engin-protein

One calibrated engine, three faces of the protein design cycle. Plans 2, 5, and 9 from the venture shortlist are not three companies — they are the same GP + Expected-Improvement + conformal-interval loop pointed at a fitness landscape instead of fermentation knobs.

The wedge

Protein campaigns share the shape the rest of the engin suite serves: few measurements, expensive assays, irreversible commitments. A model that is 5% more accurate doesn't change what you order next. A model that says "82% chance this variant clears your threshold, here's the batch that most reduces your uncertainty, and here's what's driving it" does.

Built as one package because they share a featurizer, an engine, and an uncertainty vocabulary. Splitting them would mean maintaining three copies of the same calibration story.

Use it

from engin_protein import Campaign, LowNCopilot, make_landscape

landscape = make_landscape(length=20, epistasis=0.5, seed=0)
campaign  = landscape.sample_campaign(n=40, seed=1)

copilot = LowNCopilot().fit(campaign)
batch   = copilot.recommend(landscape.library(200, seed=2), k=8)   # what to order next

Every prediction carries an interval or a probability. There are no naked point estimates in the public API — that's a suite-wide rule, not a stylistic preference.

Honest status: M0

This package is at M0: everything runs on a synthetic fitness landscape, not real assay data. The landscape is an additive + pairwise-epistatic model with a tunable epistasis knob, which is enough to prove the loops work and to show each face beating its baseline. It is not evidence about real proteins.

Two claims specifically that M0 cannot support, and which the plan's kill criteria turn on:

• evaluate beating pLDDT/ipTM. The baseline here is a simulated confidence signal, built to correlate with foldability but only weakly with function — which is what the literature reports about the real thing. Beating it on a landscape constructed that way tests the plumbing, not the hypothesis. The real test needs held-out wet data (M1).
• lown beating single-round design. M0 shows it on a landscape whose epistasis we chose. Real campaigns get to pick their own.

Read the numbers as "the loop is wired correctly and calibrated," nothing more.

Kill criteria (from the shortlist, verbatim in spirit)

• If function-aware ranking can't beat pLDDT/ipTM on held-out wet data, the evaluate value prop fails.
• If low-N can't beat single-round on <100 points, reposition.
• Cradle owns the funded low-N lane. lown wins only on the genuinely-low-N regime plus price.
• BayBE owns generic Bayesian optimization, open source. planner differentiates on biology-specific priors or it doesn't differentiate.

Featurization and the light path

Default is one-hot + physicochemical descriptors — no downloads, no PyTorch, runs anywhere. That keeps the suite's light-default-path rule (ADR 0002) intact.

For real work you want PLM embeddings (ESM-class) or structure-derived features. Rather than pulling that dependency in, PrecomputedFeaturizer accepts an embedding matrix you computed however you like. The engine never learns which featurizer produced its inputs.

Milestones

• M0 (here) — synthetic landscape with tunable epistasis; all three faces beat their baselines, calibrated coverage asserted in tests.
• M1 — real ProteinGym / public DMS sets, PINDER, Adaptyv open competition results. Report wet-correlation where public wet data exists. This is where the kill criteria actually get tested.

License

Apache-2.0.