Registry
All experiments
Every tile is an independent, self-contained experiment page loaded into this shell as its own microfrontend. Register a new one — hosted anywhere — from the Manage panel and it appears in the menu without touching any existing page.
Start Here — Interactive Pipeline Walkthrough
Open this one first. An interactive, click-through (or auto-play) flowchart of the entire ConnectionMiner pipeline — every input, every matrix it builds, what the solver does, every output. C, Ĉ, and β are live Plotly charts (hover any cell/dot for real type or gene names); every node explains what's on the x-axis, the y-axis, and what the pattern is actually telling you.
Open experiment →MorphologyMorphology — Motor
3D viewer of the motor connectome's skeletons: 707 of the 730 C-matrix types have a local skeleton (the rest are absent from the cm730 bundle), 127,839 nodes after twig pruning and contraction. Orbit, zoom and pan; colour by population, side, depth or neuron; filter by side. Depth fog fades far branches toward the background and an optional slow spin adds parallax, so near and far arbors read apart.
Open experiment →MorphologyMorphology — Visual
3D viewer of FlyWire visual-system skeletons: a stratified sample of one neuron per type per side across all 741 types (1,399 skeletons, 326,698 nodes), drawn from 95,079 reconstructions. Sampling is by side so both optic lobes appear. Colour by subsystem, side, depth or neuron; depth fog and an optional slow spin make the 3D structure readable.
Open experiment →Full Solver RunReplication — Visual disp h2000 seed 0
The grid’s rank-1 visual variant (3_dispersion_binned, h=2000, uniform init, seed 0) re-run end to end from a clean clone. Train r 0.6509 against the original 0.6590, but test r fell to 0.6044 from 0.6503 — a drop of 0.046, beyond the ~0.03 this codebase attributes to cuBLAS reduction order. The G-metacell null for the same variant is reported alongside. Rebuilt with the workbench dashboard (scripts/build_solver_tabs_dashboard.py), the same builder behind the site’s own visual pages — 14 matrix panels over 4 runs, the cell atlas, and 1,591 FlyWire skeletons (1,122 train / 469 test).
Open experiment →Full Solver RunReplication — Visual disp h2000 seed 1
The rank-2 visual variant (seed 1), same protocol. This one reproduces closely — train r 0.6622 vs 0.6599, test r 0.6266 vs 0.6245, both within ±0.002. Read with seed 0: the original rank-1/rank-2 margin on d_test was 0.005 while seed 0 moved by 0.046 on re-run, so the two are tied at the run-to-run noise level and the published ordering does not reproduce. The G-metacell null for the same variant is reported alongside. Rebuilt with the workbench dashboard (scripts/build_solver_tabs_dashboard.py), the same builder behind the site’s own visual pages — 14 matrix panels over 4 runs, the cell atlas, and 1,591 FlyWire skeletons (1,122 train / 469 test).
Open experiment →Full Solver RunReplication — Motor seurat h2000 seed 0
Motor's top-2 variant (2_seurat_vst, h=2000, uniform init, seed 0) re-run from a clean clone: train, test and the G-metacell null, 14 matrix panels over 4 runs. Both real numbers reproduce bit-exactly against the original — train r 0.723821420 and test r 0.435808603, to machine precision. The null arm is the substantive result: on train it is indistinguishable from the real fit (0.713 vs 0.724), so the train block's fit does not rest on G's metacell axis; the test gap of 0.068 is where the signal actually lives.
Open experiment →Full Solver RunReplication — Motor seurat h2000 seed 1
The other motor top-2 variant (2_seurat_vst, h=2000, uniform init, seed 1), same protocol: train, test and G-metacell null, 14 matrix panels over 4 runs. Reproduces bit-exactly — train r 0.723264675, test r 0.438981261. Here the train-side null scores slightly ABOVE the real fit (0.725 vs 0.723), which is the cleanest statement of the same point: permuting which metacell carries which transcriptome costs nothing on train. The test gap is 0.095.
