Dual-Objective Graph Search for Selective NETosis Targeting

Neutrophils — the immune system's first responders — can eject their own chromatin as sticky DNA webs decorated with enzymes. These webs, called neutrophil extracellular traps (NETs), are genuinely two-faced: they immobilize bacteria and fungi at high local antimicrobial concentration, yet when released excessively or cleared poorly they seed clots, injure vessel walls, and amplify organ-damaging inflammation (New Insights into Neutrophil Extracellular Traps; Immunothrombosis in Sepsis). This creates a hard pharmacological problem: any drug that shuts NET formation down systemically risks leaving the patient defenseless against opportunistic infection, a trade-off documented directly in the celastrol influenza study (Machine learning-driven discovery of celastrol). We argue that current computational drug-discovery tools cannot even express this trade-off, let alone optimize against it.

Meanwhile, in graph machine learning, automated architecture search — replacing hand-designed neural networks for graph data with a search procedure that selects, layer by layer, how messages are aggregated and combined — has repeatedly shown that the best architectures are counter-intuitive and dataset-dependent (GraphNAS; NAS-Bench-Graph; GRAPHDeep). But every existing search framework optimizes a single scalar number. The proposal here is to make the dual role of NETs the objective of graph architecture search: a search that jointly discovers which biological variables should be graph nodes (the schema) and which message-passing operations should run on them, under a two-headed criterion that scores a compound both on suppressing pathological NETosis and on sparing antimicrobial defense.

Core Claim

No existing automated graph search system optimizes more than one outcome, and none has been applied to the NET regulatory network. AutoEdge-CCP searches aggregation and combination operators to rank compounds on a cancer-circRNA network, but it treats every link symmetrically and scores a single ranking objective (AutoEdge-CCP). AutoCDRP searches architectures to predict scalar drug-response values, and its performance drops when tested on unseen drugs or cell lines (Cancer drug response prediction with surrogate modeling-based graph neural architecture search). The celastrol and heart-failure studies found NET-modulating compounds with ridge regression, LASSO, and random forests on flat omics tables, never modeling the multi-cellular signaling relays — platelets activated by LPS instructing neutrophils, platelet microvesicles carrying HMGB1 triggering autophagy-driven NET extrusion — that the NET literature shows are essential to the phenotype (Neutrophil extracellular traps are indirectly triggered by LPS; Gao et al., AIS; Machine learning-driven discovery of celastrol; Integrating Clinical Multi-Omics). Our claim is that a search space built around a constrained dual objective over an atomized NET graph will find architectures that rank known selective interventions (DNase, hydroxychloroquine, lactadherin, anti-Tissue-Factor) high on pathology suppression without flagging them as defense-impairing — and that single-objective search, run on the identical graph, will systematically fail the defense constraint.

Research Program

The first deliverable is a heterogeneous graph of the NET-thromboinflammation system, assembled from the mechanisms and databases used by the cited studies: cell types (neutrophil, platelet, endothelium, monocyte), relay mediators (PF4, HMGB1, PMVs, LPS, TRAIL, NLRP3/caspase-1), NET machinery (PAD4, citrullinated histone H3, MPO, elastase, Tissue Factor), outcome entities (thrombin generation, clot lysis resistance, cytokine release, infarct volume, bacterial clearance), and compounds, with edges sourced from the same repositories the network-pharmacology papers draw on — STRING, DisGeNET, GeneCards, OMIM, CTD, Swiss Target Prediction, TCMSP (Chlorogenic acid mitigates ARDS; Integrating Clinical Multi-Omics). Following Atomic HINs, continuous measurements (serum MPO, cfDNA, cytokine titers, expression of NET genes such as the celastrol target Clcn3) are atomized into distinct nodes rather than flattened into attribute vectors, because that work showed heuristic schemas discard signal worth up to 6.2% Macro-F1 and 4.9% ROC-AUC on comparable heterogeneous benchmarks. Crucially, edges are signed and directed: a NET node carries a positive edge to pathogen entrapment and a negative edge to thrombin generation, so the graph encodes the dual role structurally rather than as two labels appended afterwards.

