Scope

Every candidate on this page is a hypothesis, not a target. The release ships scored candidate parquets + manifest + validation panel for downstream wet-lab pre-selection; it has not been validated prospectively. Selecting a row from the released candidates for therapeutic application requires running the full freeze-and-validate pipeline documented at the framework repository.

v0.5 release tag + manifest

GitHub release tag

atlas-brca-v0.5.0-wg-epic-v2-hg38 ↗

Manifest sha256

94cc08baa800…

Reference assembly

hg38 (GRCh38)

Probe array

Illumina EPIC-v2 (GPL33022)

Catalog size

35,406,213 sites

Cohort

GSE322563 MCF-7 vs MCF-10A

Reviewer-access caveat: the sigaihealth/atlas repository is currently restricted to organization members; unauthenticated requests return HTTP 404. Public-read access is a held GigaScience submission gate. The release URL above resolves once the repo flips to public.

Tier breakdown (v0.5)

Tier classifications follow the v0.5 schema (Tier A–E). At v0.5, Tier A is reserved for the future v1 layer set (guide feasibility + off-target methylation + normal-tissue protection); v0.5 populates Tier B–E from the layers it currently audits. The BRCA cell-line cohort is below the framework p_trust_ramp_n=30 floor, so no candidate clears the Tier B trust requirement; the entire catalog splits between Tier C and Tier E.

Tier	candidates	shape
A	—	reserved (pending v1 guide-feasibility layer)
B	0	cohort below `p_trust_ramp_n=30` floor
C	4,917,304	trust-floor candidates
D	0	cell-line decoy archetype not populated
E	30,488,909	no-evidence remainder
total	35,406,213	matches catalog size

`panel_layer` normal-tissue coverage

panel_layer.status = unavailable_controlled_access_pending

No public EPIC-v2 / hg38 normal breast cohort surfaces in any GPL33022 series at the time of v0.5 release. Closure of the BRCA normal-tissue panel-coverage gap requires a controlled-access cohort (dbGaP / EGA). This is recorded honestly as a release limitation rather than substituted by a non-breast panel.

Validation panel composition

The BRCA v0.5 validation panel is 10 rows × 2 selection classes — a smaller, differently-classed panel than the three HM450 patient-cohort cancers (which each ship 18 rows × 3 classes). Selection classes that require enough cohort signal to populate (always_unmethylated_decoy, target_side_methylated_negative) return empty under the framework p_trust_ramp_n=30 floor for the n=2 / n=2 cell-line cohort.

selection class	rows
`delta_beta_only_comparator`	4
`feature_matched_control`	6
total	10

The full per-row schema (candidate coordinates, PAM family, score, off-target detail, feature class) lives in the release artifact atlas_validation_panel_breast.tsv under the release tag above.

Cross-platform incoherence note

The BRCA EPIC-v2 / hg38 release cannot be intersected at the candidate-id level with the three TCGA HM450 / hg19 releases. Candidate-ids are reference-coordinate-anchored; EPIC-v2 / hg38 and HM450 / hg19 differ in both probe space and reference assembly. For cross-cancer comparisons, work at the nearest-gene-symbol level and treat the BRCA cell-line contribution separately from the TCGA patient-cohort releases.

Cite or reproduce

v0.5 release manifest + scored parquet + validation panel — atlas-brca-v0.5.0-wg-epic-v2-hg38 (sha256 94cc08baa800…).
Coordinated release-set overview — /atlas/ § v0.5 Posture B.
Methods framework — PAPER.pdf at tag paper-5-10j.
Roth et al., Nature 2026 — DOI 10.1038/s41586-026-10384-z.