From raw data to biological discovery — automatically

sender_raw[i] = L[i] × mean(R in neighbors of i)
receiver_raw[i] = R[i] × mean(L in neighbors of i)
h(x) = x / (K_h + x)   where K_h = Tukey trimean of nonzero regional expression
LR score = mean( √( h(sender) × h(receiver) ) )   ∈ [0, 1]

Statistical significance via permutation test (N=200 shuffles, conservative p-value).

Step 1 — Context inference
LLM identifies tissue type and disease context from loaded dataset metadata.
Selects from curated marker panels: brain (AD, healthy), breast (tumor, normal), colon, pancreas, lung, liver, kidney.

Step 2 — Automated annotation
For each cell, compute mean z-score across panel marker genes per cell type.
score(cell, type) = mean( z(marker₁), z(marker₂), ..., z(marker_n) )
Assign cell type = argmax(score) with confidence threshold filtering.

Step 3 — User verification
Researchers inspect spatial distribution of annotated types on the canvas.
Marker genes are editable — add, remove, or adjust markers per cell type.
Re-annotation runs instantly with updated panel, enabling iterative refinement
guided by domain expertise and LLM-suggested literature evidence.

Step 1 — Context-aware L–R selection
LLM analyzes loaded tissue type, disease condition, and detected cell types to curate
a tissue-specific L–R database from published literature. AD brain prioritizes neuropeptide
and complement signaling; tumor microenvironment prioritizes immune checkpoint and
growth factor pairs.

Step 2 — Adaptive radius calibration
Interaction radius is set by signaling modality, not a single default:
Contact-dependent (Notch, Ephrin): 15–30 µm · Paracrine (TNF, TGFβ): 50–80 µm
Neuropeptide volume transmission (NPY, SST, CARTPT): 150–200 µm

Step 3 — Profile-driven updates
As the researcher draws regions, selects genes, and runs analyses, the LLM continuously
updates L–R pair recommendations based on emerging results. Custom pairs can be added
from CellPhoneDB, CellChatDB, or user-specified entries. The system learns the researcher's
focus — a user studying microglia–neuron crosstalk receives different suggestions than
one studying vascular–astrocyte signaling, even on the same dataset.

Step 4 — Evidence retrieval
Each suggested L–R pair is linked to PubMed evidence via E-utilities. The LLM provides
biological context — known signaling pathways, disease associations, and experimental
validation status — enabling informed decisions before running the spatial analysis.

The problem with cell-first trajectory
Existing tools (Monocle, Slingshot, PAGA) infer trajectories from gene expression
space — cells are ordered by transcriptomic similarity, ignoring their actual spatial
positions. This produces pseudotime orderings that often contradict tissue architecture.

Region-first approach (SpatialDraw)
Researchers define anatomically meaningful regions directly on the tissue — cortical
layers, tumor–stroma boundaries, vascular niches, or any user-defined microenvironment.
Trajectory is computed across these spatial regions using PSTD (Pseudo-Spatial Trajectory
by Distance), preserving the true tissue geometry that cell-first methods discard.

Within-region mode
Cells within a single region are ordered by spatial position along the minimum spanning
tree, revealing molecular gradients within a microenvironment — signaling molecule
concentration changes from core to boundary, or receptor expression shifts across a
cortical layer.

Across-region mode
Multiple regions are connected by their spatial centroids, computing transition
gene expression changes as the trajectory crosses tissue boundaries — capturing
how molecular programs shift as cells transition between distinct spatial domains.

LLM-guided transition gene and pathway analysis
The LLM automatically identifies genes with significant expression changes along the
spatial trajectory, maps them to known signaling pathways, and interprets the biological
meaning of observed gradients. This provides unprecedented insight into signaling
gradients — how ligand availability, receptor density, and downstream pathway activation
change across tissue architecture. The combination of spatial trajectory with LLM
interpretation reveals both the molecular and structural organization of tissue in a
single, integrated analysis that no cell-first approach can achieve.