BreakingFix
- Results change in this release. After last week’s bulk, ATAC and methylation audit we checked the single-cell RNA-seq nodes the same way, against Scanpy and the field’s best-practice guides. The standard path — QC through clustering and marker genes — already matched Scanpy exactly. Nine defects sat one step off that path and finished green. If any item below applies to your work, re-run it; opening a saved .bioflow file restores cached results without re-running, so it cannot warn you on its own.
- Sub-clustering was returning the parent’s results. After Subset Data, re-running PCA, Neighbors, Clustering and UMAP silently reused the parent dataset’s values, so a subset kept the parent’s cluster labels and embedding instead of being re-analysed (on a PBMC test set: 2 inherited clusters where a fresh analysis finds 10). Nodes now recompute whenever their input data differs. Any earlier sub-clustering should be re-run.
- After Harmony batch correction, Neighbors and t-SNE now use the corrected embedding by default. Previously the default was uncorrected PCA, so the graph, clusters and UMAP ignored the correction unless you changed the setting. The new Auto choice picks Harmony when it is present; a workflow saved with Standard (PCA) keeps it and now shows a warning, and choosing an embedding that does not exist is an error instead of a silent fallback.
- Louvain clustering now applies its resolution and random seed. Without the optional louvain package the node had been running a variant that ignored both, so changing the resolution barely changed the result and the same seed gave different clusters. Louvain labels will differ from earlier runs; Leiden, the default, was not affected.
- Top DEG Gene List applies the significance filters it offers (adjusted p-value, log2 fold change) — they were being ignored, so lists could include non-significant genes. Single-cell pathway enrichment now tests against the genes that were actually measured instead of the whole Enrichr library, which changes p-values and terms, as the bulk analysis did last release.
- Several inputs that used to finish green now stop and explain: QC or doublet detection on data that is already normalized (QC uses a counts layer when one exists), a second log transform, normalizing log data, and a Scrublet run that cannot choose a threshold — which used to report zero doublets. Data whose .raw holds raw counts no longer makes marker-gene fold changes meaningless.
- QC: the haemoglobin percentage now counts haemoglobin genes only (it also counted HBP1, HBS1L and HBEGF), and Auto-fill no longer switches on a 0% haemoglobin filter that removed about one cell in ten. A mitochondrial prefix that matches no genes — mouse data is mt-, not MT- — now shows a warning. The Methods paragraph describes the filters and parameters that actually ran.
- We then audited the remaining single-cell nodes the same way — trajectory and RNA velocity against scVelo, CellRank against CellRank 2, spatial against Squidpy, annotation and cell–cell communication against CellTypist, LIANA and CellPhoneDB. The core calculations matched; twenty defects around them finished green with wrong values. The items below are result changes.
- CellRank with CytoTRACE reported differentiation backwards. Without velocity moments it scored raw counts or scaled values instead of log-normalized expression, so on a pancreas development set the CytoTRACE pseudotime ran against the known stages (Spearman −0.83; +0.65 now, matching the CellRank 2 recipe) and a progenitor population was called terminal. The shipped CytoTRACE templates should be re-run. CellRank summaries, reports and the AI assistant also always showed zero terminal states because CellRank 2 renamed the key; they now report the real ones. Lineage drivers and CellRank plots reuse the terminal states Cell Fate chose instead of re-fitting and renaming them, and drivers use all log-normalized genes by default. The three CellRank templates now ask for four macrostates grouped by Leiden cluster, so fate probabilities no longer collapse to a single lineage.
- RNA velocity: an input that was already log-transformed was logged a second time, and the velocity graph used an approximation by default that reversed the arrow of about one cell in five on a hippocampus test set. The graph is now exact by default (the scVelo default), which changes arrows and terminal states. Diffusion pseudotime without a chosen root now starts from the first informative diffusion component instead of an uninformative one and says so, and after Subset Data it is recomputed instead of reusing the parent’s root and diffusion map.
- LIANA, the spatial ligand–receptor test and spatial bivariate scoring now score log-normalized expression. LIANA used scaled values by default (39 pairs with every magnitude rank missing on a PBMC test, 248 now), and the spatial methods used .raw even when it held counts. Label transfer (Cell Mapper) works on scaled data instead of producing missing values, and refuses independently computed PCA spaces, which cannot be compared. CellTypist confidence is now the probability of the label shown, and a model that shares too few genes with your data — wrong species, or Ensembl IDs — is an error instead of a confident wrong answer.
- Inputs that used to finish green now stop and explain: spatial coordinates whose cell IDs match none of your cells (they were attached in row order and reported as a 100% match), a density QC that would remove every cell (Visium coordinates are pixels, not micrometres), deconvolution of log-transformed data, a PAGA or diffusion-pseudotime group column that does not exist, a Cell Fate cluster column that does not exist, and CellPhoneDB basic mode, which returned zero interactions.
- Spatial statistics: Spatial Autocorrelation could report zero spatially variable genes on a whole Visium section, because the multiple-testing correction turned every value missing when a single tested gene was constant; it now reports them (about 2,800 on the test section). Permutation p-values are used when you ask for permutations, only positive autocorrelation counts as spatially variable (the opposite tail is reported separately), neighbourhood enrichment and ligand-receptor results are corrected for multiple testing, and Ripley no longer thins every group to 50 points, which had biased the G and F functions. Spatial Neighbors now picks the spot grid automatically for Visium, and BANKSY normalizes raw counts by library size first, as the reference implementation does. Spatial counts and p-values from earlier runs will differ.
- RNA velocity builds its own neighbour graph the scVelo way instead of reusing the one from the clustering chain, which had reversed several percent of arrows; it warns when connected after Scale, which removes genes velocity needs. Draw Graph and UMAP now actually start from the PAGA layout when asked (the Draw Graph default), PAGA with RNA velocity works, and Auto Annotation recomputes marker genes instead of sending the model markers from a different clustering after Subset. The CellPhoneDB p-value cutoff now changes the results and plots, Annotation Concordance compares only label columns, and the CellRank templates are wired so they run on files that do not already contain UMAP and clusters.
- The desktop app moves from Electron 28 to Electron 44, bringing two years of Chromium security fixes. Along with it: the app window and the sign-in window can no longer be navigated to other sites or local files, saved credentials are filled in only on the Oreumics sign-in page, sign-in tokens are stored encrypted, pages other than the app cannot request camera or microphone access, and the diagnostic log no longer records the connection key used by AI assistants. Dragging a file onto a path field still fills in the full path.