01 — Auracle™ Discovery Platform
Novel therapeutic targets, directly from patient data
Auracle is a deep learning platform that discovers novel drug targets directly from patient transcriptomics, ranking candidates by disease biology, not by prior druggability or target precedent.
02 — The Problem
Most target discovery platforms recycle what's already known
When computational models are trained on features like druggability scores and tractability indexes, they learn to predict what has already been targeted, sometimes missing what matters biologically. The output reflects the literature, not the disease.
Auracle removes that circularity entirely. Candidates are ranked by their position in patient-derived disease networks, scored by the biology. The result: targets that conventional approaches structurally miss.
03 — How It Works
We use patient transcriptomics to rank therapeutic hypotheses
Auracle builds patient-specific disease networks from transcriptomic data, then applies proprietary deep learning to score every gene by its topological importance instead of whether it has ever been drugged before.
Input
Patient Data
Gene expression from case-control cohorts. No proprietary formats required.
Model
Disease Networks
Patient-derived molecular interaction networks encoding disease-specific topology.
Engine
Auracle™PATENT PENDING
Proprietary deep learning trained on topology with built-in mechanistic interpretability.
Output
Ranked Candidates
Probability-scored therapeutic hypotheses with per-target mechanistic explanations.
Input
Patient Data
Gene expression from case-control cohorts. No proprietary formats required.
Model
Disease Networks
Patient-derived molecular interaction networks encoding disease-specific topology.
Engine
Auracle™PATENT PENDING
Proprietary deep learning trained on topology with built-in mechanistic interpretability.
Output
Ranked Candidates
Probability-scored therapeutic hypotheses with per-target mechanistic explanations.
04 — What Makes It Different
Built to find what hasn't been found
No druggability bias
Druggability and tractability scores are excluded from features and ranking. Candidates are scored by disease biology, not by what's convenient to target.
Cross-cohort generalization
Trained on one patient population, validated on completely independent cohorts without retraining. The model learns disease topology, not dataset artifacts.
Mechanistic interpretability
Every candidate comes with attention-derived explanations of which network interactions drove its ranking, not a black box.
05 — Validation
Tested blind against clinically validated targets
Held-out clinical targets the model never saw during training, evaluated against over 14,000 background genes across 201 patients from two independent cohorts.
0.97
Hold-out AUROC
201
Patients · 2 cohorts
14.8K
Background genes
Blind evaluation on held-out clinical targets vs. 14,819 background genes. Rankings are driven by disease network topology, not simple connectivity (degree correlation r = 0.056).
Key findings replicated in a fully independent patient cohort (n = 161) with no shared data, no shared graph, and no retraining.
06 — Work With Us
You have the data. We'll find the targets.
We partner with pharma, biotech, and research teams who have transcriptomic data and want therapeutic hypotheses that haven't been generated before.
Pilot engagements start with a single disease cohort. We deliver ranked candidates with mechanistic explanations in weeks, not months.