Does Sex Act on Cells or on Their Conversations? Locating Sex Effects in Learned Tissue Interaction Models

Many inflammatory blood scores predict disease well in men and poorly in women. A cheap ratio of white-cell counts is one example: the monocyte-to-lymphocyte ratio carries an odds ratio of 8.95 for arterial stiffness in men and much weaker signal in women (Wang et al.). The usual explanation is that estrogen "masks" inflammation. That explanation hides a question nobody has answered. Does sex change how individual cells behave internally, meaning the gene programs they run? Or does it change how cells talk to each other, meaning the signaling between neighbors that turns local inflammation into tissue damage? Epidemiology cannot answer this, because it only sees bulk blood counts. Graph-based multi-agent models can. These models treat each cell as an agent that updates its own state and exchanges messages with its neighbors, so they have separate parts for "self" and "conversation."

We propose to turn that architectural split into a measuring instrument. We fit a learned model of cell-cell interaction dynamics to spatial transcriptomics, which is gene expression measured at known positions in tissue. Donor sex is allowed to enter only the model of each cell's internal dynamics, only the model of neighbor-to-neighbor messages, both, or neither. We then ask which placement predicts held-out donors best, with the null calibrated by permuting donor sex labels. Before trusting any answer on real tissue, we check on simulators with a planted ground truth that the procedure recovers where the effect lives.

Core Claim

The claim is that the location of a sex effect in a learned multi-agent tissue model can be identified. A factorized conditioning test can tell whether sex modulates cell-intrinsic dynamics or inter-cellular communication. This is a scientific question that neither direction can pose alone. The clinical masking story (Zhou et al.; Wang et al.) invokes estrogen suppressing cytokine release, which is a communication effect, and also estrogen altering transcription, which is an intrinsic effect. These two mechanisms make different predictions inside the model. In the multi-agent literature, message-passing models such as HAMA and Relational Forward Models use one set of interaction rules for all agents, with no global field that modulates messages. The Relational Forward Models work does show that the size of edge messages can be read as a measure of influence between agents, and that reading is what makes our edge-level attribution interpretable.

In symbols, cell $i$ has expression state $x_i$ and neighbors $j$ in a spatial graph, and the donor carries sex label $s$. The model evolves as

$$\frac{dx_i}{dt} = f_\theta(x_i, s_f) + \sum_{j \in N(i)} \alpha_{ij}(s_e)\, g_\phi(x_i, x_j, s_e).$$

Here $f_\theta$ is the within-cell regulatory network, $g_\phi$ and the attention weights $\alpha_{ij}$ form the ligand–receptor-restricted messaging between cells, and $s_f, s_e$ are copies of $s$ that are either supplied or masked to a constant. The four variants (neither, $s_f$ only, $s_e$ only, both) are the hypotheses we compare.

Research Program

The method comes from STAGED. STAGED models each cell as an agent with a subcellular gene-network component and a tissue-level attention over known ligand–receptor pairs, and it evolves states with graph neural ordinary differential equations, which are continuous-time dynamics whose update rule is a learned graph network. It found microglial programs that differ near amyloid plaques in the Seattle Alzheimer's Disease Brain Cell Atlas (SEA-AD). STAGED has no sex input at all, which is the gap that motivates this work. It also already separates the self and conversation terms, which is what makes the test cheap. Our changes are small on purpose. We add a low-dimensional feature-wise modulation (a learned scale and shift driven by $s$) to either or both components. We add a sparsity penalty on the modulation so that the model does not spread a weak effect across both parts. We do all model selection on held-out donors, never on held-out cells from donors seen in training.

The program has three stages. The first is identifiability on simulators. We take the Gillespy2 multicell oscillator and the oncolytic-virus glioblastoma simulator used by STAGED, and the Boid flocking simulator used by TGV-CRN, because it is a clean agent-coordination testbed. In each we plant a binary "field" that changes either the agents' intrinsic update or their interaction kernel, at varied effect sizes and numbers of simulated donors. We measure how often the four-way comparison picks the planted location. This stage tells us the smallest effect and donor count at which the test can be trusted.

The second stage applies the test to SEA-AD, stratified by cell type and by distance to plaque, using donor sex from the atlas metadata. Microglia near plaques are the primary target because that is where STAGED found regime changes.

The third stage is interpretation. For whichever component carries the effect, we report which regulator edges or which ligand–receptor channels change most with $s$. The ranking uses message norms in the spirit of Relational Forward Models and edge-pruning deltas.

