Sisyphus

Oral structure + dose → Cmax, with a separate graph-based PBPK research engine

Methodology · Quickstart · Validation · Architecture · Limitations

Preprint: Yoon, J. M. (2026). Sisyphus: A Topology-Compiled Physiologically Based Pharmacokinetic Platform with Structure-Only Input and Bayesian Parameter Refinement. ChemRxiv. — DOI 10.26434/chemrxiv.15004452/v1

The published v1 preprint and archived Sisyphus_Preprint.pdf report an older AAFE of 2.698. The current code's source-screened development benchmark is 2.830 on 73 scored compounds; the value comes from a different model and reference set. Use the validation section below for current evidence.

Sisyphus is an oral structure-only C_max prediction system with a separate physiologically based pharmacokinetic (PBPK) research engine. The engine represents the human body as a typed directed multi-graph and derives ordinary differential equation (ODE) systems from graph topology.

The production output is C_max for a canonical parent SMILES and positive oral dose. Engine-derived T_max, AUC, half-life, multi-dose simulation, TDM, MIPD, DDI, PGx, and PK/PD are experimental research outputs and are not covered by the C_max accuracy claim. Residual/model-error and parameter-Monte-Carlo intervals are exposed separately.

Intended use. Sisyphus targets oral structure-only Cmax prediction (canonical parent SMILES + parent-active-moiety dose) when measured ADME is unavailable. On the repeatedly accessed retrospective development benchmark, Meta AAFE is 2.830 [bootstrap 95% CI 2.30–3.52, N=73 scored]. This is not an independent holdout result: the original 107-compound cohort has informed repeated system-selection decisions. A source-adjudicated diagnostic P0 pilot on 18 compounds, scored with an earlier model, found Meta AAFE 3.34, compared with 3.01 for direct ML; its labels were AI-assisted, historical VDss training membership is unverified, and it does not establish Meta superiority. The current system has no unconsumed independently curated external holdout AAFE. Error of this scale and the wide development-residual interval (÷×~10.65-fold; 90.4% coverage on the repeatedly used development set) restrict the tool to screening, ranking, and uncertainty-aware triage, not dose setting.

For a 500 mg metformin hydrochloride tablet, pass 389.93 mg with the parent metformin SMILES; the product label states that active-moiety equivalence.

$ sisyphus predict --smiles "Cn1c(=O)c2c(ncn2C)n(C)c1=O" --dose 100

Drug: Cn1c(=O)c2c(ncn2C)n(C)c1=O
Method: hybrid
Execution: ok (public profile)
Applicability: no rule-based warning flags (confidence is not calibrated)
Final oral Cmax: <value> mg/L
Development empirical residual 90% interval: <low>–<high> mg/L

Methodology

Physiological model

The body is represented as a 34-compartment directed multi-graph comprising blood pools (arterial, venous, portal vein), perfusion-limited organs (11), permeability-limited organs (4, each split into vascular and extravascular sub-compartments), GI lumen segments (8, compartmental absorption and transit model; Yu & Amidon, 1999), and mass-balance sinks (4). Physiological parameters follow the ICRP Reference Man (ICRP, 2002). Tissue compositions for partition coefficient estimation are taken from Rodgers & Rowland (2005). CYP enzyme abundances follow Shimada et al. (1994).

                      ┌─────────────────────────────────────────────┐
                      │                                             │
   ┌──────┐    ┌──────┴───┐                                   ┌─────┴────┐
   │ lung │───►│ arterial │─► brain ─────────────────────────►│  venous  │
   └──┬───┘    │  blood   │─► heart ─────────────────────────►│  blood   │
      │        │          │─► kidney ────────────────────────►│          │
      │        │          │                                   │          │
      │        │          │─► gut wall ──┐                    │          │
      │        │          │─► spleen  ───┼─► liver ──────────►│          │
      │        │          │─► pancreas ──┘   (portal,CYP450)  │          │
      │        │          │                                   │          │
      │        │          │─► muscle, adipose, skin, bone ───►│          │
      │        └──────────┘                                   └─────┬────┘
      │                                                             │
      └─────────────────────────────────────────────────────────────┘

   stomach ──► duodenum ──► jejunum ──► ileum ──► colon ──► fecal excretion
                  │            │          │
                  └────────────┴──────────┘
                       absorption ──► gut wall

ODE formulation

The ODE system is derived automatically from graph topology. Each edge type dispatches a flux function:

Perfusion-limited transport (FlowFluxSpec):

$$\frac{dA_i}{dt} = Q_i \cdot C_{in} - Q_i \cdot \frac{A_i \cdot R_{B:P}}{V_i \cdot K_{p,i}}$$

Hepatic clearance — Extended Clearance Model (ECM, default) (ClearanceFluxSpec, model="extended"):

The QSSA-derived effective intrinsic clearance, with $PS_{inf,total} = PS_{inf} + J_{max} \cdot f_u / (K_m + f_u \cdot C_{u,hep})$:

$$CL_{int,eff} = \frac{PS_{inf,total} \cdot CL_{int,h}}{PS_{eff} + CL_{int,h}}$$

is embedded in the well-stirred form:

$$CL_{h} = \frac{Q \cdot f_u \cdot CL_{int,eff}}{Q + f_u \cdot CL_{int,eff}}$$

where $PS_{inf}$ is passive sinusoidal influx, $PS_{eff}$ is sinusoidal efflux, $J_{max}/K_m$ are active uptake (OATP1B1, …) Michaelis–Menten parameters, and $CL_{int,h}$ is the sum of metabolic + biliary intrinsic clearance. Derivation: hepatocyte mass balance $PS_{inf,total} \cdot C_{u,blood} = (PS_{eff} + CL_{int,h}) \cdot C_{u,hep}$ at steady state (QSSA). For drugs without active transporter kinetics, the model reduces to the classical well-stirred form:

$$CL_{int,organ} = \sum_j \left( E_j \cdot k_j \right) \cdot S_{IVIVE}$$

$$CL_{organ} = \frac{Q \cdot f_u \cdot CL_{int,organ}}{Q + f_u \cdot CL_{int,organ}}$$

where $E_j$ is the enzyme abundance (total pmol in organ), $k_j$ is the per-enzyme intrinsic clearance (μL/min/pmol), and $S_{IVIVE}$ converts units (60/10⁶, μL/min → L/h). This formulation is enzyme-level: clearance at any organ is computed from its local enzyme profile, not from organ identity (Houston, 1994; Yang et al., 2007; Shitara et al., 2013 / Yoshikado et al., 2017 for the ECM/QSSA hepatocyte form implemented here).

Permeability-limited distribution (DiffusionFluxSpec):

$$\frac{dA_{vasc}}{dt} = Q \cdot C_{art} - Q \cdot C_{vasc} - PS \cdot (C_{u,vasc} - C_{u,tissue})$$

$$\frac{dA_{tissue}}{dt} = PS \cdot (C_{u,vasc} - C_{u,tissue})$$

GI absorption (AbsorptionFluxSpec):

$$k_a = \frac{2.88 \cdot P_{eff} \cdot f_{ka}}{r}$$

where $P_{eff}$ is effective permeability (×10⁻⁴ cm/s), $f_{ka}$ is the segment-specific absorption fraction, and $r$ is particle radius (μm). The coefficient 2.88 is an empirical absorption-scale parameter for these numerical input units; it carries the units needed to produce h⁻¹ and is not a dimensionless cm/s-to-μm/h conversion. Its transferability across formulations has not been independently established.

