gu lab / harvard medical school / neurovascular biology
following a neural signal
through a vascular network
jun 2026 - present / undergraduate researcher / computational + wet lab
A nearby neuron firing is only the start of the story. The project asks how neural-activity intensity and vascular topology shape vessel-specific responses across branching cortical arterioles. I connect the data, graph analysis, 3D cell quantification, imaging, and mouse-workflow pieces needed to make that question testable.
workstream 01 / vascular receptive fields
The core computational problem is to distinguish simple physical proximity from propagation along the vascular graph.
The graph model is deliberately interpretable. A result must become an experimentally testable statement about where a vascular signal can travel.
raw-data contract
The current packet contains 24 multiSpot imaging sessions with 120 trials each, 2,880 trials total. Stimulus identifiers separate multiSpot, blank, and full-field conditions; neural and vascular channels remain distinct through ingestion.
localGcamplevel_raw: hemoglobin-corrected neural activity, not a normalized endpointvesselTotalInten: vascular response trace- Raw transcripts and source files remain immutable; corrections live in derived manifests
activity-map construction
Each trial produces a spatial neural-activity field rather than a single point source. Blank controls and session-level scaling are evaluated explicitly before thresholding connected response regions.
- Distance-weighted convolution integrates distributed activity
- Threshold sensitivity is checked across sessions
- Full-field controls are diagnostic, not an assumed universal normalizer
topology-aware distance
Euclidean distance can make two vessel segments look equivalent even when only one lies on the propagating arterial path. The analysis therefore adds geodesic distance, branch order, vessel caliber, and penalties for crossing uncoupled branch sites.
model discipline
The data are nested across trials, vessels, sessions, and animals. Penalized or mixed-effects models are preferred over an opaque neural network because the output must identify a mechanism that can be changed experimentally.
workstreams 02 + 03 / cells, imaging, and protocol validation
The computational model sits inside a larger experimental program. I am also helping turn confocal stacks and live-animal procedures into quantitative, repeatable workflows.
3D smooth-muscle-cell quantification
Confocal z-stacks combine a nuclear channel with smooth-muscle-actin and endothelial context. Elongated nuclei are segmented in three dimensions, constrained to the vessel wall, then manually corrected where curvature or stack boundaries confuse the automated pass.
- Intensity, size, elongation, and spatial-overlap filters
- Human review remains part of the protocol
- Final measure: cell count relative to vessel length and caliber
live-animal imaging
Current work includes head-plate fixation, anesthesia and monitoring, two-photon and widefield acquisition, stimulus setup, z-bound selection, and practical quality checks such as motion, clouding, saturation, and photobleaching.
hypertension validation sequence
The planned validation is staged: establish baseline blood-pressure measurement, rehearse pump implantation with a sham, compare slow angiotensin II delivery with saline, then confirm that the dose changes both systemic pressure and cerebral vessels before scaling the experiment.
histology + documentation
Protocol changes, animal status, imaging usability, stain choices, and acquisition settings are tracked explicitly so a biological difference is not confused with a window, sampling, or processing difference.
I own the current neurovascular analysis build and contribute to the image-quantification pipeline. Experimentally, I handle and scruff mice for head-plate imaging, IP injections, and tail-cuff blood-pressure measurements; assist with perfusion, dissection, and mock osmotic-pump implantation; and acquire and process widefield, two-photon, and confocal data under lab supervision.
current status / what is actually known
This is active work. The page separates completed infrastructure from biological conclusions that still need validation.
analysis packet mapped
The current source packet, trial schema, stimulus controls, neural channel, and vascular channel have been reconciled. Derived visualizations and normalizations can now be audited back to the raw session files.
done / provenance + data orientationtopology lane is bounded
Detailed vascular branch labels exist for a subset of the current sessions. Model claims remain limited to that labeled subset until additional topology is curated.
active / graph reconstructionquantification pipeline in development
The 3D segmentation path is intentionally semi-automated. The goal is a reproducible first pass plus a small, documented manual correction step, not a false claim of perfect automation.
active / segmentation + QCexperimental conclusions pending
No coupling, hypertension, or genotype effect is presented here as established. Those claims wait on the planned controls, complete imaging, and group-level validation.
pending / biological validationmanuscript
Cerebrovascular Responses to Visual Stimulation Are Spatially Tuned by Neural Activity Intensity and Vascular Topology
Contributing author. Results remain bounded to completed controls and validated imaging data.