Characterization of inter-antibody variability in clinical pharmacokinetics of monoclonal antibodies using low-dimensional neural ordinary differential equations (NODE)

Abstract Monoclonal antibodies (mAbs) exhibit high inter-antibody variability in clinical pharmacokinetics (PK), influenced by both molecular and physiological factors. Traditional models, including PBPK and TMDD frameworks, provide mechanistic detail but can be computationally intensive and limited in their ability to generalize across diverse antibodies. Neural ordinary differential equations (NODEs) represent a flexible, low-dimensional alternative that combines mechanistic interpretability with machine learning adaptability. This study aimed to apply NODE-based modeling to characterize inter-antibody PK variability and assess predictive generalizability across a large set of clinical mAbs. A comprehensive dataset of 36 mAbs (21 approved and 15 investigational at the time of the source publication), comprising 143 dose-route concentration-time profiles, was assembled, and two NODE-based frameworks were developed: i) NODE-2CMT, a two-compartment disposition model in which a single-hidden-unit neural network replaces an explicit Michaelis-Menten/target-mediated elimination term, and ii) NODE-mPBPK, a second-generation two-pore minimal PBPK model in which plasma elimination is represented entirely by the same network. Both models were estimated in Monolix on a training set of 31 antibodies (128 profiles, 1,259 observations) and evaluated using goodness-of-fit diagnostics, prediction-corrected visual predictive checks, and an antibody-level train/test split, with five antibodies (15 profiles, 164 observations) withheld from estimation entirely. Both frameworks described the training data closely, the NODE-2CMT model giving the lower objective function value (ΔOFV = 553.7 at an equal number of estimated parameters). On the held-out antibodies, and without re-estimation of any population parameter, predictive accuracy was comparable and acceptable for both models (NODE-2CMT: absolute average fold error [AAFE] 1.40, average fold error [AFE] 1.13, 89% of predictions within 2-fold; NODE-mPBPK: AAFE 1.39, AFE 1.05, 92% within 2-fold), so the better in-sample fit did not translate into better out-of-sample prediction. Direct interrogation of the fitted networks showed that, across the observed concentration range, both learned elimination functions were effectively linear, implying apparent plasma clearances of 6.78 mL/h (NODE-2CMT, additional to an explicit linear clearance of 3.04 mL/h) and 11.2 mL/h (NODE-mPBPK); no saturable elimination component was identified in this dataset. Low-dimensional NODEs can be embedded in mechanistic antibody PK models, estimated jointly with between-antibody random effects, and applied to antibodies that did not contribute to estimation. They therefore hold promise for cross-molecule learning and for translational PK prediction in therapeutic antibody development, provided the learned term is reported and interpreted as a data-driven elimination function rather than as evidence of a specific disposition mechanism.