Tissue:plasma partition coefficients are computed via the Rodgers & Rowland method (Rodgers & Rowland, 2005, 2006), with the Berezhkovskiy correction for acids (Berezhkovskiy, 2004).

Prediction pipeline

The full pipeline combines mechanistic simulation with data-driven prediction:

SMILES → molecular profile: RDKit descriptors, structural pK_a classification, applicability domain assessment
ADME prediction: Pre-trained XGBoost models for f_u,p, P_eff, CL_int, and VD_ss (public TDC datasets; Huang et al., 2021); R_B:P defaults to 1.0. DrugBank experimental enrichment is available only in the opt-in licensed_research profile and never affects public results
IVIVE: CL_int decomposition into per-enzyme affinities, Kp calculation
PBPK simulation: 34-state ODE system solved via LSODA (Petzold, 1983)
ML direct prediction: XGBoost C_max model (906 hash-pinned public clinical PK rows derived from Omega MMPK data)
CL/F analytical track: closed-form 1-compartment C_max estimate using XGBoost CL/F + V_d predictions and k_a from Engine T_max / Peff. Decorrelates with Engine+ML residuals via different input channels.
VDss volume proxy: dose divided by predicted VDss (volume-of-distribution-at-steady-state) for a fixed 70 kg body weight. This is a simple scale estimate, not an absorption/elimination C_max model. It is included whenever a positive VDss estimate is available; this routing was selected on N=107 and is not an independently validated applicability rule.
Meta-learner: Compound-type-adaptive geometric blend of all four tracks with weights selected by LOOCV on the repeatedly accessed N=107 development set. Base compounds: engine 0.60 / ML 0.40 / CLF 0.00; non-base: engine 0.35 / ML 0.50 / CLF 0.15. VDss track weight 0.20 when activated; other weights scaled by ×0.80 so the four-track sum remains unity. These weights are frozen pending blinded external evaluation.

Uncertainty propagation

All parameters are represented as Distribution(mean, cv, dist_type). Monte Carlo sampling draws N realizations from all parameter distributions simultaneously, solves the ODE for each, and aggregates the resulting PK endpoints into distributional summaries with prediction intervals. The graph topology is compiled once; only parameter values change across MC iterations ("compile once, parameterize many").

Multi-dose regimen simulation

Multi-dose pharmacokinetics are computed by an event-driven solver that wraps the single-dose ODE engine. Dose events are injected into the state vector between integration segments; the ODE right-hand side is not modified. This preserves the identity-blind engine invariant while supporting arbitrary dosing schedules (repeated oral, IV infusion, mixed regimens).

Steady-state detection applies a trough variation criterion (<5%) across the last three dosing intervals. The solver reports C_ss,max, C_ss,min, accumulation ratio (AR = C_ss,max / C_max,first), and dose number at which steady state is reached.

Therapeutic drug monitoring

Given observed plasma concentrations, Sisyphus refines population-level parameter distributions into individual posteriors using a dispatched Bayesian router (data/sbi/method_routing.json) that selects one of three methods per drug:

SBI (Simulation-Based Inference, default, 12/13 production drugs). Amortized neural posterior estimation (Normalizing Spline Flow) conditioned on (simulator output, patient observation); the per-drug posterior is pre-trained offline. Inference is a single forward pass (milliseconds). Used when the Simulation-Based Calibration (SBC) gate passes on the offline validation profile.
IBIS (Iterative Bayesian Importance Sampling, 1/13 production drugs). Used as fallback when SBC fails for a drug (e.g. pravastatin OATP1B1 issues pre-ECM); closed-loop iteration prevents weight degeneracy.
IS (classical Importance Sampling, 0/13 production drugs post-routing). Retained for legacy compatibility; used only for compounds where SBI training data is insufficient AND IBIS has not been validated.

Effective sample size (ESS = $1/\sum w_i^2$) is monitored for importance-based methods. Morphine routes SBI with a likelihood reweighting kernel (P6, 2026-04-19) to compensate for hierarchical variance deficiency. All three methods share a unified output contract: per-PK-parameter posterior Distribution. This mechanism can update errors in the current CL_int predictor (scaffold-CV R² = 0.215): observed drug levels update population priors, reducing posterior CV by >50% in internal validation (see TDM validation).

Model-informed precision dosing

MIPD recommends adjusted doses to achieve a target steady-state concentration:

Bayesian update at the observed dose yields a posterior C_ss distribution.
Linear dose scaling: $dose_{new} = dose_{current} \times (C_{ss,target} / C_{ss,posterior})$
Clamp to clinical dose range and round to a practical increment (default 25 mg).

Linear scaling assumes non-saturable metabolism. The dosing API rejects drugs with declared saturable enzyme kinetics unless the caller explicitly opts into an experimental override; production nonlinear dosing requires forward simulation across candidate doses.

Experimental engine-as-prior posterior PK (MIPD module)

The mipd/ module explores using the mechanistic engine as a structural prior updated by measured concentrations. Bioavailability F is one candidate latent because formulation, salt or crystal form, food, particle size, and transporter genetics are absent from the SMILES input. The current experiments do not show that one observation identifies F or corrects all sources of structural error.

True F can be treated experimentally as a latent with a wide prior centered on the engine's emergent F_engine, then updated from measured data by sampling/importance-resampling (SIR). This research path is not part of the supported structure-only Cmax path, and its posterior intervals have not been independently shown to be clinically calibrated.

The module provides an a-priori-to-posterior path (predict_posterior), steady-state IV trough TDM with a renal-clearance latent (predict_tdm, vancomycin/aminoglycoside scope), patient covariate individualization (creatinine clearance via measured CrCl or a Cockcroft-Gault estimate; body weight and age via a physiology-generator graph swap), and target-attainment dose recommendation (recommend_dose) over the resulting posterior.

This module is a research hypothesis. Synthetic/simulator validation does not establish real-patient predictive accuracy or clinical utility.

Experimental drug-drug interactions

DDI is modeled by adjusting enzyme abundances in the body graph prior to ODE compilation. The engine sees modified abundances and computes clearance as usual — no engine modifications are required.

Competitive inhibition:

$$E_{eff} = \frac{E_{base}}{1 + [I] / K_i}$$

Enzyme induction (E_max model):

$$E_{eff} = E_{base} \cdot \left(1 + \frac{E_{max} \cdot [I]}{EC_{50} + [I]}\right)$$

where $E_{base}$ is baseline enzyme abundance (pmol), $[I]$ is perpetrator plasma concentration, $K_i$ is the inhibition constant, and $E_{max}$/$EC_{50}$ are induction parameters. Preset perpetrators: ketoconazole, fluconazole, quinidine (CYP inhibitors) and rifampin (CYP3A4 inducer).

Experimental PK/PD link

Pharmacodynamic effects are computed from the concentration-time profile via an effect compartment with sigmoid E_max response:

Effect-site equilibration:

$$\frac{dC_e}{dt} = k_{e0} \cdot (C_p - C_e)$$

Sigmoid E_max response:

$$E = E_0 + \frac{E_{max} \cdot C_e^n}{EC_{50}^n + C_e^n}$$

where $k_{e0}$ is the effect-site equilibration rate constant (h⁻¹), $E_0$ is the baseline effect, and $n$ is the Hill coefficient. This is implemented as analytical post-processing on the PK solution, not as additional ODE states, preserving engine isolation. Preset PD models are provided for midazolam (sedation) and warfarin (INR response).

