Hydrating the disease, compound, gene, and metaedge catalogs.
○ loading
01 · INITIALIZE
Define the investigation
Set what you're investigating and how it should run. The choices on this page govern every downstream evidence claim — the page exists to make those choices explicit before any number appears.
● ready
TOOL · INVESTIGATION PARAMETERS
What you're investigating
Form
The disease, compound, anchor target, and Hetionet relation. These four selections define the prediction task and what evidence will count.
▼
▼
▼
▼
10
DISEASES
8
COMPOUNDS
7
GENES
4
METAEDGES
TOOL · RECOMMENDATION ENGINE
Recommended combination for this investigation
Pick a starting point · 0%
Based on the current selections, this bundle is the most coherent disease-compound-gene-metaedge combination to investigate next.
Disease
Hypertension-attributed ESKD
Compound
Inaxaplin
Anchor gene
APOL1
Metaedge
CtD · Compound–treats–Disease
WHY THIS COMBINATION
APOL1 ancestry-aware kidney signal: compound mechanism, anchor gene, and disease context all point at the same causal axis.
Recommended next fields: Disease, Compound, Anchor gene, Metaedge.
Use Hybrid (Aer kernel) or Quantum HW (IBM when Settings has token+CRN, else Aer) to compare the quantum-kernel headline against classical baselines.
TOOL · CANDIDATE CONTEXT
— choose a compound —
Reference
Why this compound is on the table — mechanism, anchor target, and clinical evidence backdrop. Switch the candidate to refresh.
DESCRIPTION
Select a compound in the investigation parameters to populate mechanism, anchor target, and clinical evidence backdrop.
ANCHOR & TASK
diseasenot set
metaedgenot set
run pathhybrid
TOOL · RUN PATH
How the investigation runs
Selector
Pick a preset bundle below. The API runs one headline trainer per family (classical ensemble vs quantum kernel SVC) plus a fixed leaderboard of comparison rows — many catalog names are methodological context; non-headline rows are marked SIM on Experiment. Classical baselines always accompany hybrid / quantum paths. For drill-down metadata, use the Algorithm Catalog below.
▢
Classical
Stacking · GBDT · KGE
8 algorithms~50s · CPU
straightforward baseline run
Recommended generalist
Stacking
heterogeneous ensemble meta-learner
⌥
Hybrid
QK-SVC (Aer) + baselines
3 + 8 baselines~2m · Aer
kernel SVC · local simulator
Recommended generalist
Quantum Kernel + Metapath
metapath features → quantum kernel → SVM
⊗
Quantum HW
QK-SVC + baselines · IBM optional
2 + 8 baselines~4m · Aer or IBM
IBM when creds set; else Aer
Recommended generalist
QK-SVC (hardware path)
Cross-validated precomputed kernel SVC — backend and shots reflect Aer or IBM from the Settings you provide.
ADVANCED · DRILL-DOWN
ALGORITHM CATALOG— inspect algorithms wired in this port12 of 12 algorithms across 7 groups
alllivedevrunning now
▶Quantum kernels3of 3
QSVC (Pauli)●
Pauli ZZ feature map → quantum kernel → classical SVM
Degree-Weighted Path Count over Hetionet metapaths
classicalnp~10slive
loading knowledge-graph preview…Fetching Three.js (~first visit). Panels fill once the viewer chunk arrives.
TOOL · EVIDENCE POSTURE
Integrity guards
Reference · seed
Every switch that governs an evidence claim downstream. State is inspectable — running with any critical guard off produces a different kind of result and must be visible to anyone reading a report.
