openpkflow.ml¶
Experimental ML surrogate for PK profiles. Not validated for regulatory use.
Experimental
This module is marked EXPERIMENTAL. Do not use outputs in regulatory submissions. Results have not been validated against reference PK software.
Public API¶
| Symbol | Type | Description |
|---|---|---|
PKSurrogate |
class | MLP surrogate for 1-cmt oral PK profiles |
PKSurrogate.from_1cmt_oral(n_samples, ...) |
classmethod | Factory: generates data, trains, returns fitted surrogate |
PKSurrogate.predict(time, dose, CL_F, Vz_F, ka) |
method | Predict concentration array |
Install¶
Example¶
from openpkflow.ml import PKSurrogate
surrogate = PKSurrogate.from_1cmt_oral(n_samples=5000, seed=42)
import numpy as np
times = np.linspace(0, 24, 100)
C = surrogate.predict(times, dose=100.0, CL_F=5.0, Vz_F=20.0, ka=1.0)
Architecture¶
- Input features:
(time, dose, CL_F, Vz_F, ka)— z-score normalised - Default hidden layers: (64, 64), tanh activations
- Trained with Adam, MSE loss on log-scale concentrations
- Training data generated by
openpkflow.sim.c_1cmt_oral()(analytically exact)
Use cases¶
Appropriate for rapid sensitivity analysis, uncertainty propagation, or
parameter screening where approximate predictions are acceptable. Not a replacement
for the analytical sim module.