Fitting Complex PK/PD Models with Stan

Size: px

Start display at page:

Download "Fitting Complex PK/PD Models with Stan"

Philippa Harrison
6 years ago
Views:

1 Fitting Complex PK/PD Models with Stan University of Warwick Bayes Pharma Novartis Basel Campus May 20, 2015

2 Because of the clinical applications, precise PK/PD modeling is incredibly important. PK/PD Model

3 Because of the clinical applications, precise PK/PD modeling is incredibly important. PK/PD Model Treatment R(T, )!

4 Robust treatment decisions requires modeling both the latent PK/PD process and the measurement. PK/PD Model Treatment R(T, )!

5 Robust treatment decisions requires modeling both the latent PK/PD process and the measurement. PK/PD Model Data D Treatment R(T, )!

6 Robust treatment decisions requires modeling both the latent PK/PD process and the measurement. PK/PD Model Measurement Model (D ) Data D Treatment R(T, )!

7 Analytic models are common due to their computational convenience but ultimately too restrictive. PK/PD Model Measurement Model (D ) Data D Treatment R(T, )!

8 Analytic PK/PD models, for example, are exactly solvable. PK/PD Model Measurement Model (D ) Data D Treatment R(T, )!

9 Analytic PK/PD models, for example, are exactly solvable. dc dt = V m V C K m + C

10 These analytic models, however, can t capture many nonlinear effects in more realistic PK/PD systems. dc dt = V m V C K m + C + k V e kt

11 Similarly, analytic measurement models are straightforward to fit. PK/PD Model Measurement Model (D ) Data D Treatment R(T, )!

12 Similarly, analytic measurement models are straightforward to fit. bc t,n (V,V m )=N (C(t, V, V m ), )

13 Similarly, analytic measurement models are straightforward to fit. bc t,n (V,V m )=N (C(t, V, V m ), ) NX C(t, V, V m ) 1 N Ĉ t,n (V,V m ) n=1

14 Similarly, analytic measurement models are straightforward to fit. bc t,n (V,V m )=N (C(t, V, V m ), ) NX C(t, V, V m ) 1 N Ĉ t,n (V,V m ) V,V m C 1 1 N n=1 NX n=1 Ĉ t,n (V,V m )!

15 But for population data analytic measurement models either introduce bias or suffer from large uncertainties. dc dt = V m V C K m + C

16 But for population data analytic measurement models either introduce bias or suffer from large uncertainties. dc dt = V n m V n C K m + C

17 But for population data analytic measurement models either introduce bias or suffer from large uncertainties. dc dt = V m n V n C K m + C

18 But for population data analytic measurement models either introduce bias or suffer from large uncertainties. V 1,V 1 m V n,v n m V N,V N m......

19 If we want to maximize the utility of the data then we need non-analytic measurement models. V 1,V 1 m V n,v n m V N,V N m µ V,! V µ Vm,! Vm V n log N (µ V,! V ) V n m log N (µ Vm,! Vm )

20 Non-analytic models are practically difficult because they are computationally demanding to fit. PK/PD Model Measurement Model (D ) Data D Treatment R(T, )!

