🧬 NeoRx Report

Executive Summary

Causal Knowledge Graph

Interactive graph — drag to pan, scroll to zoom, click nodes for details.

Causal Target Analysis

Targets evaluated using Pearl's do-calculus via the backdoor criterion, effect estimation, and sensitivity analysis.

{% for t in targets.entries %} {% if t.reasoning %} {% endif %} {% endfor %}

Gene	Protein	Classification	Causal Confidence	95% CI	Robustness	Druggability	Pathways	PPIs
{{ t.gene_name }}	{{ t.protein_name }}	{% if t.is_causal %} CAUSAL {% elif t.classification == 'correlational' %} CORRELATIONAL {% else %} INCONCLUSIVE {% endif %}	{{ "%.3f"|format(t.causal_confidence) }}	[{{ "%.2f"|format(t.confidence_interval.0) }}, {{ "%.2f"|format(t.confidence_interval.1) }}]	{{ "%.3f"|format(t.robustness) }}	{{ "%.3f"|format(t.druggability) }}	{{ t.n_pathways }}	{{ t.n_interactions }}
{{ t.reasoning }}

Top Drug Candidates

Ranked by composite score: causal confidence (0.30) + binding (0.25) + QED (0.15) + SA (0.10) + ADMET (0.10) + novelty (0.10).

{% for c in candidates.entries %} {% endfor %}

#	SMILES	Target	Composite	Causal	Binding	QED	MW	Drug-Like	Novel
{{ c.rank }}	{{ c.smiles[:45] }}{% if c.smiles|length > 45 %}…{% endif %}	{{ c.target }}	{{ "%.4f"|format(c.composite) }}	{{ "%.3f"|format(c.causal_conf) }}	{% if c.binding %}{{ "%.1f"|format(c.binding) }}{% else %}N/A{% endif %}	{% if c.qed %}{{ "%.2f"|format(c.qed) }}{% else %}N/A{% endif %}	{% if c.mw %}{{ "%.0f"|format(c.mw) }}{% else %}—{% endif %}	{% if c.is_drug_like %}✓{% else %}✗{% endif %}	{% if c.is_novel %}✓{% else %}✗{% endif %}

Methodology

NeoRx applies Pearl's do-calculus to distinguish genuine causal drug targets from correlational bystanders. Unlike conventional pipelines that rank by association scores alone, NeoRx asks: "If we intervene on this protein, does it causally affect the disease outcome?"

1. Knowledge Graph Assembly — Disease causal graph built from 7 biomedical databases (Monarch Initiative, Open Targets, KEGG, Reactome, STRING, UniProt, PDB).
2. Causal Identification — Backdoor criterion applied via d-separation to test identifiability of causal effects.
3. Effect Estimation — Multi-source evidence triangulation (path strength × d-separation quality × centrality × source corroboration).
4. Sensitivity Analysis — Leave-one-source-out analysis with bootstrap 95% confidence intervals (200 resamples).
5. Molecule Generation — Novel candidates generated by GenMol (Variational Autoencoder trained on ChEMBL).
6. Drug Screening — Candidates screened for drug-likeness (MolScreen), multi-rule ADMET, and binding (DockBot).
7. Composite Scoring — Six-dimensional ranking with causal confidence weighted highest (0.30).