21/23
guards enabled
9/9
critical on
0
critical off
PASSING
posture status
allcriticalrecommendedoptionaloff
▶Bias / equity
4 on0 off
Ancestry-balanced negativescritical
Hard negatives sampled per ancestry stratum
Failing critical guards halt the audit pipeline.
config/integrity.yamlon
Equity caveat surfacedcritical
Compound equity caveat shown when present
Failing critical guards halt the audit pipeline.
config/integrity.yamlon
Hard negatives includedcritical
Negatives drawn from co-treated diseases
Failing critical guards halt the audit pipeline.
config/integrity.yamlon
Hard-negative ratiorecommended
≥1:5 hard:positive negatives
Recommended guards surface a warning when off.
config/integrity.yamlon
▶Data quality
4 on0 off
Anchor-target leakage checkcritical
Anchor gene held out from training metapaths
Failing critical guards halt the audit pipeline.
config/integrity.yamlon
Feature leakage checkcritical
No metapath includes the held-out edge
Failing critical guards halt the audit pipeline.
config/integrity.yamlon
Time-split validationrecommended
Train < cutoff < test by curation date
Recommended guards surface a warning when off.
config/integrity.yamlon
Provenance hashedrecommended
Inputs SHA-256 hashed and logged
Recommended guards surface a warning when off.
config/integrity.yamlon
▶Statistical rigor
5 on0 off
Bootstrap 95% CIrecommended
N=1000 bootstrap on metric of interest
Recommended guards surface a warning when off.
config/integrity.yamlon
Multi-seed stabilityrecommended
≥3 seeds per configuration; report std
Recommended guards surface a warning when off.
config/integrity.yamlon
Calibration checkcritical
Reliability diagram + ECE/MCE
Failing critical guards halt the audit pipeline.
config/integrity.yamlon
DWPC baseline paritycritical
Top model must beat DWPC by margin
Failing critical guards halt the audit pipeline.
config/integrity.yamlon
Random baseline paritycritical
Top model must beat random by p<0.05
Failing critical guards halt the audit pipeline.
config/integrity.yamlon
▶Reproducibility
2 on1 off
Reviewer-blind rankingcritical
Top-K ranking computed without label peek
Failing critical guards halt the audit pipeline.
config/integrity.yamlon
Unit tests greenrecommended
`pytest` and `vitest` green on this branch
Recommended guards surface a warning when off.
config/integrity.yamlon
Leave-one-disease-outoptional
Each fold withholds one disease entirely
Optional — informational; off does not gate audit.
config/integrity.yamloff
▶Quantum integrity
4 on0 off
ZNE error mitigationrecommended
Zero-noise extrapolation on quantum runs
Recommended guards surface a warning when off.
config/integrity.yamlon
Readout error mitigationrecommended
M3 readout-error mitigation
Recommended guards surface a warning when off.
config/integrity.yamlon
Shot-budget checkrecommended
Total shots ≤ tier allocation
Recommended guards surface a warning when off.
config/integrity.yamlon
Kernel spread sanityrecommended
Kernel σ within plausible range
Recommended guards surface a warning when off.
config/integrity.yamlon
▶Hetionet integrity
2 on1 off
Degree correctionrecommended
DWPC-style hub penalty applied
Recommended guards surface a warning when off.
config/integrity.yamlon
Path-length caprecommended
Metapaths capped at length 4
Recommended guards surface a warning when off.
config/integrity.yamlon
Leave-one-compound-outoptional
Each fold withholds one compound entirely
Optional — informational; off does not gate audit.
config/integrity.yamloff
TOOL · SESSION
Save or resume
Action
Investigations are reusable artifacts. Save a snapshot of the four parameters + run path, then restore it later. State persists in the browser via localStorage. Snapshots are not the same as completed job links — after a run finishes, use the handoff buttons or ?jobId= on Experiment to reopen a specific result.
CURRENT INVESTIGATION (UNSAVED)
—— → —
metaedgehybrid
RECENT (0)
No saved sessions yet. Tap "Save snapshot" to start a history.
RUN · FASTAPI
Execute investigation job
API
Posts to /investigations/run on the FastAPI service and polls /jobs/<id> until the job completes.
⊙ HOW TO READ THIS PAGE
What this view answers — and what to question
The Initialize page is six tools stacked into a contract. Read top-to-bottom: investigation parameters define what you're asking, candidate context says why this compound is worth asking about, run path selects the headline trainer family (vs roster rows marked SIM on Experiment), live KG preview shows what the model will see before compute, evidence posture guarantees the integrity guards are on, and session persists the snapshot so the contract is reproducible.