We are explicit about what this cannot show. Alzheimer's donors are likely mostly older, so female donors will largely be post-menopausal; we will verify this from metadata. The result therefore concerns sex as recorded, not circulating estradiol, and it cannot directly test the estrogen-masking story. The coronary "grace period" and post-menopausal acceleration reported by Yang et al. remain the motivating phenotype. We do not claim to explain them from brain tissue. The contribution is a validated test plus a first answer on the best available spatial dataset. If SEA-AD metadata includes age, we will also report an age-by-sex split as exploratory.

Baselines

The naive reference is the unconditioned STAGED-style model with sex masked in both components. Every other variant must beat it on held-out donors. Because adding any input can improve fit by chance with few donors, the naive model is paired with a permutation null: the same conditioned models trained with donor sex labels shuffled across donors. A variant counts as a real effect only if its held-out gain exceeds the 95th percentile of the permuted gains.

The strongest existing method is STAGED itself, run with sex concatenated as an ordinary input feature to every cell. This is how a practitioner would "add sex." It is the strongest competitor because it has the same capacity and the same biological constraints as our model, and differs only in not localizing the effect. STAGED reports Spearman agreement with observed trajectories on SEA-AD, but no reported result found on Lacuna for STAGED's exact correlation values, so we will rerun it ourselves. All variants share the same data splits, graph construction, ODE solver, parameter count (padding the unconditioned model to match), and hyperparameter search budget. Beating it means one of two things: equal or better held-out prediction together with a localization that the concatenation model cannot provide, or, on simulators, strictly higher planted-location recovery.

The key ablation compares the factorized variants against each other. The edge-only versus node-only contrast is the proposal's central ingredient. The sparsity penalty is also ablated to show that localization does not come only from the regularizer. For the Boid stage, TGV-CRN's graph variational recurrent backbone serves as a second dynamics model, to check that conclusions do not depend on STAGED's architecture. No reported result found on Lacuna for TGV-CRN's numerical errors on Boid.

Why not just…?

Why not just train separate models for women and men and compare them? This is the most natural alternative. Its problem is that two independently trained networks differ in every parameter, so comparing them cannot say whether the difference lives in intrinsic dynamics or in communication. Splitting also halves the donors available per model. We will run it anyway. If separately trained models show higher held-out error than the shared factorized model, the shared structure is earning its place. If they show equal error, the localization claim must rest on the simulator-validated test alone.

Why not just run a sex-stratified differential expression and ligand–receptor enrichment analysis? This is a real threat, because it is cheap and interpretable. It can report that some ligand pairs are more expressed in one sex. But more expression of a ligand does not mean more signaling effect on neighbors, and a sex difference in expression can arise from intrinsic programs alone. The simulators settle this directly. We plant a purely intrinsic effect that also shifts ligand expression, and check whether enrichment analysis wrongly flags communication while the factorized dynamics model does not. If enrichment performs as well at attribution, we will say so and recommend it.

Why not just add sex-by-biomarker interaction terms or splines to the clinical regressions? For the population question this is appropriate and already done. Li et al. found a threshold near 14.21 for the neutrophil-percentage-to-albumin ratio overall but a linear pattern in women, and Liu et al. found a male-skewed linear effect of the systemic immune-inflammation index on stroke. Those tools describe how risk depends on a blood score. They contain no tissue-level variables, so they cannot say where in the tissue the sex difference acts. We treat them as the phenotype to explain, not as competitors for this question.

Proposed Experiments and Decisive Checks

On the simulators, the primary metric is planted-location recovery accuracy across effect sizes and donor counts, reported as a curve with the permutation null. The secondary metric is the correlation between the planted and the recovered modulation magnitude. On SEA-AD, the primary metric is held-out-donor prediction error of expression trajectories and the Spearman correlation STAGED uses, computed per cell type and per distance to plaque. It is reported relative to the naive model and the permuted-label distribution, with confidence intervals from donor-level bootstrap.

The decisive checks are these. First, if the simulator stage cannot recover planted locations at effect sizes plausible for SEA-AD's donor count, we stop and report the test as underpowered for current data. That outcome is still useful as a design guide for future sex-balanced spatial atlases. Second, if no variant on SEA-AD beats the permuted-label null, we report that sex effects on microglial interaction dynamics near plaques are not detectable at this scale. Third, a positive result in which one component clearly wins, stable across bootstrap and across the STAGED and TGV-CRN backbones, gives a concrete and testable mechanistic hypothesis for which channels to perturb, which is what the masking literature currently lacks.