What I learned from ACoP13 (Aurora, CO)

A trainee’s perspective

Photo by Myriam Zilles on Unsplash

What an incredible experience it was to see the state-of-the-art techniques used in Model Informed Drug Development (MIDD)! Posters summarized research in fields like Pharmacodynamics, Pharmacokinetics, Pharmacometrics, Quantitative Systems Pharmacology, and many more!

Below is an overview of the many mathematical models used and various applications sorted into:

• Machine learning
• Common Techniques
• Rarest Techniques
• Inverse Problems and Parameter estimation
• Clinical trial simulation

This is not an exhaustive list. Just some insight into the discussions, topics, and fields of interest at ACOP23.

Machine Learning algorithms

• Neural ODEs — learning unknown dynamics of a set of ODEs. Interpretation of dynamics learned is not clear.
• Universal Differential Equations (UDEs) — embedding scientific knowledge into the neural network. The UDE can learn the unknown dynamics. Although interpretation of what is unknown is not clear. If there is intuition on the types of nonlinear dynamics occurring, then curve fitting can be performed.
• UDEs, closely related to Physics-Informed Neural Networks (PINNs)
• Elastic Net (a generalization of LASSO and Ridge regression)
• On interpretability: SHAP is the most common addon to ML algorithms that provide a form of interpretation of the algorithms used. This is based on a game theory concept that determines which players in a cooperative game contribute the most to the success or detriment of the game. This can suggest which input covariates are most influential in the ML task.
• (A suggestion) on finding the nonlinear terms in an ODE: Sparse Identification of Nonlinear Dynamics (SINDy). One must guess the types of nonlinearities that occur and it returns the most likely coefficients. e.g. Michaelas-Menton dynamics

Common Techniques used

• ODEs — systems of Ordinary Differential Equations describing rates of change of concentrations in chemical reactions. For an established model, coefficients are collected from literature, or found by experiment. Most popular application is in Physiologically-Based PharmacoKinetic (PBPK) modeling for drug development. Models can be made to describe any set of chemical reactions believed to occur in an organism (human and animal). Parameters are generally extracted from peer-reviewed literature or estimated through experiment or inverse problems.
• Exposure-Response Analysis, a broader field within MIDD teasing out the relationship between exposure to compounds and the biological/physiological response to it.
• Visual Predictive Checks (VPCs) — a longitudinal visualization of data with 10%, 50%, and 90% quartiles plotted and 95% confidence intervals surrounding them. Here’s a resource provided by Monolix
• Kaplan-Meier Survival Curves used often in survival analysis
• Hierarchical Bayesian Modeling — this models structural/hierarchical relationships. For example, capturing the two level doctor-patient relationship. Or, the three level person-town-virus spread.
• Item Response Theory — to determine the underlying probability distribution from multiple expert ratings.

Rarest Technique used

• Agent-Based Modelling (ABM): a 3D discretization of the brain was used to model plaque buildup to test the amyloid hypothesis. A useful tool to simulate and measure complex dynamics in a system. However, it is computationally expensive to perform.
• Control Theory — Predominantly used in mechanical engineering applications but useful in systems exhibiting hysteresis effect. This methodology creates autonomous systems that respond to changes in environment. For example, using Continuous Glucose Monitors (CGMs) to inform when to administer the appropriate amount of insulin in people with type 1 diabetes.

Inverse Problems and Parameter Estimation

• Used for model calibration: interpolating human model parameters from experimental data on animal models.

Clinical trial simulations

• Dropout Modeling: simulation of multiple trials and effect of patient dropout. Determining sensitivity of relevant metrics effect on dropout time.
• Sample size calculations: determining the minimum number of patients necessary for clinical trials using a clinical measure. One must determine the measure, its sensitivity, and validity for fit-for-purpose use.
• (Other computations) Clinical trial lengths

This was a really fun opportunity. I learned a lot about the many aspects of drug development and how much rigor is needed for technical validation, and drug trial simulations. Leave a comment below if you have any questions/comments!

Photo by Kevin Kandlbinder on Unsplash

Resources:

• Neural ODEs: breakdown of another deep learning breakthrough | by Alex Honchar | Towards Data Science
• Neural ODEs · Depth First Learning
• GitHub — ChrisRackauckas/universal_differential_equations: Repository for the Universal Differential Equations for Scientific Machine Learning paper, describing a computational basis for high performance SciML
• Physics Informed Neural Networks (PINNs): An Intuitive Guide | by Ian Henderson | Oct, 2022 | Towards Data Science
• sklearn.linear_model.ElasticNet — scikit-learn 1.1.3 documentation
• Shapley value — Wikipedia
• Visual Predictive Checks in Monolix (lixoft.com)