Sisyphus: A Topology-Compiled Physiologically Based Pharmacokinetic Platform with Structure-Only Input and Bayesian Parameter Refinement
Abstract
Physiologically based pharmacokinetic (PBPK) models provide a mechanistic framework for drug disposition, but compound-specific parameterization remains a practical constraint when experimental absorption, distribution, metabolism, and excretion (ADME) measurements are unavailable. Sisyphus is an open-source platform that compiles ordinary differential equations from a typed physiological graph and uses molecular structure plus dose to generate oral maximum plasma concentration (Cmax) predictions. Seven fitted models support ADME estimation and a four-track geometric blend of mechanistic, direct machine-learning, clearance-based, and distribution-based predictions. This revision reconciles the manuscript with public training snapshots and a source audit of clinical reference records. Of 107 repeatedly accessed development compounds, 73 retain evaluable reference Cmax values. The hybrid prediction has a geometric mean absolute fold error (AAFE) of 2.830 (conditional bootstrap 95% confidence interval 2.30-3.52), with 43.8% of predictions within twofold; engine-only and direct machine-learning AAFE are 3.872 and 3.049. These data do not establish independent generalization or superiority of the hybrid. The user-facing residual band is development-derived and is not valid split-conformal calibration. A selected ten-compound measured-input diagnostic lowers engine AAFE from 6.708 to 5.968, while new same-model synthetic experiments illustrate Bayesian updating with explicit effective-sample-size and interval diagnostics. Neither experiment establishes clinical accuracy. A previously reported temporal cohort has been consumed by repeated analysis, and no untouched external performance estimate is available. The platform's contribution is the integration of topology-compiled simulation, structure-driven parameterization, and experimental observation-informed refinement, together with an auditable account of data provenance and limitations. Negative development experiments identify unsuccessful tested strategies; they do not demonstrate an intrinsic structure-only accuracy ceiling or a verified bioavailability-specific cause of error.
Related
- A Dual Mechanism-Based Physiologically Based Pharmacokinetic Model for Predicting Mitoxantrone and Liposomal Mitoxantrone Disposition: Preclinical Validation in Rats and Human Extrapolation with Clinical Evaluation.
- Continuous perception and adaptive metrics for scalable structure refinement
- AutoSolveX: automated batch structure solution, initial refinement and guest screening from diffraction data
- Nanostructured formulations of abiraterone and abiraterone acetate: therapeutic potential, pharmacokinetic considerations and bioanalytical perspectives
- Computational Evaluation of Carbazole–Chalcone Urea and Thiourea Derivatives Against EGFR: Molecular Docking, Pharmacokinetic Profiling, Target Prediction, and Molecular Dynamics Simulation
BibTeX
@misc{yoon2026sisyphus,
title = {{Sisyphus: A Topology-Compiled Physiologically Based Pharmacokinetic Platform with Structure-Only Input and Bayesian Parameter Refinement}},
author = {Jae Min Yoon},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15004452/v5},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15004452/v5},
note = {CC BY},
}
RIS
TY - UNPB
TI - Sisyphus: A Topology-Compiled Physiologically Based Pharmacokinetic Platform with Structure-Only Input and Bayesian Parameter Refinement
AU - Jae Min Yoon
PY - 2026
DA - 2026/10/06
PB - ChemRxiv
DO - 10.26434/chemrxiv.15004452/v5
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15004452/v5
AB - Physiologically based pharmacokinetic (PBPK) models provide a mechanistic framework for drug disposition, but compound-specific parameterization remains a practical constraint when experimental absorption, distribution, metabolism, and excretion (ADME) measurements are unavailable. Sisyphus is an open-source platform that compiles ordinary differential equations from a typed physiological graph and uses molecular structure plus dose to generate oral maximum plasma concentration (Cmax) predictions. Seven fitted models support ADME estimation and a four-track geometric blend of mechanistic, direct machine-learning, clearance-based, and distribution-based predictions. This revision reconciles the manuscript with public training snapshots and a source audit of clinical reference records. Of 107 repeatedly accessed development compounds, 73 retain evaluable reference Cmax values. The hybrid prediction has a geometric mean absolute fold error (AAFE) of 2.830 (conditional bootstrap 95% confidence interval 2.30-3.52), with 43.8% of predictions within twofold; engine-only and direct machine-learning AAFE are 3.872 and 3.049. These data do not establish independent generalization or superiority of the hybrid. The user-facing residual band is development-derived and is not valid split-conformal calibration. A selected ten-compound measured-input diagnostic lowers engine AAFE from 6.708 to 5.968, while new same-model synthetic experiments illustrate Bayesian updating with explicit effective-sample-size and interval diagnostics. Neither experiment establishes clinical accuracy. A previously reported temporal cohort has been consumed by repeated analysis, and no untouched external performance estimate is available. The platform's contribution is the integration of topology-compiled simulation, structure-driven parameterization, and experimental observation-informed refinement, together with an auditable account of data provenance and limitations. Negative development experiments identify unsuccessful tested strategies; they do not demonstrate an intrinsic structure-only accuracy ceiling or a verified bioavailability-specific cause of error.
ER -