Skip to content

NeoRx

Correlation is not causation. NeoRx is built around telling the difference.

NeoRx is a drug-target discovery platform that asks an interventional question of a biomedical knowledge graph: if we inhibited this protein, would the disease outcome change? Association databases answer a different question — whether the gene and the disease occur together — and the two answers routinely disagree.

It is one Python package with several engines: graph assembly from biomedical sources, causal identification over that graph, molecule generation, drug-likeness screening, docking, a reinforcement-learning mode, and an experiment runner that refuses to report a number without a run record.

Why the distinction matters

Association databases report that TNF is strongly associated with HIV, and they are right: TNF-α is elevated during infection. But that elevation is downstream of immune activation. Inhibiting TNF does not treat HIV — it removes an immune response the patient needs.

CCR5 is the HIV-1 co-receptor. The CCR5-Δ32 loss-of-function variant confers near-complete resistance, and maraviroc, which blocks CCR5, is an approved antiretroviral. CCR5 is a target you can intervene on.

Both genes sit high in the same association tables. What separates them is not the strength of the association but its direction — and nothing in an association score carries direction.

NeoRx's answer is to refuse to treat the two kinds of claim alike:

>>> from causalphylo import is_causal_admissible
>>> is_causal_admissible({"edge_type": "associated_with",
...                       "evidence_class": "genetic_association"})
True
>>> is_causal_admissible({"edge_type": "associated_with",
...                       "evidence_class": "literature"})
False

A germline variant is randomised at conception, so a genetic association carries a natural-experiment warrant a literature co-mention does not. That warrant is what earns an arrow. Everything the causal engine does runs on the subgraph of edges that have one — see Which edges earn an arrow.

The shape of a run

  1. Build the graph. Query the biomedical sources for a disease, merge duplicate entities to gene symbols, and record the provenance of every edge. See The graph and its sources.
  2. Identify targets. Restrict to the causal-admissible subgraph, then apply Pearl's backdoor criterion per candidate, reporting why identification succeeded or failed. See Identification.
  3. Rank and classify. Combine identifiability with evidence counts, druggability and a tissue-expression gate into a causal confidence and a causal / correlational / inconclusive label. See From verdict to target.
  4. Design molecules. Sample the GenMol VAE for each surviving target, screen with MolScreen, dock with DockBot.
  5. Score. A six-dimension composite in which causal confidence carries the largest weight. See Ranking candidates.
  6. Report. An HTML report with the graph, the target reasoning, and the ranked candidate table.

Steps 1–6 are neorx run <disease>, or run_pipeline() in Python. Each is also reachable on its own.

Thirty seconds

Requires network

build_disease_graph queries the live biomedical APIs, so this block is not executed as a doctest and no output is shown for it — what comes back depends on what those sources say today.

from neorx import build_disease_graph, identify_causal_targets

graph = build_disease_graph("HIV")
for target in identify_causal_targets(graph, top_n=5):
    print(target.gene_name, target.classification.value)

For a version you can run offline and check line by line, see A worked example.

Where the causal engine lives

The graph-only part of this work — admissibility and the backdoor criterion, with no biomedical model of its own — has been extracted into a standalone package, CausalPhylo. If you want identification over your own graph rather than drug discovery, that is the smaller thing to depend on: one runtime dependency, NetworkX.

NeoRx depends on it rather than carrying a second copy, so is_causal_admissible above is CausalPhylo's. The division is along the dependency graph, not by subject: everything that needs only a graph went, and everything written against NeoRx's own data model — evidence counting, scoring, ranking, target classification — stayed.

Reproducibility is not a feature of this project

It is a constraint on it. A manuscript audit of this codebase found that no reported number could be traced to an execution artifact: a benchmark that persisted nothing, figures holding F₁ values as literals, a table entry retyped by hand from lost terminal output. The experiment runner exists so that an experiment has no code path by which it reports a result without writing a record, and the snapshot layer exists so that a run over a past release reads that release rather than today's.

Author and licence

Kelyn Paul Njeri · NeoForge Labs · ORCID 0009-0000-1068-4512

Source available, non-commercial. See Licence.