Quickstart

Installation

pip install -e ".[dev,ml]"

Pre-trained XGBoost models (f_u,p, P_eff, CL_int, VD_ss, C_max, CL/F, Vd/F) ship in models/adme/ and models/direct_pk/. Re-training scripts are provided in scripts/.

CLI

The oral predict command is the production-facing interface. The other commands below are explicitly experimental research interfaces:

# Single-dose PK prediction
sisyphus predict --smiles "Cn1c(=O)c2c(ncn2C)n(C)c1=O" --dose 100

# Multi-dose regimen simulation (atorvastatin 40 mg QD × 14 days)
sisyphus simulate --smiles "CC(C)C1=C(C(=C(N1CC[C@H](C[C@H](CC(=O)O)O)O)C2=CC=C(C=C2)F)C3=CC=CC=C3)C(=O)NC4=CC=CC=C4" \
    --dose 40 --interval 24 --doses 14

# TDM Bayesian update (midazolam 5 mg, observed 0.015 mg/L at t=1 h)
sisyphus tdm --smiles "CC1=NC=C2N1C3=C(C=C(C=C3)Cl)C(=NC2)C4=CC=CC=C4F" \
    --dose 5 --obs "1.0:0.015"

# DDI prediction (midazolam + ketoconazole inhibition)
sisyphus ddi --smiles "CC1=NC=C2N1C3=C(C=C(C=C3)Cl)C(=NC2)C4=CC=CC=C4F" \
    --dose 5 --inhibitor ketoconazole

# MIPD dose recommendation (target Css,max = 0.02 mg/L)
sisyphus dose-adjust --smiles "CC1=NC=C2N1C3=C(C=C(C=C3)Cl)C(=NC2)C4=CC=CC=C4F" \
    --dose 5 --obs "1.0:0.015" --target-css 0.02

# Holdout benchmark (add --compute-pi for empirical 90% PI coverage; diagnostic only)
sisyphus benchmark --development-set

All commands accept --verbose for debug-level logging.

Python API

from sisyphus.pipeline.predict import predict

result = predict("Cn1c(=O)c2c(ncn2C)n(C)c1=O", dose_mg=100.0)

print(result.cmax_prediction.cmax.mean)       # authoritative Meta Cmax
print(result.cmax_prediction.interval_90)     # residual/conformal interval
print(result.engine_simulation.endpoints)     # coherent engine-only endpoints
print(result.execution_status)                # "ok" / explicit fallback status
print(result.in_applicability_domain)         # rule-based flags, not calibrated accuracy

Engine-only mode (known compound parameters)

For validation or mechanistic studies, the engine can be driven directly from compound YAML files, bypassing ADME prediction:

from pathlib import Path
import numpy as np
from sisyphus.graph.builder import build_from_yaml
from sisyphus.compounds import load_compound
from sisyphus.engine.compiler import ODECompiler, ResolvedParams
from sisyphus.engine.solver import solve
from sisyphus.pk.endpoints import compute_endpoints
import sisyphus.engine.flux  # register flux functions

graph = build_from_yaml(Path("data/physiology/reference_man.yaml"))
drug = load_compound(Path("data/compounds/midazolam.yaml"))

compiled = ODECompiler().compile(graph)
rng = np.random.default_rng(42)
params = ResolvedParams(graph.sample(rng), drug.sample(rng))

y0 = np.zeros(compiled.n_states)
y0[compiled.state_index[drug.administration_node]] = drug.dose_mg

result = solve(compiled, params, y0, t_span=(0, 24))
pk = compute_endpoints(result)

print(f"Cmax: {pk.cmax.mean:.4f} mg/L")  # ~0.0028 mg/L (post FLUX-1 + RBP-2; see Validation)

Monte Carlo uncertainty

from sisyphus.engine.uncertainty import UncertaintyEngine

ue = UncertaintyEngine()
mc = ue.propagate_fast(compiled, graph, drug, n_samples=1000)

print(mc.pk.cmax)                # Distribution(mean≈0.0030, cv≈0.22)
print(mc.cmax_90ci)              # (0.0018, 0.0040) mg/L
print(len(mc.cmax_samples))      # 1000 individual realizations

Validation

Engine validation: Omega parity and post-correction snapshots

Four drugs with known compound parameters are simulated end-to-end (YAML → BodyGraph → compile → solve → C_max). They originally matched the Omega PBPK ODE engine (35-state hardcoded model) to within ~0.5% — but Omega shared two physiological bugs that Sisyphus has since corrected, so cross-engine parity is no longer the right oracle for three of the four:

FLUX-1 (2026-06-03): a flow-limitation double-count that capped hepatic/gut extraction at E→0.5. Moves high-extraction drugs (midazolam, propranolol).
RBP-2 (2026-06-04): a blood:plasma concentration-basis correction. Moves any drug with R_B:P ≠ 1 (midazolam 0.66, warfarin 0.58, propranolol 0.81).

After both fixes, only caffeine (R_B:P=1, low-extraction) remains a true cross-engine Omega-parity check. The other three are now Sisyphus self-consistency regression snapshots — their divergence from Omega is the correctness fix, not an error. Values are pinned in tests/integration/test_engine_validation.py (±5% gate; the targets carry documented macOS↔CI numerics-stack drift).

Drug	Dose	Sisyphus C_max (mg/L)	Omega C_max (mg/L)	Basis
Caffeine (Omega parity)	100 mg PO	1.6910	1.7139	Omega parity — R_B:P=1, low-extraction; invariant to both fixes (1.3% < ±5% gate, macOS-stack drift)
Midazolam (Sisyphus snapshot, post-FLUX-1/RBP-2)	2 mg PO	0.002800	0.006943	Sisyphus snapshot — FLUX-1 + RBP-2 (R_B:P 0.66)
Warfarin (Sisyphus snapshot, post-FLUX-1/RBP-2)	10 mg PO	0.343133	0.4922	Sisyphus snapshot — RBP-2 (R_B:P 0.58); FLUX-1 no-op (low-extraction)
Propranolol (Sisyphus snapshot, post-FLUX-1/RBP-2)	80 mg PO	0.059875	0.1355	Sisyphus snapshot — FLUX-1 + RBP-2 (R_B:P 0.81)

Mass balance error < 10⁻¹² for all simulations. Lesson: Omega parity is not a sufficient correctness oracle — both shared bugs survived for as long as they did precisely because parity held.

Engine-validation targets: high-extraction drugs (midazolam, propranolol) use post-FLUX-1/RBP-2 Sisyphus regression snapshots, not Omega parity — Omega shared the flow-limitation double-count bug FLUX-1 fixed, so Omega parity is no longer a correctness oracle for them. Warfarin is also a Sisyphus snapshot (RBP-2 only; FLUX-1 was a no-op for this low-extraction drug). The blood:plasma concentration basis is RBP-2 (whole-blood pools reported on a plasma basis); see docs/_internal/specs/2026-06-04-rbp-concentration-basis-design.md.

