Skip to content

example_run.md

Latest commit

 

History

History
183 lines (151 loc) · 10.4 KB

example_run.md

File metadata and controls

183 lines (151 loc) · 10.4 KB
Raw

NCypher triage: two worked examples on the real cached data

Both examples below are run on the real DMG discovery sweep (data/dmg/sweep_result.tsv) and the mining tables (data/figures/mining/), so they are demonstrable without live Modal or a GPU. The axis numbers and the memo text are the actual output of score_variant (verified against mcp/server.py); the skeptic, therapy, and honesty layers are what this skill adds on top. British English, no em dashes.

To reproduce the raw tool output:

cd /Users/faith/Desktop/NCypher
.venv/bin/python - <<'PY'
import sys; sys.path.insert(0, "mcp"); import server
# FastMCP wraps the tool in some versions; .fn is the callable, else the function itself.
fn = getattr(server.score_variant, "fn", server.score_variant)
r = fn(variant_id="chr14-33788719-A-G")
print(r.structured_content["memo"]["verdict"])
PY

Example 1 - NPAS3 (the lead "validate first" candidate)

Variant: chr14-33788719-A-G · gene NPAS3 (intron) · DMG patient PT_DS0HZCSD

1. Intake

Single somatic non-coding SNV, GRCh38, chr14-33788719-A-G. In the cached DMG sweep, so the full two-axis call is available.

2. Score (from score_variant)

  • chromatin: log2FC -0.710 (accessibility loss), JSD 0.116, high-impact yes (|log2FC| 0.710 is far above the calibrated p99 of 0.162).
  • constraint: phyloP 4.68, constrained yes (threshold 2.27).
  • function: not available (not in the MPRA/DAV set).

3. Converge

2 axes available, 2 impactful, agreement 1.0, in domain (trevino_2021.c15). Verdict from the engine: GO (promote = true).

Real memo text returned by the tool:

GO: validate chr14-33788719-A-G first. Evidence converges (log2FC=-0.71, JSD=0.116 (loss) phyloP=4.68 (constrained @ 2.27)); predicted chromatin loss.

4. Mechanism

Model-native DeepSHAP shows an 80% collapse of local contribution at the variant base (hero_candidates.tsv: deepshap_collapse = 0.802). The accessibility loss is consistent with silencing of a tumour-suppressor enhancer. The crude PWM top-motif call for this base is noisy (CTCF-like, low information), so do not assert a specific TF here (guardrail 6): report the contribution collapse, not a clean motif. Driver of the call: both axes fire, and NPAS3 sits in the constraint-driven neurodevelopmental programme.

5. Skeptic step (tried to refute it)

  • Brain-expression / length confounder? NPAS3 is one of the 6 canonical neural genes in the converged set, and it earns that credit through the constraint axis, not chromatin alone (chromatin-only neural enrichment is n.s., p = 0.363). The converged neural enrichment (2.62x, p = 0.029) is suggestive, not decisive, and is reported as such.
  • Within noise? No. |log2FC| 0.710 clears the p99 (0.162) by ~4.4x; phyloP 4.68 clears 2.27 comfortably. Robust on magnitude.
  • Constraint doing all the work? No: the chromatin axis adds a real loss-of-accessibility signal plus the 80% contribution collapse; this is not a conservation-only call.
  • Enhancer-to-gene link weak? NPAS3 is a ~0.84 Mb locus; the host gene is the likely target but Activity-by-Contact + Hi-C is needed to confirm this element regulates NPAS3 rather than a neighbour. Link confidence: Medium.
  • Mapping / recurrence artefact? The NPAS3 signal is gene-level recurrence (3 distinct high-impact variants, 3 different enhancers, 3 patients: chr14-33788719-A-G converged, plus chr14-33057410-A-G and chr14-33282787-C-A high-impact), and it is externally validated: NPAS3 is a tumour suppressor in glioma (Moreira et al., Am J Pathol 2011, PMID 21703424) and a paediatric-medulloblastoma tumour-suppressor master regulator (Michaelsen et al., Sci Rep 2025). But the per-gene binomial p = 0.014 does not survive Bonferroni across ~6,800 genes, and only 1 of the 3 is fully converged. So: elevate as a biologically-motivated lead, not a genome-wide driver.

What would kill it: base-edit chr14-33788719 in an OPC-like DMG line; if NPAS3 expression and local accessibility do not move beyond noise, and CRISPRi of the element does not lower NPAS3 either, the call is dead.

Survives as the strongest single lead to validate first, at MEDIUM confidence.

6. Go/no-go memo

  • Verdict: GO, validate first. Confidence MEDIUM (two-axis; function axis absent).
  • Decisive experiment: base-edit the A>G base at chr14:33788719 in SU-DIPG-XIII or HSJD-DIPG-007, in parallel with CRISPRi of the element; read NPAS3 mRNA, local ATAC accessibility, and the OPC-differentiation phenotype.
  • Kill criterion: no change in NPAS3 beyond noise after both the base edit and CRISPRi.
  • Therapy angle (Axis D, tumour-suppressor lost): NPAS3 cannot be directly restored; nominate the EZH2 / EZHIP reactivation axis (H3 K27M silencing is epigenetic; residual PRC2 is a dependency, Mohammad et al., Nat Med 2017, PMID 28263309) or use the variant as a prognostic / stratification biomarker. Therapy confidence Low, biomarker Medium. NCypher does not predict drug response; CNS delivery (BBB) is the binding constraint.

