Method Profile
PhINs uses an equation-constrained network to reconstruct a synthetic cell count and its hidden time-varying net-effect function.
Benchmark observed-state and hidden-function recovery in a controlled partially observed pharmacodynamic problem.
Example 3 synthetic cancer-cell count N(t), generated from proliferation, time-varying death and a carrying-capacity term; only N(t) is observed.
Compare physics-informed multilayer perceptrons and Kolmogorov–Arnold networks with different sampling and residual-weighting schemes over a 600-hour generating interval.
Synthetic observations
Known cell-count and net-function ground truth; Table 3 compares 20/120 observation settings and 0%/5% noise.
Mechanistic constraint
dN/dt = FD(t) N(t); physical FD equals 0.03 times the learned hidden output, with an initial-value anchor in the described loss.
Evaluate the observed trajectory and hidden function separately against their known generating values. FD is the net count-change function, not an independently measured drug-killing rate.
Source-reported synthetic benchmark compares architectures and training settings; no real drug-response experiment is claimed.
One controlled synthetic model; Table 3 compares 20 or 120 observations with 0% or 5% noise.
Known generating N(t)/FD(t) and the other listed network/sampling/weighting configurations.
At 120 observations/5% noise, PINN N/FD RMSE is 7.78 × 10−3 / 8.12 × 10−4; PIKAN is 3.89 × 10−3 / 4.70 × 10−4 in the reported configurations.
PhINs is an open-source library for physics-informed pharmacometric modeling. Three synthetic examples distinguish fitting observed trajectories from recovering unknown parameters and a hidden time-varying chemotherapy-response function.
