1 of 36

0

CMINNs: Compartment model informed neural networks—Unlocking drug dynamics

Nazanin Ahmadi

Brown University

2 of 36

1

Paper

3 of 36

2

CMINNs: Compartment model Informed Neural Networks

Compartmental Modeling

PINNs: Physics-Informed Neural Networks

Scientific Machine Learning

PK/PD Modeling

Outline

4 of 36

3

Pharmacokinetics and Pharmacodynamics Models

The relationship between dose, concentration, and effect is fundamental in clinical pharmacology.

5 of 36

4

Plasma Drug Concentration

Time

Distribution phase

Elimination Phase

a

b

 

 

Plasma

C1 , V1

Peripheral

C2 , V2

K12

K21

K10

 

Traditional Compartmental PK/PD Modeling

6 of 36

5

In Pharmacometrics, "compartments" are abstract mathematical constructs designed to simplify the representation of drug distribution within the human body. These compartments do not correspond to specific anatomical or physiological spaces; rather, they serve as valuable tools for modeling the behavior of substances introduced into the body.

A. Talevi and C. L. Bellera, Compartmental Pharmacokinetic Models, p. Chapter 8. Springer, Cham,2024.

We can use a different model if the calculations are feasible.

We will not lose information if we reduce the number of compartments.

If we have a better tool, we can model them using a different approach.

Traditional Compartmental PK/PD Modeling

7 of 36

6

  • Lack of generalizability across different drugs
  • Structural Identifiability
  • Inability to Model Non-Exponential Decay

Limitations of Traditional Compartmental Modeling

Fractional Calculus

8 of 36

7

Ahmadi, N. , Wang, S., Karniadakis, G. (2024). CMINNs: Compartment Model Informed Neural Networks-Unlocking Drug Dynamics. arXiv preprint arXiv:2409.12998.

Time-varying parameter

Time-varying parameter

Fractional Calculus

Fractional Calculus

Solver

CMINNs Workflow

9 of 36

8

AI-Powered Solver:

Physics-Informed Neural Networks(PINNs)

10 of 36

9

Mathematical Equations

DATA

Partial Physics

Interactions

Neural Networks

How do we blend our understanding of physics and the interactions within the system into the workings of a neural network?

Physics-Informed Neural networks (PINNs)

11 of 36

10

The model hasn’t ‘understood’ the scientific problem.

u

x

Compare to training Data

Output

Input

A damped harmonic oscillator

Neural networks

12 of 36

11

Raissi, M., Perdikaris, P., & Karniadakis, G. E. (2019). Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378, 686-707.

u

x

Compare to training Data

Output

Input

Compute derivatives and minimize underlying equation residual

 

Loss Function

Physics-Informed Neural networks (PINNs)

13 of 36

12

PINNs vs Neural ODEs

NODEs

t

Dynamics

ODE solver

Minimizing Loss

 

14 of 36

13

Data

Data +

Physical Laws

Physics-Informed Neural networks (PINNs)

Raissi, M., Perdikaris, P., & Karniadakis, G. E. (2019). Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational physics, 378, 686-707.

15 of 36

14

Physics-Informed Neural networks (PINNs)

Raissi, M., Perdikaris, P., & Karniadakis, G. E. (2019). Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational physics, 378, 686-707.

16 of 36

15

Fractional Calculus in Pharmacokinetics and Pharmacodynamics Modeling

 

 

 

 

17 of 36

16

Fractional Calculus in Pharmacokinetics and Pharmacodynamics Modeling

Caputo fractional derivative:

Caputo, Michele (1967). "Linear model of dissipation whose Q is almost frequency independent. II". Geophysical Journal International. 13 (5): 529–539. doi:10.1111/j.1365-246x.1967.tb02303.x

18 of 36

17

Huang, Y., & Kim, H. K. (2022). Fractional transit compartment model for describing drug delayed response to tumors using Mittag-Leffler distribution on age-structured PKPD model. PLOS ONE, 17(11), e0276654. https://doi.org/10.1371/journal.pone.0276654

Fractional Calculus for Pharmacodynamics Modeling

Total tumor Cells are given by W = u + y1 + y2 + … + yn

Total tumor Cells are given by W = u + y

19 of 36

18

Fractional Physics-Informed Neural networks (fPINNs)

20 of 36

19

CMINNs: Compartment model Informed Neural Networks

21 of 36

20

Ahmadi, N. , Wang, S., & Karniadakis, G. (2024). CMINNs: Compartment Model Informed Neural Networks--Unlocking Drug Dynamics. arXiv preprint arXiv:2409.12998.