Open experiment →Full Solver RunConnectome Atlas — Type Split (seurat_vst)
All three blocks of the type-level 80/20 split of C on one page, with the FlyWire morphology behind every type. Switch between the held-out block (148 types, β frozen — every connection shown was predicted without the model seeing the block, r +0.588), the training block β was fitted on (593 types, r +0.670), and the whole 741-type catalogue solved in one piece on the same gene set (r +0.699). Hover the C matrix for a pair, or pick a type to light up its neurons in 3D along with its connectome partners. Test and train share no type and the catalogue is their union, so all three blocks are contiguous slices of a single shared skeleton pool in one global frame.
Open experiment →UMAP ViewsRaw MultiomeNN Clusters on UMAP
Stage 1 of the pipeline on its own: every cell placed on the UMAP and colored by its raw MultiomeNN cluster, before any connectome constraint is applied. This is the starting point the solver has to work from — hover any point for its cluster.
Open experiment →UMAP ViewsCell Type Constraints (P support) on UMAP
Stage 2: the same UMAP colored by the cell-type constraints the solver's P matrix supports — which connectome types each cell is still allowed to be after the constraint pass. Compare against the raw clusters view to see what the constraints buy you.
Open experiment →UMAP ViewsMetacell → Cluster Map
Every dot is a metacell, positioned at the UMAP centroid of its member cells and colored by its dominant subtype. Hover any metacell to see its raw cluster; pick a raw cluster to spotlight its individual cells alongside the metacells they collapsed into.
Open experiment →MatricesCluster → Metacell Count Matrix
241 raw clusters × 5,310 metacells, block-diagonal by construction — confirms zero metacells span more than one cluster. Every one of the 5,310 metacells gets its own real column (no representative sampling), labeled m1..m5310, with cell counts labeled inside each block. Scroll to zoom, drag to pan, double-click to reset.
Open experiment →MatricesWiring vs. Distance — C, Soma & Neurite Matrices
Three independent 741×741 matrices, same alphabetical type order on all axes so a row/column lines up across panels: C is the binary synaptic connectome (does type i synapse onto type j, both hemispheres pooled); d(soma) is the mean soma-to-soma Euclidean distance between every cell of type i and every cell of type j; d(neurite) is the true closest-approach distance instead — the minimum distance between any point on any neurite of type i and any point on any neurite of type j, using each cell's full traced SWC skeleton (95,079 skeletons) rather than just its cell body, so it reflects how close the actual axon/dendrite arbors get. Scroll or drag to zoom on any heatmap — full cell-type labels fill in as you get closer — or use the jump-to-type search box to zoom all three panels to the same neighborhood at once.
Open experiment →MatricesMetacell Size Distributions — All 25 Combos
Cells-per-metacell histogram for every gene-selection × metacell-generation combination in the rebalanced grid (real production merge/re-split logic, not a fixed target size). Metacell counts range 3,780–6,695 across combos — this shows the actual shape of that variation: most metacells are small (long right tail), and the tail's heaviness differs a lot by method — graph_ward_supercell and agglomerative_ward_full both show a sharp spike of very small metacells the other three methods don't. Dashed line marks the mean size per combo.
Open experiment →MatricesPhysical Terms vs. Gene Model
Tests whether extending the reconstruction formula (Chat = b0·gene_term + b2/d_soma + b3/d_neurite + b4·overlap_volume) with physical/spatial terms improves the fit to the true connectome C beyond the existing gene-expression model, across all 25 combos, with weights fit via non-negative-constrained least squares. Finding: b2 and b3 are zero in 25/25 combos, b4 is negligible in the 4/25 where it's nonzero — the gene-based reconstruction already correlates with neurite distance and overlap volume at r≈0.31–0.37 in every combo (gene-compatible types tend to be anatomically co-located), so it implicitly re-derives most of the physical proximity signal on its own.
Open experiment →