scCCVGBen Benchmark Atlas #

scCCVGBen links the curated benchmark roster to reproducible, paper-ready result panels.

Benchmark at a glance #

Benchmark composition #

Model architecture #

Figure 2 summarizes the model-first path through scCCVGBen: input preprocessing, five graph-construction choices, the 14-encoder registry, variational latent flow, dual reconstruction outputs, and the grouped 20-display-metric evaluation table. The publication image is rendered by the reproducible figure pipeline; generated binaries are refreshed from the curated data artifacts.

Metadata distributions #

Explore #

• Datasets → Browse all 200 benchmark datasets with per-dataset detail pages (manifest metadata, PubMed links, GEO accessions, and public-safe redaction notes where applicable).

• Methods → 14 scCCVGBen graph encoders (attention + message-passing), 5 graph construction methods, and 13 comparator entries (12 external baselines plus the scCCVGBen reference row).

• Metrics → 20 publication-display numeric metrics: 3 clustering-compactness scores, 10 DRE neighbourhood/scale diagnostics, and 7 latent-space intrinsic-geometry scores, with two retained restricted-access record handling documented separately.

• About → Paper citation, reproducibility pointers, license.

• Next.js companion explorer → Open the responsive companion browser for the same benchmark roster. It is published separately from this Hugo atlas, so /scCCVGBen/ remains the documentation atlas while /scccvgben-next/ remains the interactive companion route.

• scPortal discovery hub → Navigate the broader public graph that links the homepage, scPortal, this Hugo atlas, and the Next.js companion.