The second deliverable is the search procedure. The architecture space follows GraphNAS's decoupling of per-layer attention, aggregation, and activation choices (GraphNAS), extended with signed-aggregation variants so the search can discover operators that respect edge sign and direction — a capability the GNAS literature explicitly lacks. The schema space is the subset of atomized feature-nodes retained, searched genetically as in Atomic HINs. To keep compute tractable — full tabulation of a 26,206-architecture space cost roughly 8,000 GPU hours in NAS-Bench-Graph — candidate evaluation uses a lightweight surrogate GNN trained on DAG-encoded candidates, the technique AutoCDRP introduced (AutoCDRP). The search criterion is the new part:

$$ \max_{a,\,s}\; \mathrm{AUC}_{\mathrm{path}}(a,s) \quad\text{subject to}\quad \mathrm{AUC}_{\mathrm{def}}(a,s)\;\ge\;\tau, $$

where $a$ indexes architectures, $s$ indexes schemas, $\mathrm{AUC}_{\mathrm{path}}$ measures ranking of compounds by evidence of suppressing thromboinflammatory NETosis, and $\mathrm{AUC}_{\mathrm{def}}$ measures ranking of compounds by preserving antimicrobial function, with the defense threshold $\tau$ set a priori. Both heads share the searched encoder; compounds violating the constraint are reported as immunosuppression risks rather than silently ranked well.

Labels come in two tiers. A small, curated, high-confidence tier encodes the interventions validated in the file's in vivo and ex vivo studies: DNase I with tPA enhanced clot lysis by 34.5% over tPA alone in CVST (Jin et al.); DNase I plus anti-Tissue-Factor antibody cut thrombin generation by about 69% in the AIS cohort (Gao et al.); hydroxychloroquine and PF4 neutralization suppressed NETosis in CVST; lactadherin reduced infarct volume in MCAO mice; celastrol suppressed Cit-H3/NET-DNA and improved survival with baloxavir in influenza (Machine learning-driven discovery of celastrol); chlorogenic acid reduced PAD4, citH3, and MPO in the ARDS rat model (Chlorogenic acid mitigates ARDS). A larger weakly supervised tier scores compounds on NET-relevant gene modules (Clcn3, PAD4, ELANE, MPO) via the CTRP and PRISM screens, the same resources the celastrol pipeline used. We are explicit that the curated tier is small — dozens of compounds, not thousands — which is precisely why the evaluation protocol leans on blind splits and retrospective recovery rather than headline accuracy.

Proposed Experiments

Three evaluations, in escalating difficulty. First, retrospective recovery: the searched model must rank the known selective interventions into the top decile of the pathology head while scoring them above $\tau$ on the defense head, and must flag blanket PAD4 inhibition — which the sepsis and influenza literature identifies as carrying secondary-infection risk (Immunothrombosis in Sepsis; Machine learning-driven discovery of celastrol) — as defense-impairing. Second, blind splits in the AutoCDRP sense: hold out entire compounds, and then entire diseases (leave-CVST-out, leave-AIS-out, leave-influenza-out), because both AutoCDRP and NAS-Bench-Graph — whose authors call transferability "a trap" — show this is where such systems break. Third, a wet-lab-facing check on the ischemic heart failure cohort structure of Wei et al.: using pre/post-treatment multi-omics as a perturbation test, the model's predicted direction of NET-marker change (MPO, cfDNA) should match the observed suppression. Primary metrics are AUC per head and recall at the top 10 and 50 ranked compounds; the honest null hypothesis is that the curated label set is too small for any method to separate from chance, which the naive baselines below will expose immediately.

Baselines

The naive reference point is the standard network-pharmacology setup of the CGA and YQHX studies: a static, unweighted STRING-derived graph with no search, plus a fixed uniform GCN on a heuristic schema (Chlorogenic acid mitigates ARDS; Integrating Clinical Multi-Omics). On symmetric biomedical link-ranking tasks, competing classical methods hover near the 0.50 AUC chance level, which is the number to beat first (AutoEdge-CCP). The strongest existing competitor is the AutoEdge-CCP pipeline — automated edge-embedding architecture search coupled to LambdaMART ranking — because it is the closest published system: automated GNN search over a static heterogeneous biomedical network producing a ranked compound list. Its authors report AUC 0.700 for drug-cancer prioritization on their own task (and 0.989 for circRNA-cancer association), so it demonstrably clears chance on exactly this kind of ranking problem; no reported result found on Lacuna for AutoEdge-CCP or any GNAS system on a NET or thromboinflammation task, so we reimplement it on our graph. Fairness conditions: identical graph inputs, identical label tiers, identical compute cap (2,000 surrogate evaluations, full training of only the top five candidates), and identical tuning budget via random hyperparameter search. Beating it means at least five AUC points higher on the pathology head under blind compound splits, materially smaller degradation from mixed to blind splits, and a defense head it structurally lacks. The third baseline is shallow ML — dual ridge regression on CTRP/PRISM features exactly as the celastrol pipeline built it, plus the LASSO/SVM/RF ensemble of the YQHX study — run on the same labels and splits. Two ablations isolate our ingredients: schema search only, with the architecture fixed to a simple relational GCN as in Atomic HINs, and architecture search only, with the schema fixed to the heuristic STRING-style graph. The single most important ablation is single-objective search: the identical space and surrogate optimizing pathology AUC alone, which isolates the dual constraint as the causal ingredient. Per NAS-Bench-Graph, random architecture sampling is included as a controller baseline alongside the surrogate-guided search, given it proved competitive with RL and evolutionary controllers there.

