James Lu (1), Dan Lu (1), Jin Y. Jin (1)
(1) Modeling & Simulation, Clinical Pharmacology, Genentech
Introduction/Objectives: The introduction of nonlinear mixed-effects modeling has led to a revolution in the analysis of pharmacokinetic (PK) and pharmacodynamic (PD) data. However, the specialized modeling expertise and the analysis time do not allow for an easy adoption of modeling tools by experimental scientists and clinicians at the bench-/bed-side to generate impactful insights in real-time. With the increasing amount and complexity of data being generated and the acceleration of drug development, such challenge is growing.
Methods: We propose a deep-learning framework whereby neural networks are trained by PK/PD and Quantitative System Pharmacology (QSP) models to enable the automated identification of specific patterns in data. The identified patterns can be used to classify the type of drug effect, as well as quantify the relevant parameters. This approach relies on modelers providing the appropriate training data (simulated from existing models) and utilize pertinent methodologies to develop and train the neural networks, which subsequently can be deployed as AI solutions to process and generate insights (e.g., via parametric characterization) from newly observed data. We demonstrate a prototype of the proposed approach utilizing Mathematica (version 11.3) [1] as the deep-learning framework.
Results: We show how a number of different neural network architectures [2] can be utilized to perform various modeling tasks. Within the context of analyzing PK and PK/PD data, neural networks can give rise to: (1) improved accuracy in the estimation of PK parameters (e.g., AUC) as compared to the traditional NCA calculation based on sparse observed data; (2) the ability to identify population PK and PK/PD parameters without conducting traditional modeling. In the context of safety, neural network can be used to infer hematological toxicity mechanisms and parameters from an in-vitro multilineage assay, which have previously required solution via global optimization techniques based on a QSP model [3]. The network dimensions chosen depend on the sizes of the inputs and outputs as well as the task, with representative examples shown in Table 1. To increase the reliability of neural network predictions, we show that adding L2 regularization with an appropriate choice of regularization parameters and drop-out layers can improve their tolerance to data noise while retaining the accuracy of predictions. These deep-learning problems can be solved effectively using the adaptive moment estimation (Adam) method [2]. In all these applications, while the training of the neural network may require some significant computational modeling effort upfront, at deployment the networks can make inferences (e.g., parameter estimates) at a fraction of a second.
Conclusions: With an appropriate choice of network architecture and regularization, we demonstrate that deep-learning has the potential to perform a number of routine PK/PD and QSP modeling tasks accurately, reliably and efficiently. These early explorations suggest that the proposed deep-learning framework could have broad applications in generating insights from data. We envision a future of human-machine partnership whereby repetitive modeling tasks are done by machines so as to free human modelers to carry out more scientific innovations and real-time decision impacts [4].
Table 1
|
Examples |
Training Data |
|
Neural Network |
|
|
|
|
Input Layer |
Intermediary Layers |
Output Layer |
|
PK: Estimate AUC from 5 time points |
Simulated time-concentration in 300,000 virtual subjects from compartmental PK models |
2×5 matrix (time – concentration) |
2 convolutional, 12 rectified linear units (ReLU), 10 fully connected [2] |
AUC estimate |
|
Safety: Estimate 28 QSP parameters from in-vitro multilineage assay [3] |
Simulated concentration-response in 100,000 virtual experiments from QSP model [3] |
14×6 matrix (concentration – response) |
2 convolutional, 4 ReLU, 3 fully connected layers [2] |
28 parameters |
References:
[1] https://www.wolfram.com/mathematica/
[2] Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep learning. MIT press.
[3] Lu, J., Corr, N., Fullerton, A., Miles, D., Jin, J., Lu, D. (2019). Systems modeling of a novel multilineage hematopoiesis assay for the deconvolution of mechanisms of drug-induced myelosuppression, ASCPT Annual Meeting, abstract QSP-009.
[4] Altman, R. B. (2017). Artificial intelligence (AI) systems for interpreting complex medical datasets. Clinical Pharmacology & Therapeutics, 101(5), 585-586.
Reference: PAGE 28 (2019) Abstr 8921 [www.page-meeting.org/?abstract=8921]
Poster: Methodology - New Modelling Approaches