DMD Direction Prior
DMD Evidence Explorer
A frozen legacy prior, now separated from its registered successor
Two genes are explicitly not assessed; missingness is never converted to zero.
The five-rule multiverse and gene-level leave-one-source-out analysis are registered, not computed.
Exact-zero direction remains a legacy audit convention, not a biological ROPE.
C1–C7 exist as a governed taxonomy only; no gene receives a class prematurely.
| Algorithm | Role | Status | Freeze |
|---|---|---|---|
dmd-prior-qweighted-context-v0.9 | Published audit baseline | Computed and frozen | Input + result 2026-07-25 |
dmd-prior-source-balanced-v1.0 | Planned source-balanced primary | Preregistered, not computed | No result freeze yet |
Inspect the independent algorithm identity · DMD API 1.2 manifest.
How the current integrated prior was assembled
Four heterogeneous pipelines
Bulk, single-cell and regulatory-context evidence retain their native units.
Context score
sign(effect) × |effect| × min(−log10(q), 50); q falls back to the available P value and is clipped to [10−300, 1].
Integrated direction
For each gene, use significant context rows when any exist, otherwise all available rows; the median signed score defines up, down or mixed/zero.
Integrated prior + conflict
The integrated direction is a prior, not perturbation ground truth.
Source-direction audit
- Within each source and gene, take the median DMD-versus-control effect across its contexts.
- Map positive, negative and exact-zero medians to +1, −1 and 0. Zero directions are excluded from the directional-agreement denominator but the source remains in coverage.
source_direction_agreement = max(n_positive, n_negative) / (n_positive + n_negative). Missing sources do not enter either count.- Both positive and negative sources produce
source_direction_state=conflicted; otherwise the state is direction-consistent up/down. Module membership is reported separately.
Prior-variant stability
The nine frozen prior variants alter the upstream prior/scoring specification used by the historical sensitivity analysis. A candidate is same sign across variants only when all nine counteralignment values are strictly positive or strictly negative; crossing or touching both signs is reported as sign varied across variants. This is not source consensus.
Practical-effect boundary
Unavailable contexts and sources are omitted, never converted to zero. The successor method requires a calibrated practical-effect threshold before positive, negative, negligible or uncertain states can be assigned. Inspect the uncalibrated registry.
Statistical boundary
Context-row q values are retained, but independent source-level standard errors are unavailable. No random-effects meta-analysis, pooled standard error or ground-truth consensus is claimed.
Observed expression and regulatory inference stay separate
| Evidence channel | Pipelines | Current treatment | Independence status |
|---|---|---|---|
| Observed expression | Baseline pseudobulk; delta/DID; DESeq2 source-state | Three source medians exposed separately | Donor/cohort overlap unresolved |
| Regulatory inference | NicheNet target-state | Separate channel; not pooled as observed expression | Donor/cohort overlap unresolved |
The legacy integrated prior historically included the NicheNet target-state channel. The registered successor keeps that channel separate and will not treat it as observed expression. Four pipelines are not presented as four independent cohorts. Inspect typed source and dependence metadata.
Search the gene-level source audit
Coverage reconciliation
| Total gene records | Legacy integrated DMD prior | No integrated prior | Reason |
|---|---|---|---|
| 17,921 | 17,919 | 2 | GARS1: two source-specific observational summaries remain auditable, but no row survived the frozen integrated-prior reconciliation. |
| NEFL: no usable source-specific or integrated DMD-context row is available. Neither missing prior is zero or a negative result. |
Download the machine-readable reconciliation.
Why can Spearman and cosine disagree after source removal?
Spearman measures rank ordering on the intersected genes; cosine measures vector angle and is sensitive to scale, sign and many shared near-zero values. A removal can reorder small effects while leaving the overall direction nearly parallel, or preserve ranks while shrinking/rotating the vector. The metric pair must therefore be interpreted with the exact gene intersection and zero policy, not as interchangeable validation scores.
Registered C1–C7 conflict taxonomy
Near-consensus, one-source outlier, evidence-channel split, context heterogeneity, rule sensitivity, source domination and insufficient coverage are registered labels. They remain unassigned until practical-effect states, source-dependence groups and the five-rule analysis exist. Inspect the taxonomy.
Baseline pseudobulk, delta/DID, DESeq2 source-state and NicheNet target-state evidence.
Source heterogeneity and conflicts remain explicit at gene level.
