Answers are grounded in this paper and cite their sources.
This study investigates Neural ordinary differential equation (NODE) model to describe clinical pharmacokinetic (PK) data from Monomethyl auristatin E (MMAE)-based antibody-drug conjugates (ADCs). ADC PK is composed of four analytes: total antibody (tmAbs), antibody-conjugated MMAE, unconjugated free MMAE (fMMAE), and ADC. Two different model architectures were explored, one that predicts single-state with one NODE model (single-state NODE) and the other that predicts all states with one single NODE model (multiple-states NODE). All the models were trained with single-dose regimen PK datasets at dose levels of 0.1, 0.5, 1.0, 1.8, 2.4 mg/kg. After training, the models could extrapolate to unseen single-dose regimen datasets at dose levels of 0.15 to 3.3 mg/kg and multiple-dose regimen datasets at dose levels of 0.3 to 3.0 mg/kg with various dosing frequencies (Q1W and Q3W) without introducing covariates such as inter-product variability or inter-study variability. The final model performance, assessed as percent prediction error (%PE) was acceptable. For single-state NODE, the average pooled %PE values across studies were 9.4, -11, 6.6, and 40 for total antibodies, antibody-conjugated MMAE, free MMAE, and ADC, respectively. For multiple-states NODE, the average pooled %PE values across studies were 5.2, -14, 17, and 31 for total antibodies, antibody-conjugated MMAE, free MMAE, and ADCs, respectively. The training process and good training practices for a NODE model development and the limitations of the current model were discussed. This study provides proof of concept for training a NODE model on high-dimensional, highly variable real-world PK data.
No takes yet. Share an insight, caveat, or question.
Huang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: