Duke Postdoctoral Research · Ongoing methodological development

Mechanistic and Multiscale Modeling for CMV and Immune Translation

I develop quantitative links across biological scales and study systems. The program combines maternal-placental-fetal CMV modeling with relaxed optimal transport for mapping immune-cell populations between human and non-human primate data.

Multiscale modelingUncertaintyOptimal transportScientific ML

One program, two translation problems

Choose a model that respects biology, data limits, and uncertainty.

Congenital CMV transmission spans maternal immune dynamics, placental infection and barrier function, and fetal exposure. Cross-species immune studies add a second translation problem: cell proportions, technical effects, marker panels, and even cell subsets can differ between humans and non-human primates.

The unifying contribution is a disciplined modeling process. I compare alternative mechanisms, identify what available observations can constrain, and retain ambiguity when the evidence does not support a forced match or a single preferred model.

Track 01

Congenital CMV

Maternal viral and immune dynamics, placental passage, fetal exposure, model comparison, sensitivity, calibration, and surrogate modeling.

Track 02

Human-NHP immune mapping

Shared-marker preprocessing, independent clustering, Gaussian summaries, relaxed optimal transport, bidirectional mapping, and uncertainty diagnostics.

Module 1 · Congenital CMV

A modular model separates maternal, placental, and fetal processes.

Ordinary differential equations represent viral and immune dynamics in maternal blood. Spatial or agent-based components describe placental infection and passage. Stochastic components represent variation in fetal exposure. Clinical cohort measurements and placental-organoid experiments provide complementary constraints.

Maternal compartmentViral and immune ODEs

→

Placental interfaceSpatial or agent-based passage

→

Fetal compartmentStochastic exposure and transmission

↕

EvidenceClinical cohorts and organoid experiments

Exploratory comparison of candidate viral dynamics formulations — **Exploratory model-formulation comparison.** Candidate formulations are compared to identify which assumptions materially change target-cell depletion, infected-cell dynamics, and viral load before project-specific calibration.

Parameter sensitivity diagnostics for a CMV model — **Parameter sensitivity diagnostics.** The curves identify which assumptions most strongly affect viral and cellular trajectories and therefore need stronger calibration or wider uncertainty bounds.

Status

These figures are model diagnostics, not validated vaccine-efficacy predictions. Claims about maternal protection or fetal transmission should wait for project-specific calibration and external validation.

Model selection and computation

Faster surrogates are useful only when they preserve the dynamics that matter.

I compare candidate mechanistic formulations before choosing a calibration target, use sensitivity analysis to identify influential and weakly identifiable parameters, and evaluate surrogate models for reducing simulation cost.

Gaussian Process Regression, gradient-boosted trees with Bayesian optimization, and physics-informed neural networks are considered for specific tasks. They are not interchangeable. The choice depends on data volume, computational cost, uncertainty needs, and whether biological structure must remain explicit.

Mechanistic modelsODEs, agent-based simulation, stochastic processes

UncertaintySensitivity, calibration, identifiability, emulation error

SurrogatesGaussian processes, boosted trees, PINNs

InfrastructureDocker, SLURM/HPC, versioned workflows

Module 2 · Cross-species immune mapping

Relaxed transport allows ambiguity and unmatched cell populations to remain visible.

Marker panels, cell proportions, and technical effects differ across species, so nearest-centroid matching can produce misleading one-to-one correspondences. The workflow uses shared-marker preprocessing, independent clustering, covariance-regularized Gaussian representations, Wasserstein distances, and optimal transport with relaxed marginals.

The mapping is bidirectional and reports top probability, entropy, and unmatched mass. These diagnostics show where a correspondence is concentrated, ambiguous, or unsupported.

PreprocessShared markers and batch correction

→

ClusterIndependent Leiden structure

→

SummarizeGaussian means and covariance

→

TransportRelaxed marginals and unmatched mass

→

AuditEntropy, probability, bidirectionality

Hierarchically clustered Wasserstein distance matrix for paired human and non-human primate clusters — **Cross-species cluster-distance diagnostic.** Lower Wasserstein distances indicate closer multivariate cluster summaries. The paired Human and NHP labels are pipeline diagnostics; biological correspondence still requires reviewed annotations and external validation.

Takeaway

The contribution is a process for choosing and testing models.

The goal is not to present one preferred simulator or one forced cross-species map. It is to decide which model is appropriate, what the data can identify, and where uncertainty should remain visible.

View Duke postdoctoral role