Retrospective development benchmark (SMILES → C_max)

Retrospective evaluation on a Murcko scaffold-stratified development split (107 compounds, seed=42; 73 currently have source-supported scored Cmax). All seven active fitted models now have SHA-pinned public training snapshots with compound-level exclusions, but this cohort has been used for repeated weight, track, routing, and mechanism feedback and therefore is not an independent system holdout. It integrates observed concentration–time profiles from OSP, curated literature PK data, and FDA DailyMed labels, with mixed formulations and populations. Performance is reported using AAFE with bootstrap confidence intervals conditional on the selected system; those intervals do not include adaptive model-selection bias:

$$AAFE = 10^{\operatorname{mean}\left(\left|\log_{10}\frac{C_{max,pred}}{C_{max,obs}}\right|\right)}$$

Track	AAFE	95% CI	%2-fold	%3-fold	N
Meta-learner (production)	2.830†	[2.30, 3.52]	43.8%	67.1%	73
Engine only	3.872	[2.99, 5.09]	38.4%	50.7%	73
ML only	3.049	[2.47, 3.78]	41.1%	57.5%	73
Meta, in-domain	2.898	[2.34, 3.67]	40.0%	66.7%	60

The paired compound-bootstrap Meta/ML AAFE ratio is 0.928 (95% CI 0.816–1.053, 10,000 resamples, seed 20260422). This conditional interval does not account for repeated system selection. The development data cannot establish independent Meta superiority. The in-domain ratio is 1.008 (0.898–1.137).

Reference and training-label audit (2026-09-24/25). Before the current scores were produced, every scored development arm and the direct-model training aggregates were re-checked against original sources. The audit removed 165 synthetic or arm-mixed concentration profiles; quarantined untraceable, multi-day, steady-state, prodrug-metabolite, and whole-blood labels; replaced digitized or mixed-arm peaks with directly reported single-dose fasted arms where available; and converted salt-form doses to parent active-moiety mass. Per-compound decisions, sources, and hashes are in the dated development_* notes under docs/validation/, starting from the reference source audit. A subsequent metadata-level source-consistency filter for the direct C_max training set was a no-go (3.5% higher out-of-fold AAFE; ablation), so the production model is unchanged.

Reproducibility. The table uses public-only TDC f_u,p, P_eff, hepatocyte CL_int, and Lombardo VD_ss models plus Omega-derived C_max, CL/F, and Vd/F retrains. The fitted datasets contain 1,557 f_u,p, 874 P_eff, 906 C_max, 995 CL_int, and 1,055 VD_ss hash-pinned rows; the CL/F and Vd/F tracks fit 900 and 831 rows. Per-drug predictions are in data/training/4track_holdout_predictions.json; bootstrap intervals are in data/validation/4track_ci_2026-09-24_audited_reference.json. Aggregate AAFE reproduces from a fresh clone with requirements-lock.txt; per-drug bit-identity additionally requires the same numerics stack (Python minor version, BLAS, libomp). †This repeatedly used development set and its conditional bootstrap CI do not establish independent generalization. Earlier benchmark lineage and numerics-drift measurements are in docs/research/experiment-log.md.

The 4-track meta-learner combines mechanistic PBPK (Engine), data-driven XGBoost C_max (ML), a closed-form CL/F analytical (CLF), and a conditional VDss analytical track. Weights are compound-type-adaptive and were LOOCV-selected on the original N=107 cohort: base compounds blend Engine 0.60 / ML 0.40; other compounds use Engine 0.35 / ML 0.50 / CLF 0.15, with VDss 0.20 added when applicability criteria are satisfied. The current in-domain N=60 slice is descriptive only: applicability flags have not demonstrated reliable error stratification, and neither slice is independent evidence.

Consumed temporal challenge (FDA NMEs approved 2024–2025, single-active-ingredient oral small molecules, production-clean at construction, re-scored on the 2026-07-05 engine, N=28). These are historical predictions made before the public-only fup, Peff, Cmax, CLint, and VDss retrains; the set has informed diagnosis and is no longer independent evidence for the current system:

Slice	AAFE	95% CI	%2-fold	%3-fold	N
All	3.286	[2.45, 4.48]†	25.0%	50.0%	28
In-domain	3.318	[2.12, 5.39]†	37.5%	50.0%	16

On the 2026-07-05 system, the temporal cohort was directionally worse than its same-version development benchmark (3.286 versus 2.743), reversing the earlier favorable N=15 reading. It was assembled by decontaminating previously proposed candidates and expanding FDA-NME discovery before that re-score:

Decontaminated. Three drugs that had leaked into production training were removed: vorasidenib (clinical_pk.json gold reference), and aficamten + gepotidacin (clf_training.csv → the CLF track, which has no prospective-exclusion filter). A reusable production-aware gate (scripts/check_prospective_eligibility.py) now enforces this; it additionally rejected 9 of 26 newly-discovered candidates as already-in-training (e.g. ensartinib in holdout.json['train'], deuruxolitinib in clinical_pk.json, 7 in clf_training.csv). Membership in non-production files (e.g. mmpk_expanded_*, vdss_v2_training — models the pipeline never loads) is not treated as contamination.
Expanded. Exhaustive FDA-NME discovery (101 unique 2024–2025 NMEs across 3 cross-checked sources) → 37 new oral small-molecule candidates → adversarial per-drug Cmax verification (FDA label / EMA EPAR / peer-reviewed PK, ≥2 independent sources agreeing within ~1.5×). Excluded with reasons: 4 verification failures, 7 combination products, 9 production-contaminated, 1 prodrug (sepiapterin, parent-Cmax fold ~3000 — consistent with the prior vadadustat prodrug exclusion). 16 added.

The same-machine re-score moved Meta AAFE 3.208 → 3.286 and Engine AAFE 4.302 → 4.551, while the direct ML track was bit-identical. This supports an absorption/first-pass error diagnosis but does not establish a statistically separated generalization gap. †Compound bootstrap, 100,000 resamples of absolute log-fold error, seed 20260422; diagnostic because the set is consumed. Current artifact: data/validation/prospective_N28_current_engine_2026-07-05.json.

Adaptive-selection caveat. N=107 has been used for dozens of configuration feedback cycles (track weights, routing, and meta variants), including inspection of the public-only fup, Peff, Cmax, CLint, and VDss candidates. The historical 2.85–3.10 selection-bias sensitivity range has not been recalibrated for this model. The current scored N=73 bootstrap CI ([2.30, 3.52], point estimate 2.830) does not account for adaptive search. The attempted 2026Q2 N50 was invalidated after 21/50 repository-corpus collisions and must not be cited. The replacement strategy is the outcome-blind N=260 protocol in docs/validation/external_holdout_v1_protocol.md (N=120 resource-limited fallback; its combined release gate has about 80% pass probability only near a 19% true improvement).

Historical diagnostic P0 source-adjudicated pilot

This pilot used the earlier DrugBank-trained fup artifact, so its scores do not measure the current public-only model. All 186 frozen candidates were reviewed against original PK sources before the sealed predictions were opened. Eighteen compounds (58 arms) met the source rules; 47 were excluded and 121 remain unresolved. Compound-cluster Meta AAFE was 3.34 (95% bootstrap CI 2.43–4.79), versus 3.01 (2.26–4.18) for direct ML. The paired Meta/ML ratio was 1.11 (95% CI 0.93–1.34); only 33.3% of compounds were within twofold for Meta. These results do not show Meta superiority. Source eligibility was AI-assisted with coordinator checks, and the clarification of eligibility rules followed first-pass viewing of source Cmax values. Two scored compounds occur in the current VDss fitted snapshot; the earlier model's fitted-row membership cannot be reconstructed. This is diagnostic development evidence, not independent External Holdout V1 or a clinical release gate. See the P0 result for hashes, exclusions, and remaining source uncertainty.

AI-assisted blind P1 diagnostic (earlier public model)

Eight previously unseen compounds were source-screened by a separate AI-assisted worker, and anonymous inputs/predictions were committed before their Cmax labels were opened. On this small, nonconsecutive, development-grade set, a second source review found one of the eight labels was whole-blood Cmax, not plasma. Excluding that arm, frozen-prediction Meta AAFE is 6.63 (95% compound-bootstrap CI 3.10–13.84, N=7), versus 5.25 (2.22–13.97) for direct ML; the paired Meta/ML ratio is 1.26 (0.90–1.75). These post-label diagnostic values are a serious error signal, not a population accuracy estimate or a proof of track inferiority. Immediate-release and active-moiety dose evidence is incomplete, one paper has an inconsistent capsule strength, and two independent human curators were unavailable. No P1 arm qualifies for External Holdout V1. See the P1 diagnostic for the freeze chain, source cells, sensitivity, and limitations.

AI-assisted P2 FDA acquisition check

A separate AI-assisted worker consecutively screened all 248 oral FDA NME rows from 2015–2025 against the repository exclusion union and strict original-arm criteria. No new eligible compound remained, so no predictions were run. Repository name collisions removed 235 rows; the two structure-clean parents that reached original-source arm review lacked stated post-dose meal timing. A follow-up found an explicit oral-solution arm for one parent but no matching post-dose fasting statement; both exposed identities are now excluded from a future blinded cohort. This is a single-agent acquisition check, not External Holdout V1 evidence; the P2 screen records ordered attrition and limitations. A qualifying external cohort needs other source windows and independent human curation.

Experimental multi-dose regression checks

Three drugs were simulated at clinical dosing regimens and compared against FDA-label steady-state C_max values:

Drug	Regimen	Predicted C_ss,max (mg/L)	FDA label C_ss,max (mg/L)	Fold error
Atorvastatin	40 mg QD	0.070	0.029	2.42
Metformin	500 mg BID	2.39	1.0	2.39
Warfarin	5 mg QD	0.21	1.4	0.15

(Regenerated on the current engine — post-FLUX-1 + paracellular — via scripts/verify_v2.py; supersedes the pre-FLUX-1 values.) Atorvastatin and metformin are each over-predicted by ~2.4×. The historical warfarin comparison suggested severe under-prediction, but its cited Cmax source was later quarantined for a dose/analyte mismatch, so that comparison is not current accuracy evidence. Predicted accumulation ratios tracked the theoretical values (metformin 1.11 vs 1.35; atorvastatin 1.50 vs 1.44; warfarin 1.81 vs 2.94) and steady-state detection operated correctly in all cases. This is a solver-mechanics and limitation-exposure exercise rather than a multi-dose accuracy claim: each compound’s single-dose error propagates into its steady-state estimate.

Experimental TDM regression checks

Bayesian update was tested with a single-drug functional check and a multi-drug synthetic-observation benchmark. The tables below report the Importance Sampling baseline (legacy); the experimental SBI/IS/IBIS router (data/sbi/method_routing.json) sends 12/13 configured drugs to SBI after simulation-based calibration checks. The reported SBI inference is millisecond-scale and reduced posterior CV in those tests; detailed calibration and per-drug coverage reports are tracked separately. These checks do not establish clinical TDM validity.

The morphine row and pooled summaries below are historical: they used a superseded 30 mg / 18.65 ng/mL reference. See the reference adjudication. They require a dose-matched rerun before being used as current TDM evidence.

Single-drug functional check (midazolam, 5 mg PO, one observation at t = 1 h, 10% assay CV):

Metric	Prior	Posterior
C_max CV	44.3%	19.8%
ESS	—	586.6 (29.3% of 2,000)
CV reduction	—	55.3%

Multi-drug benchmark (5 holdout drugs, synthetic patient observations scaled from engine C(t) profiles to observed C_max, 10% log-normal assay noise, seed = 42):

Drug	Type	1-obs CV reduction	2-obs CV reduction	3-obs CV reduction	1-obs ESS
Morphine	base	76.3%	77.0%	74.6%	428
Amantadine	base	74.0%	74.5%	74.7%	514
Ketorolac	acid	87.7%	92.6%	90.1%	2.8
Clozapine	neutral	68.7%	76.2%	76.9%	482
Rivaroxaban	neutral	83.8%	93.3%	98.3%	7.1

Across all 15 runs (5 drugs × 3 observation scenarios): mean CV reduction 81.2%, mean error reduction 79.8%, 90% CI coverage 67%. A single observation suffices to reduce C_max CV by 70–88% for drugs where the population prior fold error is below 2.5×. For drugs with larger prior errors (ketorolac, fold error 3.25×) or multi-observation scenarios, effective sample size degrades below 10, indicating particle degeneracy in the importance sampler. Sequential Bayesian methods (ensemble Kalman filter, particle filter) would be required for these cases.

Timepoint sensitivity analysis (morphine, single observation): t = 1.0 h (near T_max) yielded maximal CV reduction (76.3%); observations beyond 4 h post-dose provided diminishing information (CV reduction 34%). Seed sensitivity across three random seeds was 0.8%, confirming robustness at N = 2,000 prior samples.

Performance

Operation	Time	Configuration
Full prediction (SMILES → C_max)	414 ms	Prior warm-run benchmark, deterministic, single core
ODE solve (full fidelity)	106 ms	LSODA, rtol=10⁻⁸, atol=10⁻¹⁰
ODE solve (MC fast path)	33 ms	LSODA, rtol=10⁻⁴, atol=10⁻⁶
MC N=1,000	33.5 s	Pure Python RHS (no JIT compilation)
RHS evaluation	31 μs	54 flux specs per call

The prior warm-run benchmark supports interactive research screening, but first-call model loading and hardware can change latency. A local macOS/Python 3.10 caffeine check on 2026-09-23 took 1.11 s on the first call and 0.16–0.17 s on five warm calls. MC propagation at N=1,000 required ~34 s in the prior benchmark due to pure Python ODE evaluation; JIT compilation (e.g., via Numba) is an optimization path not yet pursued.

Test suite

The full test suite covers graph construction, ODE compilation, flux functions, solver correctness, mass balance, ADME prediction, the Cmax ensemble, dosing research modules, applicability flags, phenotype scaling, leakage guards, and development-benchmark reproducibility. CI output is the authoritative current count; historical pass counts are not used as a scientific validation claim.

Expected failures (3): Rosuvastatin and atorvastatin still miss their ECM-forced Cmax gates; the separate axial PGx test deliberately retains a strict expected failure because its well-stirred analytic oracle does not apply to parallel-tube extraction. Fluvastatin now passes its numerical gate, but ECM remains marked not applicable for it in production. Three prodrug clinical gates are skipped in the public clone because their conditional disposition data are absent.

Test status. The current public-only fup/Peff/Cmax/CLint/VDss benchmark is pinned by test_cached_development_aafe_is_2p830; historical benchmark changes and resolved failures are recorded in docs/research/experiment-log.md. The cached headline is reproducible with scripts/run_engine_benchmark.py on the pinned public profile.

Architecture

SMILES + dose
    │
    ▼
 predict ──► DrugOnGraph (enzyme-level, all values are Distribution)
                  │
                  ▼
             engine ◄── BodyGraph (from YAML)
             (compile graph → ODE → solve → MC propagate)
                  │
                  ▼
               pk (Cmax, AUC, t½ from SimResult)
                  │
    ml ───────────┤
    (direct PK)   │
                  ▼
             pipeline (meta-learner → final PredictionResult)
                  │
       ┌──────────┼──────────┬──────────┐
       ▼          ▼          ▼          ▼
    regimen      ddi       pkpd       mipd
   (multi-dose, (enzyme    (effect    (engine-as-prior
    TDM)        adj.)     compartment) posterior, MIPD)

Layer	Responsibility	Depends on
`graph/`	BodyGraph, node/edge types, YAML builder	`core`
`engine/`	ODE compiler, flux registry (including ECM), solver, MC	`core`, `graph`
`predict/`	SMILES → chemistry → ADME → DrugOnGraph + transporter DB	`core`
`ml/`	XGBoost C_max, CL/F, VDss predictors, 4-track meta-learner	`core`
`pk/`	SimResult → PKEndpoints (route-aware)	`core`
`regimen/`	Multi-dose solver, TDM method dispatch (SBI/IS/IBIS/EnKF), linear-scaling dose adjust	`core`, `engine`, `graph`, `sbi`
`sbi/`	Simulation-based inference training + amortized posterior, physiology generator	`core`, `engine`
`mipd/`	Engine-as-prior posterior PK: F/CL/renal latents via SIR, covariate individualization (CrCl, weight/age), target-attainment dose recommendation	`core`, `engine`, `graph`, `regimen`, `sbi`
`pipeline/`	Orchestrator wiring all layers	all layers
`ddi.py`	Drug-drug interactions (competitive inhibition, E_max induction)	`core`, `graph`
`pkpd.py`	PK/PD effect modeling (effect compartment, sigmoid E_max)	`core`

Only the structure-only Cmax prediction path is production-supported. regimen, sbi, mipd, ddi, and pkpd are experimental research modules and are not validated for clinical decisions. See docs/architecture/product_scope.md.

Layer isolation. No cross-layer imports outside designated dependencies. predict does not import engine. engine does not import predict. regimen wraps engine without modifying it. Shared data types live in core.py.

Design principles

Identity-blind engine. The ODE compiler and flux functions operate on node/edge types, never on identities. No string matching on organ names, enzyme names, or drug names exists in engine/. Replacing all organ names with random strings produces identical numerical results.
Distribution-native. All physiological and drug parameters are Distribution objects. Point estimates are represented as Distribution(mean=x, cv=0). The uncertainty system is not an add-on; it is the system’s native representation.
Compile once, parameterize many. Graph topology is compiled into an ODE skeleton once. MC iterations change only parameter values, not structure. 1,000 MC samples = 1 compilation + 1,000 ODE solves.

Extending the Model

The architecture is designed so that new compartments, routes, populations, interaction models, and clinical workflows require zero changes to the ODE engine. This was validated empirically: subcutaneous injection, pediatric physiology, tumor compartment, DDI, PK/PD, multi-dose regimen, and TDM were each implemented with 0 lines changed in src/sisyphus/engine/.

New organ (tumor compartment)

nodes:
  - name: tumor
    type: organ
    volume: 0.05
    composition: {fn: 0.013, fp: 0.010, fw: 0.700, pH: 6.8}
edges:
  - {source: arterial_blood, target: tumor, type: flow, flow_fraction: 0.005}
  - {source: tumor, target: venous_blood, type: flow}

New route (subcutaneous injection)

graph.add_node(Node(name="sc_depot", node_type="lumen", volume=Distribution(0.01)))
graph.add_edge(AbsorptionEdge(source="sc_depot", target="venous_blood",
                               ka_fraction=Distribution(1.0)))

New population (pediatric)

Allometrically scaled physiology (cardiac output ∝ BW^0.75) with ontogeny-adjusted enzyme abundances (e.g., CYP3A4 at 50% of adult at age 5). Same graph structure, different YAML parameters.

Experimental drug-drug interactions

Competitive CYP inhibition via pre-simulation enzyme abundance adjustment:

from sisyphus.ddi import apply_inhibition, KETOCONAZOLE

inhibited_graph = apply_inhibition(graph, KETOCONAZOLE)
# Mechanistic scenario only; external clinical-pair validation is required.

Experimental PK/PD modeling

Effect compartment with sigmoid E_max response, computed as post-processing on the concentration-time profile:

from sisyphus.pkpd import compute_effect, PDModel

pd = PDModel(ke0=0.5, emax=100.0, ec50=0.05, hill=2.0)
effect = compute_effect(sim_result, pd)

Limitations

Stereoisomer discrimination is unvalidated. The direct-ML track uses a chirality-blind Morgan fingerprint, so enantiomers with the same connectivity receive the same ML features. Curated full-InChIKey registries can distinguish specific structures, but their isomer-specific C_max accuracy has not been independently tested.
Evaluated scope is narrower than the engine's route support. The development C_max claim concerns oral small-molecule parent drugs. Biologics and non-oral routes have no independent C_max accuracy evaluation here.
Prodrug activation remains experimental. The seven-substrate registry supports parent-to-active routing, but its three-drug clinical gate still fails (sepiapterin 4748×, tebipenem pivoxil 9.05×, fostamatinib 4.50× in the recorded v3 test). Prodrugs are flagged outside the applicability domain. Clopidogrel parent C_max remains in the consumed development benchmark; that label does not validate active-metabolite prediction. See CHANGELOG.md for mechanism and version history.
Simplified pK_a. Ionization state is classified by structural rules (carboxylic acid → 4.5, aliphatic amine → 9.0), not computed quantum-mechanically. This limits Kp accuracy for highly ionized compounds.
Phase II metabolism is incomplete. NAT2, UGT1A1, UGT1A9, and UGT2B7 have modeled abundances; SULT and other UGT routes remain unmodeled. Parent-drug exposure may be overpredicted when omitted pathways materially contribute to clearance, all else equal. The modeled pathway effects have not been independently validated for clinical dosing.
Transporter-mediated disposition: OATP1B1 only. Hepatic uptake by OATP1B1 is modeled mechanistically via the ECM (closed-form QSSA hepatocyte flux) with per-drug kinetic parameters in data/transporters/oatp1b1.json. Other hepatic transporters (OATP1B3, NTCP, BSEP), intestinal transporters (P-gp, BCRP), and renal transporters (OAT1/3, MATE1/2-K) are not mechanistically modeled. P-gp efflux at the gut wall is approximated via a binary permeability correction.
CL_int prediction is limited on the tested data. The public-only XGBoost model achieved scaffold-CV R² = 0.215 on 995 TDC Hepatocyte_AZ compounds. Tested representation, dataset, and architecture changes did not reliably improve end-to-end C_max on the repeatedly used development cohort. This does not prove an intrinsic R² ceiling or identify assay noise as the sole cause; see the corrected accuracy diagnosis. TDM results are internal, observation-conditioned research and do not establish structure-only or clinical dosing accuracy.
Novel-drug underprediction has an unresolved mechanism. Some 2024–2025 compounds were severely underpredicted in a consumed prospective diagnostic set. The earlier claim that F alone caused these errors is withdrawn: the cited oral CL/F is apparent clearance, not systemic CL, and the ten-drug literature-F comparison mixed incompatible endpoints. Low predicted F and engine↔ML divergence did not provide a useful error flag in the tested sets. A matched human oral/IV study and untouched C_max cohort are needed to distinguish F, clearance, formulation, and other causes; see diagnosis §8 and the F source audit.
Error cancellation constrains component-level improvements. Development experiments show strong residual correlation among many blending variants and sensitivity to compensating ADME errors. This does not prove the current blend is optimal. Further selection on N=107 would deepen adaptive overfitting; changes must be hypothesis-driven, trained without external labels, and judged once on the blinded external protocol.
IV-Cmax observation convention. For intravenous bolus dosing, engine Cmax is extracted as the maximum concentration over t ≥ 5 min, rather than the instantaneous t = 0 spike. Oral drugs use the full-interval maximum. All 73 currently scored development compounds are oral, so this convention does not affect that AAFE; IV accuracy remains unvalidated.
ECM (Extended Clearance Model) generalization unverified for non-statins. Under the current public profile, pravastatin, pitavastatin, and fluvastatin pass their numerical ECM-forced Cmax gates; rosuvastatin and atorvastatin remain expected failures. Fluvastatin is marked ecm_applicable=false in production because its CYP2C9-dominant clearance makes the ECM-forced result biologically inapplicable. A pre-registered generalization test on valsartan + glimepiride (2026-04-22, N=2) returned Mode C with systematic 2.5× underprediction under V3 methodology. The fup override hypothesis was ruled out (DE-33); candidates remain for Jmax calibration, Vss/Kp over-distribution, and ECM architectural limits for Km > 1 µM substrates. Users should not rely on ECM accuracy for non-statin OATP1B1 substrates without independent verification.
R_B:P defaults to 1.0. The RBP model (R² = −0.08 on external data) is effectively disabled; all drugs are assumed to have equal blood and plasma concentrations.
90% residual interval is not externally calibrated. The parameter-only Monte Carlo interval covered only 29.9% of N=107 observations at nominal 90%. The user-facing Meta band is a much wider empirical residual quantile whose component predictions are not fully out-of-sample for its calibration records. It is shown only for the public, default-Kp, four-track oral path after its calibration cache and model hashes pass a first-use check; altered partition methods, licensed-profile predictions, track fallbacks, and stale artifacts have no such band. It is not a valid split-conformal guarantee. Independent calibration must use nested OOF predictions or a separate untouched calibration cohort, followed by one-time evaluation on a different blinded holdout.
Dose recommendation and TDM are experimental. Linear dose scaling is rejected automatically for modeled saturable metabolism unless explicitly overridden. That guard prevents a known misuse but does not clinically validate MIPD, TDM, or any posterior interval; these modules remain outside the main product claim.