Why not just…?

Why not just use a well-tuned handcrafted GNN (GCN, GAT, GraphSAGE)? This is partly a real threat: with a few dozen gold-labeled compounds, a tuned GCN might match any searched architecture on the curated tier. The literature says the gap appears when the task is hard and the space is unfamiliar: top-tier architectures in NAS-Bench-Graph consistently beat uniform baselines but are nearly impossible to find manually; AutoCDRP outperformed handcrafted GNNs on blind splits; GRAPHDeep found no single architecture fit all four spatial-omics modalities, and a signed, directed, relay-structured NET graph is more exotic still. Our design makes the comparison decisive rather than rhetorical: the handcrafted baselines run under the identical dual objective, so if they match the searched model, the contribution reduces to the objective and schema, not the search — and we would report it that way.

Why not just use random search as the controller? We concede the strongest version of this objection in advance: random search will be our default controller for architecture sampling, because NAS-Bench-Graph showed it is genuinely competitive on GNN spaces. The proposal's contribution is not a fancier controller; it is the search space (atomized schema, signed operations) and the constrained dual objective. The surrogate-guided variant from AutoCDRP is reported as an efficiency comparison, not a claimed advance; if random search within our space matches it, that is the expected outcome, not a failure.

Why not just use ridge regression or random forests, since that found celastrol? This is the most serious alternative, and we do not dismiss it: the celastrol study succeeded with linear models on bulk transcriptomics and drug screens (Machine learning-driven discovery of celastrol). But it produced one compound, chosen by a pipeline with no representation of the platelet–microvesicle–neutrophil relays that the NET studies show are mandatory intermediaries (Liu et al.; Gao et al.), and no notion of the defense–pathology trade-off; the same shallow-toolbox study of heart failure likewise collapsed multi-omics into flat tables (Integrating Clinical Multi-Omics), and Atomic HINs measured the cost of that flattening at up to 6.2% Macro-F1. The experiment settles it: shallow baselines run on the same labels and the same blind splits, including the defense head. If ridge regression matches our model on both heads under leave-disease-out splits, the graph machinery is unjustified and we will say so; if it wins the pathology head while blind to the defense trade-off, that asymmetry is itself a result.

Why not just optimize the schema and keep a simple fixed model? Atomic HINs makes this appealing: schema search with a simple relational GCN beat complex handcrafted models across eight benchmarks. But that work fixed the message-passing rule and explicitly left schema–architecture co-optimization open, and both AutoEdge-CCP and GRAPHDeep show the aggregation operators themselves decide whether deep topological structure is usable at all. The schema-only ablation is run precisely so this question is answered by data: if schema search alone recovers the dual-objective performance, the architecture component is unnecessary and the proposal simplifies accordingly.

Decisive Checks

The proposal succeeds if three things hold simultaneously: the dual-objective searched model beats the AutoEdge-CCP-style single-objective pipeline on blind compound and leave-disease-out splits by at least five AUC points; the single-objective ablation, run on the same graph, ranks a measurable fraction of compounds that violate the defense constraint (recovering the PAD4-inhibition risk the NET literature predicts); and the schema-atomized graph beats its heuristic-schema ablation by a margin consistent with the 4.9–6.2% range Atomic HINs observed on conventional heterogeneous benchmarks. It fails, informatively, if the curated label set proves too small for any method to separate from the 0.50 chance level — in which case the deliverable becomes the labeled dual-outcome resource itself, since no reported result found on Lacuna for any method, graph-based or shallow, ranking compounds against both faces of NETosis.