Skip to content

README.md

Latest commit

 

History

History
70 lines (55 loc) · 3.72 KB

README.md

File metadata and controls

70 lines (55 loc) · 3.72 KB
Raw

NCypher DMG non-coding regulatory map

A downloadable, agent-queryable map of every scored somatic non-coding variant in paediatric H3 K27M diffuse midline glioma (DMG), with its regulatory score, the mechanism, the two-axis convergence verdict, and every orthogonal hardening check.

This is the resource leg of NCypher: not a leaderboard, but a reproducible, mechanism-explained shortlist of which non-coding variants to validate next, and why.

Files

file what it is
ncypher_dmg_regulatory_map.tsv The full map — 10,869 scored variants × 37 columns.
ncypher_dmg_regulatory_map.parquet Same, columnar (fast for agents / dataframes).
ncypher_dmg_regulatory_map.converged.tsv The 164 two-axis-converged hits — the useful shortlist, fully annotated.
data_dictionary.md Every column defined, with source analysis, coverage, and caveats. Read this first.
manifest.json Version, row counts, source files + md5, thresholds, coverage.

Quickstart

import pandas as pd
m = pd.read_parquet("ncypher_dmg_regulatory_map.parquet")   # preferred: preserves dtypes

# The validation shortlist: two-axis GO, OPC-context-specific, strongest first.
go = m[(m.verdict == "GO") & (m.a3_context_label == "OPC-specific")]
go.sort_values("chromatin_abs_log2fc", ascending=False)[
    ["host_gene", "chromatin_log2fc", "phylop", "motif_top_tf",
     "a9_target_gene", "a9_eqtl_supported"]
].head(10)

Use the Parquet for programmatic queries. It preserves real booleans/numbers, so m[m.a9_eqtl_supported] works. In the TSV, booleans are the strings "True"/"False", so a naive == True filter silently returns nothing (compare to "True", or load the Parquet).

For agents: the map is also served through the MCP tool get_regulatory_map (manifest + schema + the 164 converged rows, already typed) and the resource ncypher://regulatory-map. Use the tool on hosts that surface tools but not MCP resources.

How to read it (one paragraph)

Each row is one variant. Two axes carry the call: chromatin (chromatin_log2fc, high-impact at |log2FC| ≥ 0.162, fetal-OPC ChromBPNet) and constraint (phylop, constrained at ≥ 2.27, Zoonomia). A variant is converged (verdict GO) only when both fire — 164 of 10,869. For those, the map adds the model-native mechanism (motif_*), the cell-type specificity (a3_context_label), the target gene (a9_*, a9b_*), and an independent second opinion (a5_*). Nulls mean "not assessed at this tier", never "assessed negative".

The honesty rules baked in (so you can trust it)

  • Confidence is highest where axes and orthogonal evidence agree; disagreements are surfaced (e.g. NPAS3 is a top nomination, not cross-confirmed by AlphaGenome).
  • "OPC-specific" means neural, not cardiomyocyte — a developing-brain-progenitor signal, not OPC-exclusive (see A3 caveat in the dictionary).
  • Target genes are mostly the nearest-TSS default (143/164); only 21 are HIGH-confidence reassignments. The NPAS3 Hi-C loop is confirmed by a targeted probe (obs/exp 5.6), even though its ABC-top gene reads AMBIGUOUS — read the dictionary's NPAS3 caveat.
  • These are hypotheses to validate, not proven drivers, and NCypher does not claim to out-predict general variant-effect models.

Provenance / rebuild

Built by scripts/build_regulatory_map.py — a deterministic LEFT-join of the scored sweep with the Claude Science hardening tables (A3/A5/A9/A9b) and the motif calls, keyed by variant_id. Rebuilding needs the local data/ + results/ (fetched, not committed). Cohort: OpenPedCan (open). Model: developing-brain ChromBPNet (fetal-OPC c15). Licence: Apache-2.0.