Spatial genomics · Regulatory dynamics
StageBridge
What can a single cross-sectional tissue sample legitimately tell us about biological progression?
I developed StageBridge for my M.S. thesis at Johns Hopkins to model niche-conditioned regulatory transport across premalignant epithelial states while making recoverability, falsification, and identifiability part of the method itself.
Premalignant progression unfolds over time, but human tissue is typically sampled once. Cross-sectional data can reveal structure, association, and ordering, but they do not uniquely determine an individual's trajectory.
I developed StageBridge to ask a narrower question: which features of regulatory progression can be constrained by stage, spatial context, regulatory structure, and distributional relationships across many lesions?
The model
Context enters the dynamics explicitly.
StageBridge separates an epithelial-state field from a niche-associated field and learns how local tissue context redirects regulatory transport.
The framework combines spatial and single-nucleus transcriptomics with transcription-factor activity, pathway activity, and gene regulatory structure. Local immune, stromal, extracellular-matrix, inflammatory, and spatial context are represented alongside epithelial state.
The goal is to move beyond a generic pseudotime coordinate toward a representation in which changes can be discussed in terms of regulatory programs and tissue context.
Geometry + regulation
Where does context actually bend the field?
The project uses geometric and regulatory readouts to ask whether niche-associated dynamics are localized, structured, and biologically interpretable.
An early context-neutral transport assumption looked acceptable in a coarse-grained latent benchmark. After projection into biologically meaningful regulatory space, prespecified Tier-2 recoverability fell to a cosine of 0.27.
I treated that as a failure of the model rather than a visualization problem. Reformulating the coupling around conditional transport raised Tier-2 recovery to 0.84, but it also changed the scientific interpretation. The supported estimand is a context-conditioned transport relationship, not an individual's reconstructed causal path.
Biological readout
Stable programs, not just attractive trajectories.
Held-out regulatory effects are summarized as stable niche-redirected programs rather than relying on a single latent visualization.
In lung adenocarcinoma, StageBridge recovered structured progression under donor-held-out evaluation and negative controls. In an external PanIN setting, the same framework returned a calibrated null rather than forcing a directional progression signal.
That negative result is important. A flexible trajectory method can almost always produce an ordering; a useful inference framework should also be able to say when the available data do not identify one reliably.
Cross-sectional tissue does not uniquely determine longitudinal dynamics. StageBridge therefore does not claim to recover a person's true causal trajectory.
I evaluate whether context-associated regulatory structure is reproducible using donor-held-out analysis, recoverability tests, matched-context shuffles, degeneracy controls, and cross-disease testing. The broader objective is methodological: make identifiability and failure modes part of the biological analysis rather than an afterthought.
View public research implementation → · Explore thesis figures →
Selected outputs
Related work
StageBridge: Niche-Conditioned Regulatory Transport in Premalignant Epithelial Progression
StageBridge Resolves Inflammatory-Stromal Transition Niches in Lung Adenocarcinoma Progression