7. Honesty

NPAS3 is a MEDIUM lead, not a driver. What this is not: not a recurrent non-coding driver (gene-level, p does not survive Bonferroni), not a validated target, not a restore-therapy nomination.

Example 2 - SOX2-OT (the star mechanism chain)

Variant: chr3-181317760-C-T · gene SOX2-OT (intron) · a single DMG patient

1. Intake

Single somatic non-coding SNV, GRCh38, chr3-181317760-C-T. In the cached sweep.

2. Score (from score_variant)

  • chromatin: log2FC +0.26 (accessibility gain), JSD 0.067, high-impact yes (clears the p99 of 0.162; impact percentile 0.999 in the cohort).
  • constraint: phyloP 3.54, constrained yes.
  • function: not available.

3. Converge

2 axes available, 2 impactful, agreement 1.0, in domain. Verdict: GO.

Real memo text returned by the tool:

GO: validate chr3-181317760-C-T first. Evidence converges (log2FC=+0.26, JSD=0.067 (gain) phyloP=3.54 (constrained @ 2.27)); predicted chromatin gain.

4. Mechanism

An accessibility gain in an evolutionarily constrained element inside the SOX2-OT body, 394 kb 5' of SOX2, the master TF of the OPC-like / glioma-stem state that defines this tumour (Suva et al., Cell 2014, PMID 24726434; Filbin et al., Science 2018, PMID 29674595). A gain fits a variant that reinforces the core stemness circuit. This variant is a small-effect GAIN, so it is not in the top-40 DeepSHAP hero set; running model-native contribution scoring on it (the accessibility-gain / motif-creation analysis) is the next step to show a created motif at the alt base. Driver: chromatin gain plus constraint; the mechanistic weight is the target-gene biology, not effect size.

5. Skeptic step (tried to refute it)

  • Brain-expression / length confounder? SOX2-OT is SOX2's antisense lncRNA, not on the strict canonical neural gene list, so it does not inflate the neural-enrichment count; its relevance is the OPC core circuit, argued from biology, not from the enrichment test.
  • Within noise? Partly a concern. The chromatin effect is modest (+0.26); it clears the p99 threshold but is small in absolute terms. phyloP 3.54 clears 2.27 but not deeply. Honestly a Medium, not a slam-dunk, on magnitude.
  • Constraint doing all the work? No, the gain is a real chromatin signal, but it is modest.
  • Enhancer-to-gene link weak? This is the key weakness. The element is 394 kb from the SOX2 promoter. It is within the same locus / TAD and a canonical long-range enhancer-in-lncRNA case, but the 394 kb link must be confirmed by fetal-brain / DIPG Hi-C. Even the conservative reading (the element regulates SOX2-OT, a positive SOX2 regulator) is pro-SOX2. Link confidence: Medium, pending Hi-C.
  • Mapping / recurrence artefact? Private, single-patient; not claimed as recurrent. The chain rests on the core-circuit biology, not recurrence.

What would kill it: (a) base-edit chr3:181317760 in an OPC-like line and CRISPRi the element; if SOX2 accessibility and mRNA do not move, dead. (b) If Hi-C shows no contact between the element and SOX2, re-route the memo to the actual Activity-by-Contact target before any wet-lab spend.

Survives as the cleanest single-variant-to-drug logic on the core circuit, at MEDIUM confidence, with the 394 kb link flagged as the thing to confirm first.

6. Go/no-go memo

  • Verdict: GO, validate first. Confidence MEDIUM.
  • Decisive experiment: base-edit the C>T base at chr3:181317760 in an OPC-like DMG line, in parallel with CRISPRi of the element; read element and SOX2-promoter accessibility, SOX2 mRNA/protein, the core-regulatory-circuit programme, and proliferation / self-renewal.
  • Kill criterion: base edit does not move SOX2 beyond noise AND CRISPRi of the element does not lower SOX2. Second kill: Hi-C shows no element-to-SOX2 contact.
  • Therapy angle (Axis A, OPC/stem core circuit): a tumour whose stem state is reinforced by this circuit is a rational candidate for the circuit's transcriptional dependency: BET (BRD4) or CDK7 inhibition, synergistic with HDAC, all validated in DIPG (Nagaraja et al., Cancer Cell 2017, PMID 28434841). Therapy confidence High for the axis, Medium that this variant marks a drug-responsive tumour. NCypher does not predict drug response; brain-penetrant BET/CDK7 chemotypes or CED are the realistic delivery routes (BBB caveat).

7. Honesty

A MEDIUM star chain: the boldest defensible single-variant story, but the variant effect is modest and the 394 kb enhancer-to-gene link needs Hi-C. What this is not: not a proven SOX2 enhancer, not a validated target, not a claim that NCypher out-predicts AlphaGenome.

Why these two, together

NPAS3 is the strongest lead by evidence (large effect, deep constraint, 80% contribution collapse, gene-level recurrence, external validation), routing to a biomarker / epigenetic axis. SOX2-OT is the cleanest mechanism-to-drug chain (a gain that reinforces the OPC core circuit, routing to clinical-stage BET/CDK7), but with a modest effect and a long-range link to confirm. Presenting both, each with its honest confidence and its kill criterion, is the NCypher shape: a ranked, mechanism-explained, druggable-linked shortlist of what to validate first, not a leaderboard number.