engin-pathway

Metabolic route manufacturability ranking — stage [3] of the engin-suite strain-to-scale funnel. Rank candidate routes to a target by predicted manufacturability (not just feasibility), with a calibrated interval, so a team spends its foundry cycles on the routes most likely to hit titer.

Why

Route-finding tools exist (novoStoic, RetroBioCat, RetroPath) and FBA predicts feasibility well — but it does not predict titer. The whitespace isn't finding routes; it's ranking them by manufacturability. A single toxic or thermodynamically-uphill step tanks a route, and step-count is blind to it.

What it does

• represents each route as a graph (networkx) of enzymatic steps with goodness features (thermodynamics, enzyme availability, cofactor balance, toxicity, expression);
• embeds it with a message-passing GCN whose max/min-pooling captures the worst step — the structural signal step-count misses;
• predicts manufacturability with a ridge head and a split-conformal interval (reusing engin_core's calibration), and ranks routes;
• is measured against the honest baseline it must beat: step-count.

Results (synthetic routes, python examples/run_demo.py)

• Manufacturability forecast (held-out): R² 0.73, RMSE 0.05, 90% coverage 0.88.
• Ranking: Spearman ρ 0.85 (graph model) vs 0.51 (step-count).
• Best-of-6-route selection: regret-vs-oracle 0.008 (model) vs 0.074 (step-count) — roughly 9× lower regret.

How much of that is a finding, and how much is the generator

Audited 2026-08-15 for issue #124, and the "beats step-count" line above says less than it first appears. Same shape as #88 in engin-materials, and the numbers are kept with the caveat rather than deleted, for the same reason.

make_dataset produces both the routes and their labels, and simulate.py:37 builds the label as

manuf = (0.6 * worst_step + 0.4 * mean_step) * 0.96 ** (length - 2)

Step-count sees only length. So its score is not a result — it is a fixed property of the generator, and no length-only heuristic can beat it. Over 2000 routes: the worst-step term carries r² 0.81 of the label, the length term 0.12, and step-count's ceiling is ρ ≈ 0.41.

The decisive check is that the margin moves with the constants. Relabel the same routes with the length base at 0.70 instead of 0.96 and step-count rises to ρ 0.95 while the worst-step signal falls to 0.31 — the ranking flips with nothing about either method changed. Reproduce with python benchmarks/generator_audit.py.

So the claim this package makes is the narrow one: the graph model does recover a worst-step signal that a length heuristic is blind to. That is a correct check that the implementation recovers a signal it ought to recover. It is not a discovery about metabolic routes.

Whether real routes are worst-step dominated is an empirical question this package has not touched. It is a design assumption of the generator — stated plainly in its docstring, and defensible, since a toxic intermediate or a thermodynamic wall does plausibly tank a route — but assumed, not measured. M1's real routes from KEGG/MetaCyc/BiGG are what would test it.

Quickstart

from engin_pathway import make_dataset, PathwayRanker, labels, spearman

data = make_dataset(500, seed=1)
ranker = PathwayRanker(lam=1.0).fit(data[:320])
ranker.calibrate(data[320:410], level=0.90)

test = data[410:]
pred = ranker.predict(test)                 # ranking scores
lo, hi = ranker.predict_interval(test)      # calibrated 90% interval
print("Spearman vs truth:", spearman(pred, labels(test)))

Status & roadmap

M0 (this package): a random-weight GCN + ridge + conformal beats step-count on synthetic routes. The weights are untrained — a deliberate stand-in.

M1 swaps in a trained GNN on PyTorch Geometric, real routes from KEGG/MetaCyc/BiGG via COBRApy/networkx, and ΔG node features from eQuilibrator. The route-as-graph interface stays the same, so the upgrade is local to embed.py. Titer-ranking validation needs wet outcomes (M3, via the suite and partners).

Install

pip install -e "packages/engin-pathway[dev]"   # from an engin-suite checkout

Apache-2.0.