Behavior-Matched Latent Spaces for Patient-Specific Tau Spreading Models

Computational models of Alzheimer's disease often treat tau pathology as something that spreads along the brain's white-matter wiring and grows locally within each region. Each patient gets a handful of fitted parameters, and patients are then grouped into "mechanistic subtypes" by comparing those parameters. This step is fragile. Different parameter settings can produce nearly the same disease trajectory, so two patients whose disease behaves alike can land far apart in parameter space. Work on compressing neural-network weights ran into the same trap and found a fix: compare and reconstruct models by what they do, not by their raw numbers. We propose to carry that fix over to disease models. A small encoder maps each patient's scans and covariates to a low-dimensional code. A decoder turns the code into the parameters of a spreading model, and training penalizes the code only by how well the decoded model reproduces observed and simulated tau trajectories.

The vascular direction supplies the mechanism we most want this latent space to expose. Existing spreading models run on a connectome averaged over young healthy people, but vascular and endothelial damage plausibly degrades the transport routes themselves. We let the decoder output a patient-specific, low-rank weakening of the normative connectome alongside regional growth rates. We then ask a falsifiable question: does a "transport degradation" direction in the learned space predict future tau better than the normative wiring, and does it track independent vascular measures rather than simply soaking up growth-rate variation?

Core Claim

Fitting each patient separately and then clustering the parameter estimates produces subtypes that reflect how the parameters happen to be written down as much as how the disease actually behaves. A latent space trained with a behavioral loss, meaning one scored on simulated trajectories, should give subtypes that are more stable and better at forecasting, at no cost in fit. With patient-specific transport degradation added, the same space can separate vascular-type spreading from local-production-type spreading. Whether that separation is identifiable is something we test directly, on synthetic patients with known ground truth, before reading anything into real data.

Where This Sits in Prior Work

The neurodegeneration side provides both the model we build on and the gap we target. A coupled-mechanisms modelling framework for neurodegeneration fits a network reaction-diffusion model to longitudinal tau-PET in 110 ADNI subjects. A reaction-diffusion model here is one in which pathology flows between connected regions and also grows within each region. In that paper, each region's growth rate is a sparse combination of six network metrics, estimated per subject with variational inference under a sparsity-inducing horseshoe prior, and 83 subjects are sorted into six subtypes. We reuse its dynamics and data regime. We change two things: we replace independent per-subject fits with a shared behavioral latent space, and we drop the assumption of a normative connectome, which that paper itself names as a limitation. Effects of regional neurotransmitter receptor densities on modelling amyloid and tau accumulation in Alzheimer's disease with network spreading models scales spreading and production by healthy receptor-density maps and reports cross-sectional regional correlations. Its weighting is static and identical for every patient. Our decoder makes the modulation patient-specific and learned, and we keep its receptor maps as candidate regional inputs. On the vascular side, The Endothelial Cell-Related Genes EIF1 and HSPA1B Contribute to the Pathogenesis of Alzheimer's Disease by Modulating Peripheral Immunoinflammatory Responses motivates treating blood-brain-barrier and endothelial dysfunction as an active driver, but works at the molecular level and says nothing about how such damage would reshape spreading. Supplying that link at the scale of the whole brain network is our contribution. The broader mechanistic accounts in A Model for the Development of Alzheimer's Disease are out of scope; they would need measurements we do not have.

The weight-compression side supplies the method. Structure Is Not Enough: Leveraging Behavior for Neural Network Weight Reconstruction shows that an autoencoder over model weights trained only to match the weight values reconstructs classifiers that lose much of their accuracy (74.5% versus 90.4% on the SVHN model zoo, against a 91.0% original average). The fix is to add a loss that matches outputs on realistic query inputs. The same paper reports that noise queries hurt performance. From Parameters to Behaviors: Unsupervised Compression of the Policy Space goes further and drops parameter reconstruction entirely, training decoders of reinforcement-learning policies to reproduce actions. We borrow both the behavioral objective and the lesson that queries must be realistic.

