BulkSeq Studio v0.34.0

Scientific output changes. The pooled-effect rem_padj values now use Benjamini–Hochberg adjustment across every estimable pooled test, including opposite-sign and neutral genes. Recompute meta-analysis results produced before 0.34.0 before interpreting rem_padj. The matching-sign combined-p FDR calls, meta_sig, raw pooled estimates and enrichment selection retain their prior definitions. Synthetic regressions establish these corrected contracts, with independent numerical checks; the nine bundled biological datasets have not been rerun under 0.34.0. Their published counts remain the measured 0.33.0 baseline.

Changed

• Meta-analysis methods and evidence (scientific). The eligibility ledger records both testing families and their sizes, included and excluded studies, post-filter losses, direction categories, and the executed factor, contrast and per-study formula. Each eligible study uses a separate DESeq2 fit for the selected factor; the supported requested meta design permits an intercept, that factor and optionally additive dataset. Extra covariates, interactions and transformations are refused before fitting. Keep a scientifically needed covariate; turning off meta-analysis retains the joint design. The strict input gate refuses missing, malformed, REVIEW_REQUIRED or FAIL check 01 evidence.
• Reports and interpretation. The cross-study report separates combined-p evidence from pooled-effect adjustment and per-study estimates, states the recorded comparison and threshold, and warns when a legacy result lacks the corrected-method marker or orientation record. Comparative figures explain their plotted subsets and link matching source data where available. A combined hit with matching signs is not proof of independent replication; the two-study pooled fit uses a common-effect model and omits heterogeneity fields by reporting policy. Custom ORA reporting distinguishes the supplied background and selected list from the effective annotated populations and shows GeneRatio/BgRatio; its selection and p-values are unchanged.
• Interface and validation. The Samples page exposes study-of-origin assignments; Analysis settings previews eligibility from current editor values while run validation checks saved inputs. Saved reports open directly when their files exist. Figure controls have label relationships and keyboard zoom; compact views can hide the inspector. Recorded checks preserve malformed evidence as failure. Native tool lookup follows the runner environment and local WSL readiness is probed off the UI thread. The CLI retains shared project and output options before or after nested commands.
• Figure typography and spacing. The default-serif figure setting prefers Times New Roman when available to the R renderer; otherwise it names an installed serif fallback and fails if none is available. Explicitly configured families should be installed in the analysis backend. Volcano ranked-key text and heatmap/correlation guides have more clearance. These presentation changes do not alter model results or enrichment calculations.
• Documentation website. The handbook now follows input selection, first run, analysis, results and upgrade decisions. Its synthetic lessons explain BH rejection-set limits, effect screens and STRING associations without presenting them as new biological validation.
• Packaging and upgrades. The x86-64 Windows installer updates an existing installation in place, without automatically running its old uninstaller or recursively deleting the install directory. Windows builds use isolated staging and a restricted dependency search path so earlier release artifacts and the root launcher remain untouched and unrelated PATH libraries are not collected from the ambient shell. Frozen application tests require the complete GUI/WebEngine sentinel followed by a clean process exit. The release package comprises the Windows installer and portable ZIP plus the Linux x86-64 AppImage, its functional zsync companion and portable tarball; optional checksum manifests are omitted while integrity checks remain local. Every binary bundle includes the MIT LICENSE.
• Full backend lock verification. The linux-64 lock declares exact Conda package versions and builds, with the same four pip pins. After tool and R load probes, setup checks the installed inventory before writing a lock success marker. Drift triggers one bounded reinstall and recheck; unresolved drift fails setup. Extra packages are reported but not deleted by the validator. A floating-spec installation remains labelled fallback and makes no exact-lock claim. Clean-system execution is a separate release gate.
• g:Profiler transport. The enrichment route sets the pinned gprofiler2 client to HTTPS before service requests, preserving any selected archive path and all query and statistical settings. This transport repair has no new biological rerun claim.
• Direct Entrez identifier mapping (scientific). OrgDb-backed enrichment now retains directly supplied Entrez IDs during mapping; a run using them previously could fail before enrichment. Revalidate and rerun an affected analysis rather than treating its saved empty or failed output as a negative result. No bundled biological dataset was rerun for this correction.