Project Structure

src/sisyphus/
├── core.py              # Distribution, TissueComposition, data contracts
├── descriptors.py       # Morgan fingerprints + RDKit descriptors
├── compounds.py         # Compound YAML → DrugOnGraph loader
├── ddi.py               # Drug-drug interaction modeling
├── pkpd.py              # PK/PD effect compartment + Emax
├── cli.py               # Cmax CLI + explicitly experimental research commands
├── resources.py         # explicit public/licensed model-data profiles
│
├── graph/               # Body graph definition and construction
│   ├── types.py         # Node, Edge type hierarchy (frozen dataclasses)
│   ├── body.py          # BodyGraph (add/remove/validate/sample)
│   ├── builder.py       # YAML → BodyGraph with flow conservation check
│   └── presets.py       # reference_man() (reference_woman YAML not shipped)
│
├── engine/              # ODE compilation and solving (identity-blind)
│   ├── compiler.py      # ODECompiler, CompiledODE, ResolvedParams
│   ├── flux.py          # FluxSpec implementations (8 transport types,
│   │                    #   incl. ECM ActiveTransport, ProdrugActivation,
│   │                    #   OneCompartmentElimination)
│   ├── result.py        # SimResult dataclass
│   ├── solver.py        # LSODA wrapper (solve, solve_mc)
│   └── uncertainty.py   # Monte Carlo propagation
│
├── predict/             # SMILES → drug parameterization
│   ├── chemistry.py     # Molecular profiling, pKa, AD assessment
│   ├── adme.py          # XGBoost ADME property prediction
│   ├── ivive.py         # In vitro → in vivo extrapolation, Kp
│   ├── hepatic_fu_correction.py # hepatic intracellular fu correction registry
│   ├── drugbank.py      # DrugBank enrichment (licensed_research profile only)
│   ├── phenotype.py     # Pharmacogenomic phenotype (e.g., SLCO1B1)
│   ├── registry.py      # Prodrug activation registry (SMILES-keyed)
│   ├── cyp_clearance_overrides.py # metabolic_fraction registry (OATP1B1 substrates, ECM)
│   ├── non_cyp_substrates.py # NAT2 / UGT1A1 substrate-class registry (SMILES-keyed)
│   └── transporter_db.py # OATP1B1 + hepatic ECM kinetic parameters
│
├── ml/                  # Data-driven PK prediction
│   ├── features.py      # Feature vector construction
│   ├── models.py        # XGBoost Cmax predictor
│   ├── clf_predictor.py # CL/F analytical track
│   ├── registry.py      # Model manifest + feature-schema hash (H2)
│   ├── surrogate.py     # Optional ML surrogate solver (experimental; relocated from engine/ in PR #39)
│   └── ensemble.py      # 4-track meta-learner (_W_VDSS=0.20)
│
├── pk/                  # PK endpoint extraction
│   ├── endpoints.py     # SimResult → PKEndpoints (route-aware t_min_h)
│   ├── nca.py           # Non-compartmental analysis (AUC, t½)
│   └── analytical.py    # Closed-form 1-cpt and 2-cpt solutions
│
├── regimen/             # Multi-dose and clinical pharmacology
│   ├── types.py         # DosingEvent, DosingRegimen (frozen dataclasses)
│   ├── solver.py        # Event-driven multi-dose solver
│   ├── profile.py       # Steady-state detection, Css metrics
│   ├── tdm.py           # Bayesian TDM dispatcher (SBI/IS/IBIS routing)
│   ├── tdm_sbi.py       # SBI (neural posterior) TDM method
│   ├── tdm_ibis.py      # Iterative Bayesian Importance Sampling
│   ├── tdm_enkf.py      # Ensemble Kalman filter (sequential alternative)
│   └── dosing.py        # Linear-scaling dose adjustment (see mipd/ for engine-as-prior MIPD)
│
├── sbi/                 # Amortized neural posterior estimation
│   ├── priors.py        # Drug/physiology prior distributions
│   ├── simulator.py     # Forward model for SBI training
│   ├── amortizer.py     # Neural posterior trainer (NSF)
│   ├── multi_drug.py    # Multi-drug amortizer + continuous (bw, age)
│   ├── physiology_generator.py  # Achour correlated physiology sampler
│   └── sbc.py           # Simulation-Based Calibration gate
│
├── mipd/                # Engine-as-prior posterior PK (model-informed precision dosing)
│   ├── core.py          # F latent + SIR posterior (FPrior, MeasuredF/Cmax/AUC, PosteriorPK)
│   ├── clgrid.py        # CL latent + MeasuredConc via engine clint-scale grid surrogate
│   ├── grid.py          # build_cl_grid (CrCl-aware drug renal_clearance scaling)
│   ├── renal_grid.py    # Renal-CL latent grid (steady-state IV)
│   ├── tdm.py           # predict_tdm — steady-state IV trough TDM (renal-CL latent)
│   ├── covariates.py    # Covariates(crcl_ml_min, ...) + Cockcroft-Gault CrCl estimate
│   ├── dosing.py        # recommend_dose — target-attainment dose recommendation
│   ├── meta.py          # Route the posterior through the meta blend (product interval)
│   ├── api.py           # predict_posterior entry point
│   └── amortizer.py     # (shared amortizer hooks)
│
├── physiology/          # Achour correlated-physiology registry (source package)
│   └── correlation_registry.py  # log-space correlation matrices + correlated-lognormal sampling
│
├── validation/          # Benchmarking infrastructure
│   ├── reference.py     # Clinical PK reference loader (331 drugs)
│   ├── benchmark.py     # Holdout benchmark runner (--compute-pi from H3)
│   ├── metrics.py       # AAFE, fold error, PI coverage
│   └── oatp_generalization.py  # ECM substrate generalization classifier
│
└── pipeline/            # End-to-end orchestration
    ├── predict.py       # SMILES → PredictionResult (route-conditional V3)
    └── config.py        # Pipeline configuration

data/
├── physiology/          # BodyGraph YAML definitions
├── compounds/           # Curated compound configurations
└── reference/           # Clinical PK reference data, holdout split

models/                  # Pre-trained XGBoost models (committed; ~31MB)