CMINNs Workflow

Time-varying parameter

Time-varying parameter

22 of 36

21

PINNs Block in CMINNs method with time-varying parameter

23 of 36

22

Published Real-World Data

Two-Compartment Model for Gentamicin Pharmacokinetics

Two-Compartment Model for Amiodarone Pharmacokinetics

Three-Compartment Model for Talaporfin Sodium Pharmacokinetics

PK-PD Model

A PK-PD Model for Tumor Growth and Treatment in Animal Models

PK Models

Two-Compartment Model for Amiodarone Pharmacokinetics

A PK-PD Model for Tumor Growth and Treatment in Animal Models

CMINNs’ case study

24 of 36

23

 

 

 

or

The pharmacokinetics of a single dose of amiodarone, an antiarrhythmic drug known for its non-exponential pharmacokinetics have been analyzed using a general two-compartment model as follows:

Two-Compartment Model for Amiodarone Pharmacokinetics

25 of 36

24

PINNs with Time-varying parameter

fPINNs results

Data source: M. D. Freedman and J. C. Somberg, “Pharmacology and pharmacokinetics of amiodarone.,” Journal of clinical pharmacology, vol. 31, no. 11, pp. 1061–1069, 1991.

Two-Compartment Model for Amiodarone Pharmacokinetics - Results

26 of 36

25

 

Pharmacokinetics

Pharmacodynamics

A PK-PD Model for Tumor Growth and Treatment in Animal Models

 

Data and Model: M. Simeoni, P. Magni, C. Cammia, G. D. Nicolao, V. Croci, E. Pesenti, M. Germani, I. Poggesi, and M. Rocchetti, “Predictive pharmacokinetic pharmacodynamic modeling of tumor growth kinetics in xenograft models after administration of anticancer agents,” Cancer Research, vol. 64, no. 3, pp. 1094– 1101, 2004.

27 of 36

26

 

Pharmacokinetics

Pharmacodynamics

Data and Model: M. Simeoni, P. Magni, C. Cammia, G. D. Nicolao, V. Croci, E. Pesenti, M. Germani, I. Poggesi, and M. Rocchetti, “Predictive pharmacokinetic pharmacodynamic modeling of tumor growth kinetics in xenograft models after administration of anticancer agents,” Cancer Research, vol. 64, no. 3, pp. 1094– 1101, 2004.

A PK-PD Model for Tumor Growth and Treatment in Animal Models

 

Data

28 of 36

27

 

Time-varying parameter

Fractional Compartment

 

Data : M. Simeoni, P. Magni, C. Cammia, G. D. Nicolao, V. Croci, E. Pesenti, M. Germani, I. Poggesi, and M. Rocchetti, “Predictive pharmacokinetic pharmacodynamic modeling of tumor growth kinetics in xenograft models after administration of anticancer agents,” Cancer Research, vol. 64, no. 3, pp. 1094– 1101, 2004.

A PK-PD Model for Tumor Growth and Treatment in Animal Models

29 of 36

28

Model Parameters

Rate of cell death k1(t)

Drug potency k2

1st dose

2nd dose

3rd dose

1st dose

2nd dose

3rd dose

A PK-PD Model for Tumor Growth and Treatment in Animal Models - Results

PINNs with Time-varying parameter

30 of 36

29

1st dose

2nd dose

3rd dose

A PK-PD Model for Tumor Growth and Treatment in Animal Models - Results

PINNs with Time-varying parameter

31 of 36

30

A PK-PD Model for Tumor Growth and Treatment in Animal Models - Results

fPINNs results

32 of 36

31

Summary

Challenges

  • Sparse data
  • Nonlinear dynamics
  • Lack of generalizability across different drugs
  • Structural Identifiability
  • Non Exponential Decay

Framework Introduced

  • CMINNs

Applications Demonstrated

  • Amiodarone PK
  • Tumor growth PK-PD model

Outcomes

  • Interpretable results
  • Accurate prediction
  • High-fidelity recovery of hidden biological mechanisms

33 of 36

32

Acknowledgement

Prof. George Karniadakis

Dr. Shupeng Wang

34 of 36

33

Thank you for listening

https://www.linkedin.com/in/ahmadinaz/

Nazanin@Brown.edu

https://www.researchgate.net/profile/Nazanin-Ahmadi-Daryakenari

35 of 36

34

3-Compartment Model

36 of 36

35

2-Compartment Model