The most obvious cross-over would be the one suggested by Deep Learning on Implicit Neural Representations of Shapes. That work fits small coordinate networks to 3D shapes and classifies their weights directly, which requires a shared random initialization to keep the weights comparable across shapes. Applied here, one would fit such networks to PET scans and compress their weights. We deliberately do not do this. Tau-PET in these cohorts is analyzed at roughly 68–80 regions, so a coordinate network adds no information over the regional vector. It also inherits the weight-alignment problem without any matching benefit. Our "weights" are instead the interpretable parameters of a mechanistic model, and the symmetry we must handle is not neuron permutation. It is the many-to-one map from parameters to trajectories.

Finally, Inference in Spreading Processes with Neural-Network Priors shows that a neural prior over node features improves source recovery in graph epidemics beyond what topology alone gives. That result supports feeding high-dimensional regional covariates into spreading inference. Its experiments are on synthetic random graphs with discrete infection states, however, so it motivates our approach rather than supplying a method we can reuse for continuous tau dynamics.

Research Program

Tau in region $i$ evolves under a patient-specific version of the network reaction-diffusion model:

$$\frac{dx_i}{dt} = \kappa \sum_j A_{ij}\, e^{-[U_z U_z^\top]_{ij}} (x_j - x_i) + \alpha_i(z)\, x_i (1 - x_i).$$

In this equation, $x_i$ is normalized regional tau, $A_{ij}$ is the normative connectome, and $\kappa$ is a global spread rate. The decoder maps the code $z$ to regional growth rates $\alpha_i(z)$ and to a low-rank matrix $U_z$ with $r \le 3$ columns. The factor $e^{-[U_z U_z^\top]_{ij}}$ therefore lies between 0 and 1 and can only weaken a connection, which is the form we want for transport degradation. The growth-rate head receives the six network metrics from He et al. and the receptor maps from Soskic et al. as regional inputs, so the linear-weighting models from those papers are special cases of our decoder.

The encoder reads a patient's baseline scan, any follow-up scans, and covariates. It outputs a distribution over $z$, with dimension 4–8. Training minimizes

$$\mathcal{L} = \sum_{s} \sum_{t \in T_s} \| x^{obs}_s(t) - \hat{x}_s(t) \|^2 + \lambda_B \sum_{s} \mathbb{E}_{q \sim Q} \| \hat{x}_s(\cdot \mid q) - \hat{x}_{s'}(\cdot \mid q) \|^2_{\text{if } z_s \approx z_{s'}} + \beta\, \mathrm{KL}.$$

Each term does a specific job. The first term fits each patient's observed scans at their scan times $T_s$; $\hat{x}_s(t)$ is the trajectory simulated from the patient's decoded parameters. The second term is the behavioral consistency term, weighted by $\lambda_B$. It takes pairs of patients $s, s'$ whose codes are close and requires their decoded models to produce similar trajectories when started from the same query states $q$. The query set $Q$ consists of real baseline tau patterns from other patients, which follows Meynent et al.'s finding that realistic queries matter. The third term is the usual variational regularizer, weighted by $\beta$. Taken together, distance in $z$ is shaped to mean "these patients' disease models behave alike," whatever their raw parameters look like. We integrate the dynamics with a differentiable ODE solver and train on ADNI longitudinal tau-PET, the same data regime as He et al. Because about 110 subjects is small, we pre-train on a large synthetic population simulated from the model with known parameters, then fine-tune on real data.

For vascular validation we need a measure that the model never sees during training. We will use white-matter lesion burden from structural MRI where the cohort provides it, and verifying that availability is the first data task. This independent measure is held out from the encoder and used only to test whether the learned degradation direction tracks vascular damage.

Baselines

The naive reference point is a single-mechanism diffusion model on the normative connectome with one growth rate shared by all patients. This is the baseline He et al. improved upon ($p \approx 3.57 \times 10^{-22}$ for the improvement in fit). It sets the floor every other method must clear.

The strongest existing method is the coupled-mechanisms model of A coupled-mechanisms modelling framework for neurodegeneration: per-subject variational inference with a horseshoe prior over six network metrics. We reimplement it on the same subjects, the same parcellation, and the same connectome. Its prior scale and optimizer get the same number of hyperparameter trials as our model, chosen on the same validation split. Its reported result is a significance test of fit improvement. I found no reported result on Lacuna for its held-out future-scan forecasting error or its subtype stability, so we measure both ourselves. Beating it means two things. First, lower error predicting each patient's last scan from their earlier scans. Second, higher agreement of subtype assignments across bootstrap resamples and random restarts. Both must hold with fit to the training scans no worse.

