Structure-only Cmax.
Evidence stated plainly.
Sisyphus predicts single-dose oral Cmax from a parent-drug SMILES and dose. A separate graph-based PBPK solve provides mechanistic research outputs; those endpoints are not presented as if they came from the Cmax ensemble.
Three ideas that define Sisyphus
The design choices that make it extensible and honest, not just another PBPK script.
The body is a graph
Organs are nodes; vessels, transit, and clearance are typed edges. The ODE system is derived from graph topology — you extend the model by editing YAML, never the engine.
Intervals keep their meaning
Parameter Monte Carlo and empirical model-residual bands are separate outputs. The current residual band is development evidence, not an independent conformal guarantee.
The engine knows types, not identities
No organ or drug names live in the engine. Identity lives in data, so new organs, enzymes, and routes never require an engine change.
One production question, one evidence view
The public console is intentionally narrow: oral Cmax prediction and transparent development-benchmark inspection. Open the console →
Parent SMILES + oral dose → Cmax, component tracks, applicability flags, and explicitly labeled interval sources.
N=73 retrospective development evidence — predicted vs observed by track, never labeled as an independent holdout.
Experimental Python and CLI workflows include engine-as-prior MIPD with forward dose search (sisyphus mipd), time-dependent DDI scenarios (sisyphus ddi), and effect-site PK/PD scenarios (sisyphus pkpd). The graph supports opt-in enterohepatic recirculation overlays, a reference_woman physiology preset, and CYP2C19, CYP3A5, and CYP2B6 phenotypes. Tissue partitioning now uses full Rodgers–Rowland Kp by default, with a Berezhkovskiy option.
Retrospective development evidence
The source-audited N=73 benchmark is compound-disjoint from fitted-model training, but it was repeatedly used for system selection and calibration. Meta AAFE has a conditional bootstrap 95% CI of 2.12–3.09. This is development data, not an independent holdout.
Explore the current N=73 predicted-vs-observed scatter by track in the console.
View development benchmark →| Track | AAFE | within 2-fold |
|---|---|---|
| Meta · development | 2.543 | 45.2% |
| Engine only | 4.259 | 34.2% |
| ML only | 2.786 | 42.5% |
| Meta, structurally in-scope | 2.463 | 42.4% |
A four-track meta-learner blends a mechanistic engine, a data-driven ML Cmax, a CL/F analytical, and a conditional VDss track. Prospective FDA NMEs (2024–25) are harder still at AAFE 3.29 — generalization, stated plainly.
Or run it from the command line
The console is one face of a Python library + CLI.