Predecessor

Sisyphus inherits curated data assets from Omega PBPK (591 commits) but not its architecture:

Inherited (data)	Not inherited (architecture)
331-drug clinical reference	35-state hardcoded ODE system
Scaffold-stratified holdout split	Organ-specific CL_int fields
ICRP physiology values	Sequential ADME → IVIVE chain
Pre-trained XGBoost models	Point-estimate pipeline
Rodgers & Rowland tissue compositions	Post-hoc hybrid selector

Key empirical findings from Omega that informed Sisyphus:

Data quality dominates model choice. 14 reference corrections reduced AAFE by 47.5% with zero model changes.
Gut CL_int > hepatic CL_int for C_max. Global sensitivity analysis (Sobol): gut S_T=0.47, hepatic S_T=0.00.
Meta-learner > fixed ensemble. Feature importance: ML C_max 50%, PBPK C_max 26%.

References

Berezhkovskiy, L. M. (2004). Volume of distribution at steady state for a linear pharmacokinetic system with peripheral elimination. J Pharm Sci, 93(6), 1628–1640.
Houston, J. B. (1994). Utility of in vitro drug metabolism data in predicting in vivo metabolic clearance. Biochem Pharmacol, 47(9), 1469–1479.
Huang, K., et al. (2021). Therapeutics Data Commons: Machine learning datasets and tasks for drug discovery and development. NeurIPS Datasets and Benchmarks.
ICRP (2002). Basic anatomical and physiological data for use in radiological protection: reference values. ICRP Publication 89.
Obach, R. S., et al. (1997). The prediction of human pharmacokinetic parameters from preclinical and in vitro metabolism data. J Pharmacol Exp Ther, 283(1), 46–58.
Petzold, L. R. (1983). Automatic selection of methods for solving stiff and nonstiff systems of ordinary differential equations. SIAM J Sci Stat Comput, 4(1), 136–148.
Poulin, P., & Theil, F. P. (2002). Prediction of pharmacokinetics prior to in vivo studies. J Pharm Sci, 91(4), 940–951.
Rodgers, T., & Rowland, M. (2005). Physiologically based pharmacokinetic modelling 2: Predicting the tissue distribution of acids, very weak bases, neutrals and zwitterions. J Pharm Sci, 95(6), 1238–1257.
Rodgers, T., & Rowland, M. (2006). Mechanistic approaches to volume of distribution predictions: Understanding the processes. Pharm Res, 24(5), 918–933.
Shimada, T., et al. (1994). Interindividual variations in human liver cytochrome P-450 enzymes involved in the oxidation of drugs, carcinogens and toxic chemicals. J Pharmacol Exp Ther, 270(1), 414–423.
Shitara, Y., Maeda, K., Ikejiri, K., Yoshida, K., Horie, T., & Sugiyama, Y. (2013). Clinical significance of organic anion transporting polypeptides (OATPs) in drug disposition: their roles in hepatic clearance and intestinal absorption. Biopharm Drug Dispos, 34(1), 45–78.
Yang, J., Jamei, M., Yeo, K. R., Tucker, G. T., & Rostami-Hodjegan, A. (2007). Prediction of intestinal first-pass drug metabolism. Curr Drug Metab, 8(7), 676–684.
Yoshikado, T., Toshimoto, K., Nakada, T., Ikejiri, K., Kusuhara, H., Maeda, K., & Sugiyama, Y. (2017). Comparison of methods for estimating unbound intracellular-to-medium concentration ratios in rat and human hepatocytes using statins. Drug Metab Dispos, 45(7), 779–789.
Yu, L. X., & Amidon, G. L. (1999). A compartmental absorption and transit model for estimating oral drug absorption. Int J Pharm, 186(2), 119–125.

How to Cite

If you use Sisyphus in your research, please cite the preprint:

Yoon, J. M. (2026). Sisyphus: A Topology-Compiled Physiologically Based
Pharmacokinetic Platform with Structure-Only Input and Bayesian Parameter
Refinement. ChemRxiv. https://doi.org/10.26434/chemrxiv.15004452/v1

Software (this repository):

Yoon, J. M. (2026). Sisyphus (0.4.0): Graph-based whole-body PBPK
simulation with native uncertainty propagation.
https://github.com/jam-sudo/Sisyphus

The git tag v1.0.0 (commit d6ee9a6) records an earlier feature-branch milestone that has been superseded by the current main line under a more conservative measurement regime. See CHANGELOG.md for details.

Requirements

Python ≥ 3.10
numpy, scipy, pyyaml (core)
rdkit, xgboost, scikit-learn (prediction)

License

MIT

Name		Name	Last commit message	Last commit date
.github/workflows		.github/workflows
app		app
archive/contaminated		archive/contaminated
data		data
docs		docs
models		models
scripts		scripts
server		server
src/sisyphus		src/sisyphus
tests		tests
web		web
.dockerignore		.dockerignore
.gitignore		.gitignore
CHANGELOG.md		CHANGELOG.md
CNAME		CNAME
DESIGN.md		DESIGN.md
LICENSE		LICENSE
README.md		README.md
Sisyphus_Preprint.pdf		Sisyphus_Preprint.pdf
_config.yml		_config.yml
favicon.svg		favicon.svg
index.html		index.html
pyproject.toml		pyproject.toml
requirements-lock.txt		requirements-lock.txt

Latest commit

History

Folders and files

Repository files navigation

Sisyphus

Methodology

Physiological model

ODE formulation

Prediction pipeline

Uncertainty propagation

Multi-dose regimen simulation

Therapeutic drug monitoring

Model-informed precision dosing

Experimental engine-as-prior posterior PK (MIPD module)

Experimental drug-drug interactions

Experimental PK/PD link

Quickstart

Installation

CLI

Python API

Engine-only mode (known compound parameters)

Monte Carlo uncertainty

Validation

Engine validation: Omega parity and post-correction snapshots

Retrospective development benchmark (SMILES → Cmax)

Historical diagnostic P0 source-adjudicated pilot

AI-assisted blind P1 diagnostic (earlier public model)

AI-assisted P2 FDA acquisition check

Experimental multi-dose regression checks

Experimental TDM regression checks

Performance

Test suite

Architecture

Design principles

Extending the Model

New organ (tumor compartment)

New route (subcutaneous injection)

New population (pediatric)

Experimental drug-drug interactions

Experimental PK/PD modeling

Limitations

Project Structure

Predecessor

References

How to Cite

Requirements

License

About

Topics

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages

Retrospective development benchmark (SMILES → C_max)