fto-admet

a modular admet / dmpk screening pipeline. pass one smiles and get back one consolidated card across every endpoint: herg, lipophilicity, solubility, clearance, distribution, permeability, plasma protein binding, metabolism, toxicity, druglikeness, synthesizability, structural alerts, and a triage summary.

core is a thin, model-agnostic layer: a curated registry of models, a dispatcher that runs each model in its own isolated environment, and per-endpoint aggregators that fuse the model outputs into one verdict per endpoint. every model adapter hides its upstream code behind one uniform interface (run.py --input <path> --output <path> [--gpu N]), so core never imports a model, it shells out. adding a model is a registry entry plus a folder; adding an endpoint is a folder plus an aggregator.

setup

cp .env.example .env    # set the two storage paths (code + outputs, and envs + caches + weights)
pixi install            # solve the core environment (no gpu or model dependencies)
pip install -e .        # make `core` importable

run

one smiles, full card:

python -m core.screen --smiles "CC(C)NCC(O)COc1cccc2ccccc12" --out card.json

many molecules at once (fast: each model loads once for the whole set, not once per molecule):

python -m core.screen --input molecules.smi --out cards.json

one endpoint at a time:

python -m core.run --endpoint herg --input mol.json

--input to core.screen accepts a .smi file or an InputRecord json (object or array). each model runs in its own environment; the first use of a model installs that environment from its lockfile, then it is cached.

how it works

models are grouped by endpoint in the registry. for each endpoint the runner selects its models, dispatches each one in isolation, and passes the collected outputs to that endpoint's aggregator. the aggregators never average across incompatible scales: they harmonize onto a common quantity, surface cross-model spread as a confidence signal, and keep separate reads separate. every output reserves uncertainty and applicability-domain fields, so native signals (bayesian variance, fold-error, domain flags) are carried through rather than dropped.

tests

pixi run pytest -m "not model" -q   # fast tier: core and aggregators, no gpu
pixi run pytest -m "model" -q       # opt-in: per-model smoke tests, each in its own environment

note

the default input fixture is a placeholder smiles. swap in your target structure before a real screen; the fast test tier passes either way.