GDSS's 15.21 s per graph on Enzymes. A further ablation removes the dual-timescale node-lifecycle tokens (vital versus suicidal), collapsing both NETosis modes into one; if the ablated model matches the full one, the dual-mode encoding is decoration and should be cut.
Why not just…?
Why not just run classic CFD or compartmental ODEs? Classic solvers are forward tools on fixed meshes: Graflow's own comparison shows even a neural solver only trims ~15% off a ~12-minute L-BFGS solve for a modest 1,700-node tree, and a single occlusion trajectory needs thousands of such solves. Compartmental ODEs, the systems-biology default, assume well-mixed compartments and cannot represent that clots initiate at specific bifurcations and that a DNA web snares cells along particular paths — the localization that the PMV-HMGB1, PF4, and NLRP3-tissue-factor studies all emphasize. The experiments will make this concrete by measuring spatial variance of occlusion, a quantity an ODE model is structurally unable to produce.
Why not just use a dense diffusion model like DiGress or GDSS? The file's evidence is unambiguous: GDSS needed 15.21 seconds per graph on Enzymes versus 0.12 for GEEL, dense models hit out-of-memory errors on graphs of a few thousand nodes (Road-Minnesota, PPI in the EDGE paper; past 50,000 nodes for GRAN against BiGG's 100,000 at ~2 GB), and physiological microvascular beds are exactly that size. EDGE's degree-guided discrete diffusion, with degree MMD of 0.0175 versus DiGress's 0.0409 on Community and Cora generation in ~1.7 seconds, is the closest diffusion relative; but EDGE diffuses toward an empty graph without any conditioning on a physical flow field, and its active-node-set convergence is reported to degrade at high step counts — precisely the many-step regime a growing clot requires. I adopt its linear-cost message passing for the solver side, not its generation logic.
Why not just hand-craft occlusion rules instead of learning them? This is the most serious objection, because the oracle simulator is hand-crafted rules. The honest answer is that the rules suffice for forward simulation but not for the inverse questions the proposal cares about — which intervention dose, timing, and target preserve microbial trapping (the protective NET role) while restoring perfusion, across a distribution of anatomies. The learned conditional model amortizes thousands of oracle trajectories into a fast sampler, enabling that search; the decisive experiment is a calibration check where surrogate-guided intervention choices are replayed in the oracle and must match the oracle's own optimal choices. If the surrogate merely memorizes a few trajectory modes, this transfer fails and the learning layer earns nothing — the experiment is designed to expose exactly that.
Why not just extend Graflow directly? Graflow solves pressure on static, rigid graphs; adding hard-coded occlusion heuristics to it is essentially the oracle simulator again, with the same per-step cost and no ability to generate collateral topology or unseen anatomies. The differentiating experiment is generation under conditioning the oracle never saw — anatomy changes, stimulus schedules outside the training grid — where the sequence model's generalization can be tested and the rule-based extension cannot go at all.
Proposed Experiments and Decisive Checks
Experiment one establishes cost: surrogate versus oracle trajectories on vessel trees scaled from 1,700 to 50,000 segments, reporting wall-clock per trajectory and Hemo-PGD distance between surrogate and held-out oracle runs; the scaling curve should be near-linear in $M$, echoing BiGG's $O((N+M)\log N)$ and GEEL's $O(M)$ behavior, while the oracle's cost grows with solves per step. Experiment two establishes biology: intervention replay under DNase I, lactadherin, anti-TF, and PAD4 inhibition, requiring correct direction and rank order against the published effects, including the ~69% thrombin-generation reduction for combined DNase I plus anti-TF in the AIS study and DNase I's reduction of Cit-H3, BALF NET-DNA, wet/dry ratio, and IL-6/TNF-α in the ALI work. Experiment three establishes the dual role as a measurable tradeoff: sweep the vital-versus-suicidal conditioning mix and the pathogen-burden input, and plot pathogen-entrapment proxy against perfusion loss — the model should exhibit a regime where trapping dominates (protective) and a regime where occlusion dominates (pathological), consistent with the double-edged-sword concern raised for broad NET inhibition in the DVT and sepsis literature (Neutrophil Extracellular Traps in Sepsis and Sepsis-Related Organ Dysfunction). Experiment four is the metric contribution: construct topology-matched but perfusion-mismatched network pairs and verify that Hemo-PGD separates them while standard PGD does not, using the PGD finding that estimates stabilize at 256 samples on large-sample benchmarks to fix evaluation sample sizes. The kill criterion is stated in advance: if the hemodynamically-conditioned surrogate does not beat the unconditioned generator-plus-post-hoc-solve ablation on Hemo-PGD, the feedback loop is not earning its complexity, and the project should retreat to the metric and simulator contributions, which stand on their own.
References
- Neutrophil extracellular traps mediated by platelet microvesicles promote thrombosis and brain injury in acute ischemic stroke
- Neutrophil extracellular traps mediate deep vein thrombosis: from mechanism to therapy
- Graflow: Neural Blood Flow Solver for Vascular Graph
- New Insights into Neutrophil Extracellular Traps: Mechanisms of Formation and Role in Inflammation
- Flatten Graphs as Sequences: Transformers are Scalable Graph Generators
- Bridging inflammation and venous thrombosis: the NLRP3 inflammasome connection
- Neutrophil extracellular traps are indirectly triggered by lipopolysaccharide and contribute to acute lung injury
- PolyGraph Discrepancy: a classifier-based metric for graph generation
- SLIME: robust, high-speed 3D microvascular mapping
- Microvascular Dynamics from 4D Microscopy Using Temporal Segmentation
- Neutrophil Extracellular Traps in Sepsis and Sepsis-Related Organ Dysfunction