The second mechanistic baseline is receptor-weighted spreading from Effects of regional neurotransmitter receptor densities on modelling amyloid and tau accumulation in Alzheimer's disease with network spreading models. It reports regional Pearson $r = 0.75$ for early tau (amyloid plus inverted NMDA weighting, versus 0.51 for amyloid alone) and $r = 0.81$ for late tau (amyloid plus inverted 5HTT weighting, versus 0.75). We reproduce its cross-sectional correlations on our split to check that our pipeline matches, then compare on longitudinal forecasting, where it has no reported number.

A purely statistical reference is principal components analysis, together with a small variational autoencoder, applied directly to regional tau trajectories. It gets the same encoder size and tuning budget, and a linear forecaster in latent space. This tests whether the mechanistic decoder earns its place at all.

Three ablations isolate the proposal's ingredients. The first is a linear-Gaussian decoder, which amounts to a hierarchical Bayesian mixed-effects version of He et al.'s model. The second sets $\lambda_B = 0$ and replaces the behavioral term with a loss that matches fitted parameter values, the direct analog of Meynent et al.'s weights-only baseline. The third removes the connectome-degradation head, setting $U_z = 0$. A fourth variant swaps the realistic queries for random tau patterns, to test whether Meynent et al.'s query-sensitivity finding carries over.

Why not just…?

Why not just fit He et al.'s model per patient and cluster the posteriors? This is the main competitor, and it may well match us on forecasting a single patient. Our argument rests on stability. If parameter combinations trade off, as with growth versus spread rate, clustering posterior means mixes together patients whose disease behaves alike. The bootstrap-and-restart stability experiment settles this. If He et al.'s subtypes prove just as stable, the behavioral latent space adds nothing for subtyping, and we will report that.

Why not just use a hierarchical Bayesian model with a population prior? This is a real threat, because it shares strength across patients exactly as our encoder does. It is included as the linear-decoder ablation. If it equals the full model, the honest conclusion is that population pooling, not behavioral geometry or a nonlinear decoder, is what helps. The $\lambda_B = 0$ ablation then tells us whether the behavioral term contributes anything on top of pooling.

Why not just compress the regional tau vectors with PCA or an autoencoder? With about 68 regions and few scans per patient, this could forecast as well as any mechanistic model over short horizons. What it cannot do is name a transport-degradation axis or extrapolate beyond the range of tau burden seen in training. We test this by training on early-stage subjects only and forecasting late-stage ones. If the statistical model still wins there, our case reduces to interpretability alone, and we will say so.

Why not just fit coordinate networks to PET and compress their weights, as in inr2vec? The data are regional, so this approach adds parameters, and a weight-alignment problem from Deep Learning on Implicit Neural Representations of Shapes, without adding information. We therefore do not run it as a main baseline.

Decisive Checks

The first check is synthetic. We simulate patients with known growth rates and known connectome degradation, including deliberately confounded cases where degradation and lower growth produce similar early trajectories. We then measure how well each method recovers the true degradation and the true subtype. If our model cannot separate the two mechanisms in simulation, any vascular reading of real data is off the table.

The second check is held-out forecasting on ADNI: for patients with three or more scans, predict the last scan from the earlier ones. We report regional error and Pearson $r$ against all baselines and ablations, with confidence intervals from resampling patients.

The third check is subtype stability, measured by agreement of subtype assignments (adjusted Rand index) across 50 bootstrap resamples and random seeds, compared against He et al.'s pipeline.

The fourth check is vascular validity. We test whether the learned degradation coordinate correlates with the held-out white-matter lesion burden, and whether the degradation head improves forecasting specifically in high-burden patients. If the degradation head helps forecasting but tracks lesion burden no better than the growth-rate coordinates do, we conclude it is flexible but not vascular. That negative outcome would itself be useful information for the vascular-spreading hypothesis.