engin-host

Host / chassis selection — stage [4] of the engin-suite strain-to-scale funnel. Given a target molecule's capability profile, recommend the microbial host with an explainable rationale, an honest confidence band, and hard-constraint flags — so teams don't spend 18 months discovering E. coli was the wrong chassis.

Why

Host selection is a crucial design parameter still chosen empirically ("just use E. coli" or tribal knowledge). No published tool scores candidate chassis against a production requirement with calibrated uncertainty — searched 2026-08-13 and written up, with the near-misses, in docs/design/host-selection.md. The honest competitor is not other software: it is building the constructs in six hosts and measuring. Circuit failures are frequently host-interaction failures rather than failures of the circuit itself — the contextual causes are catalogued in Cardinale & Arkin (2012).

Corrected 2026-08-13 by the D23 pass (#91), then searched under #107. This read "with no standard commercial tool — genuine whitespace": an absence claim about a market, never searched, on the day the equivalent claim about the data convention proved flatly wrong. The search found no published tool, so the absence holds — but "commercial" was dropped, because a literature and package search cannot settle what sits inside a proprietary platform. It also read "Most circuit failures are host-interaction failures"; the mechanism is well supported, the proportion is not.

What it does

A multi-criteria decision engine over a curated host-capability knowledge base:

• scores candidate hosts for a target profile (weighted suitability),
• shows why (per-capability contributions),
• flags hard requirements (glycosylation, secretion, GRAS, scale) and demotes infeasible hosts below every feasible one regardless of raw score,
• attaches a confidence band that widens honestly where the KB is thin.

Uncertainty is first-class and shared with the rest of the suite: P(suitability ≥ threshold) is computed with engin_core's primitive.

Quickstart

from engin_host import default_kb, HostQuery, score, render_memo

kb = default_kb()
q = HostQuery(
    weights=dict(glyco=1.0, secretion=0.9, protein=1.0, titer=0.6, scaleup=0.7),
    hard=dict(glyco=0.6),          # a hard glycosylation requirement
)
ranked = score(kb, q)
print(ranked[0].host, ranked[0].score, "±", round(ranked[0].band90, 2))
print(render_memo("my target", ranked))

On two contrasting queries the engine picks the right hosts — CHO for a secreted human glycoprotein, S. cerevisiae for a food-grade small molecule — correctly demoting higher-scoring-but-infeasible chassis via hard constraints. Full demo:

python examples/run_demo.py     # writes memos + score charts to outputs/

Status & roadmap

M0 (scoring + uncertainty + flags) works on an illustrative KB. Next: replace the illustrative values with literature-grounded, cited capability data (M1); calibrate the confidence band against retrospective host-choice cases, reusing engin_core's conformal machinery (M2); backtest and sign design partners (M3).

Install

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

Apache-2.0.