Prediction of Plasma Concentration-Time Profiles in Mice from Chemical Structure Using a Kinetically Constrained Hybrid Neural ODE

Predicting plasma concentration-time profiles directly from chemical structure remains intrinsically challenging in early drug discovery. Existing approaches typically fall into two paradigms: mechanistic reconstruction via predicted PK parameters, which is constrained by the assumed model structure, and direct data-driven prediction of concentrations, which often suffers from instability and poor temporal coherence under sparse and irregular sampling. Here, we present a kinetically constrained hybrid Neural ordinary differential equation (KCH-NODE) framework for structure-to-PK prediction that bridges these limitations. KCH-NODE anchors predictions to a shared, learnable two-compartment typical trajectory while modeling compound-specific PK behavior as continuous-time deviations governed by a Neural ODE. This hybrid design generates flexible yet pharmacokinetically coherent trajectories, even under sparse observations. Trained on mouse intravenous (IV) and oral (PO) PK data from 1159 compounds across 30 projects, KCH-NODE generally provided more accurate predictions than a purely data-driven Neural ODE in both compound-level and scaffold-based cross-validation (median RMSE 0.374 and 0.388 for IV; 0.564 and 0.570 for PO) and in a temporally held-out evaluation (223 compounds; median RMSE 0.637 for IV and 0.731 for PO), with a smaller loss of accuracy on the temporal split. Predictive accuracy for oral administration nevertheless remained limited on the temporal test set for both methods. Integrated Gradients supported mechanistically consistent feature contributions, and structural similarity to the model-development dataset together with ensemble disagreement provided complementary indicators of prediction reliability. These results indicate that kinetically constrained hybrid Neural ODEs offer an effective and interpretable framework for structure-to-PK profile prediction.