Independent source-level standard errors are unavailable.
Source matrix and heterogeneity
Each pipeline contributes a distinct observational contrast. Negative cross-source concordance is retained as evidence of heterogeneity rather than averaged away.
| source | n_rows | n_genes | n_contexts | independent_source_level_standard_error_available | random_effects_eligibility | reason |
|---|---|---|---|---|---|---|
| dmd_single_cell_baseline_pseudobulk | 242,609 | 21421 | 21 | no | not_eligible_current_harmonized_table | context rows are nested within source and cannot be treated as independent studies |
| dmd_single_cell_delta_did_final_labels | 240,810 | 20017 | 20 | no | not_eligible_current_harmonized_table | context rows are nested within source and cannot be treated as independent studies |
| sema3c_deseq2_source_state | 157,959 | 21567 | 9 | no | not_eligible_current_harmonized_table | context rows are nested within source and cannot be treated as independent studies |
| sema3c_formal_nichenet_target_state | 24,511 | 14553 | 2 | no | not_eligible_current_harmonized_table | context rows are nested within source and cannot be treated as independent studies |
Pairwise source concordance
Pairwise concordance is descriptive; source rows are not independent studies.
| source_a | source_b | n_shared_genes | pearson | spearman | cosine |
|---|---|---|---|---|---|
| dmd_single_cell_baseline_pseudobulk | dmd_single_cell_delta_did_final_labels | 15980 | -0.223 | -0.446 | -0.218 |
| dmd_single_cell_baseline_pseudobulk | sema3c_deseq2_source_state | 13769 | 0.0183 | 0.163 | 0.0212 |
| dmd_single_cell_baseline_pseudobulk | sema3c_formal_nichenet_target_state | 11456 | 0.162 | 0.218 | 0.164 |
| dmd_single_cell_delta_did_final_labels | sema3c_deseq2_source_state | 13473 | 0.00216 | -0.0623 | 0.00324 |
| dmd_single_cell_delta_did_final_labels | sema3c_formal_nichenet_target_state | 11363 | -0.0336 | -0.0412 | -0.0324 |
| sema3c_deseq2_source_state | sema3c_formal_nichenet_target_state | 11835 | 0.233 | 0.478 | 0.234 |
Source-removal sensitivity
Leave-one-source-out and source-only variants expose how the integrated direction changes.
| variant | n_signature_genes | n_shared_with_full | spearman_vs_full | cosine_vs_full | n_positive | n_negative |
|---|---|---|---|---|---|---|
| full_all_sources | 17919 | 17919 | 1 | 1 | 8,935 | 8,983 |
| leave_out__dmd_single_cell_baseline_pseudobulk | 17544 | 17544 | 0.67 | 1 | 8,214 | 9,329 |
| source_only__dmd_single_cell_baseline_pseudobulk | 16687 | 16687 | 0.499 | 0.0134 | 9,098 | 7,589 |
| leave_out__dmd_single_cell_delta_did_final_labels | 17876 | 17876 | 0.823 | 1 | 9,264 | 8,611 |
| source_only__dmd_single_cell_delta_did_final_labels | 16058 | 16058 | 0.145 | 0.00332 | 7,397 | 8,661 |
| leave_out__sema3c_deseq2_source_state | 17153 | 17153 | 0.732 | 0.0403 | 9,025 | 8,128 |
| source_only__sema3c_deseq2_source_state | 14949 | 14949 | 0.744 | 0.998 | 7,214 | 7,734 |
| leave_out__sema3c_formal_nichenet_target_state | 17910 | 17910 | 0.984 | 1 | 8,861 | 9,048 |
| source_only__sema3c_formal_nichenet_target_state | 11851 | 11851 | 0.527 | 0.226 | 6,584 | 5,267 |
Exploratory source-audit module counts
The current TSV assigns one exploratory module label to each of 17,920 source-audit genes. These counts are mutually exclusive in this export; they are not the 17,919-gene integrated-prior denominator.
| exploratory_evidence_module | n_genes |
|---|---|
| direction_conflicted | 13,616 |
| exploratory_multisource_core | 1,818 |
| multisource_direction_consistent | 1,293 |
| source_limited | 1,193 |
Machine-readable access
DMD API 1.2 manifest · OpenAPI · dynamic gene route api/v1.2/dmd-prior/genes/{gene}.json · Gene modules TSV · Versioned TSV/SQLite downloads