Each point is a cell profiled in one of 3 studies (batches) and belonging to one of 6 true cell types (clusters). A study-specific technical offset vector b pushes every cell in that batch away from its true cell-type centroid — this is the "batch effect" that plagues raw atlas merges. Integration methods (Harmony, Seurat CCA, scVI, etc.) estimate and subtract that offset; the α slider plays the role of that correction.
position:
p_i = c[type_i]·S + ε_i + (1-α)·b[batch_i]·M
mixing score (kNN batch entropy, LISI-like):
score = mean over sampled i of
(1/k) Σ_{j∈kNN(i)} 1[batch_j ≠ batch_i]
annotation transfer accuracy:
ŷ_i = argmin_k ‖p_i - c_k‖²
acc = (1/N) Σ_i 1[ŷ_i = type_i]
- α (integration strength): fraction of the batch offset removed — 0 = raw unintegrated atlas, 1 = fully harmonized.
- Batch effect magnitude: how strong the technical/platform shift is before correction.
- Cluster separation: how distinct the underlying cell types are in expression space.
- Colour toggle: re-colours cells by biological type or by originating study, to visually spot batch-driven vs biology-driven structure.
Real atlas projects (Human Cell Atlas, Tabula Sapiens) run this exact trade-off at scale: too little correction leaves studies as separate islands; too much erases genuine biological differences between cell types.