21 Recall that in Bayesian inference our complete model is specified with a posterior distribution.

22 Recall that in Bayesian inference our complete model is specified with a posterior distribution. ( )

23 Recall that in Bayesian inference our complete model is specified with a posterior distribution. (D ) ( )

24 Recall that in Bayesian inference our complete model is specified with a posterior distribution. ( D) / (D ) ( )

25 And all well-posed statistical manipulations are expectations with respect to the posterior. Z E[f] = d ( D)f( )

26 Expectations, however, are computationally demanding when the model is complex. Z E[f] = d ( D)f( )

27 The problem is that probability concentrates on a nonlinear surface called the typical set.

28 In order to estimate expectations we can use Markov chain Monte Carlo to find and explore the typical set.

29 In order to estimate expectations we can use Markov chain Monte Carlo to find and explore the typical set.

30 Given the complexity of the PK/PD model, this exploration has to be extremely efficient to be practical.

31 Random Walk Metropolis explores with an ineffective guided diffusion.

32 Random Walk Metropolis explores with an ineffective guided diffusion.

33 Random Walk Metropolis explores with an ineffective guided diffusion.

34 The Gibbs sampler updates each parameter one-at-a-time to no greater success.

35 The Gibbs sampler updates each parameter one-at-a-time to no greater success.

36 In order to explore efficiently we need coherent exploration.

37 In order to explore efficiently we need coherent exploration.

39 Modeling Language

40 Modeling Language Hamiltonian Monte Carlo

41 Modeling Language Hamiltonian Monte Carlo

42 A strongly typed modeling language allows users to specify complex models with minimal effort.

43 Stan not only admits the specification of a system of ordinary differential equations functions { real[] onecomp_mm_infusion(real t, real[] y, real[] theta, real[] x_r, int[] x_i) {! real dydt[1];! if (t < x_r[2]) dydt[1] <- - theta[2] * y[1] / ( theta[1] * (theta[3] + y[1]) ) + x_r[1] / (theta[1] * x_r[2]); else dydt[1] <- - theta[2] * y[1] / ( theta[1] * (theta[3] + y[1]) );! return dydt;! } }

44 But also high-performance integrators. transformed parameters { for (p in 1:N_patients) { V[p] <- exp(mu_v + eta_v[p] * omega_v); V_m[p] <- exp(mu_v_m + eta_v_m[p] * omega_v_m);! } theta[1] <- V[p]; theta[2] <- V_m[p]; theta[3] <- K_m; C <- integrate_ode(onecomp_mm_infusion, C0, t0, t, theta, x_r, x_i);

45 Consequently, even complex statistical modeling of PK/PD systems is straightforward in Stan. model { eta_v ~ normal(0, 1); mu_v ~ normal(log(5), 1); sigma_v ~ cauchy(0, 1); eta_v_m ~ normal(0, 1); mu_v_m ~ cauchy(0, 1); sigma_v_m ~ cauchy(0, 1);! } K_m ~ cauchy(0, 1); for (p in 1:N_patients) for (n in 1:N_t) C_hat[p, n] ~ lognormal(log(c[p, n]), 0.15);

46 Modeling Language Hamiltonian Monte Carlo

47 Once a posterior has been specified, Stan implements Hamiltonian Monte Carlo to estimate expectations.

48 Let s consider a one-compartment Michaelis-Menten clearance model with a constant infusion. dc dt = V m V C K m + C + I(t) I(t) = 8 < : D VT, 0, t apple 0 0 <t<t 0, T apple t

49 Data is simulated for ten patients, with each patient given their own V and V m. dc n dt = V n m V n C n K m + C n + I(t) Ĉ n (t) log N (C n (t), )

50 Data is simulated for ten patients, with each patient given their own V and V m. dc n dt = V n m V n C n K m + C n + I(t) Ĉ n (t) log N (C n (t), ) V n log N (µ V,! V ) V n m log N (µ Vm,! Vm )

51 Within a few minutes Stan generates all of the samples we need to infer the individual patient concentrations Concentration(mg/L) Time(days)

52 As well as simulated data for constructing posterior predictive checks Concentration(mg/L) Time(days)

53 We can also analyze the system parameters for each individual patient V(L)

54 As well as the population parameters µ V (L) V (L)

55 As well as the population parameters µ Vm (mg/day) Vm (mg/day)

Contents. Part I: Fundamentals of Bayesian Inference 1

Contents. Part I: Fundamentals of Bayesian Inference 1 Contents Preface xiii Part I: Fundamentals of Bayesian Inference 1 1 Probability and inference 3 1.1 The three steps of Bayesian data analysis 3 1.2 General notation for statistical inference 4 1.3 Bayesian