Application of a single-objective, hybrid genetic algorithm approach to pharmacokinetic model building

Abstract

A limitation in traditional stepwise population pharmacokinetic model building is the difficulty in handling interactions between model components. To address this issue, a method was previously introduced which couples NONMEM parameter estimation and model fitness evaluation to a single-objective, hybrid genetic algorithm for global optimization of the model structure. In this study, the generalizability of this approach for pharmacokinetic model building is evaluated by comparing (1) correct and spurious covariate relationships in a simulated dataset resulting from automated stepwise covariate modeling, Lasso methods, and single-objective hybrid genetic algorithm approaches to covariate identification and (2) information criteria values, model structures, convergence, and model parameter values resulting from manual stepwise versus single-objective, hybrid genetic algorithm approaches to model building for seven compounds. Both manual stepwise and single-objective, hybrid genetic algorithm approaches to model building were applied, blinded to the results of the other approach, for selection of the compartment structure as well as inclusion and model form of inter-individual and inter-occasion variability, residual error, and covariates from a common set of model options. For the simulated dataset, stepwise covariate modeling identified three of four true covariates and two spurious covariates; Lasso identified two of four true and 0 spurious covariates; and the single-objective, hybrid genetic algorithm identified three of four true covariates and one spurious covariate. For the clinical datasets, the Akaike information criterion was a median of 22.3 points lower (range of 470.5 point decrease to 0.1 point decrease) for the best single-objective hybrid genetic-algorithm candidate model versus the final manual stepwise model: the Akaike information criterion was lower by greater than 10 points for four compounds and differed by less than 10 points for three compounds. The root mean squared error and absolute mean prediction error of the best single-objective hybrid genetic algorithm candidates were a median of 0.2 points higher (range of 38.9 point decrease to 27.3 point increase) and 0.02 points lower (range of 0.98 point decrease to 0.74 point increase), respectively, than that of the final stepwise models. In addition, the best single-objective, hybrid genetic algorithm candidate models had successful convergence and covariance steps for each compound, used the same compartment structure as the manual stepwise approach for 6 of 7 (86 %) compounds, and identified 54 % (7 of 13) of covariates included by the manual stepwise approach and 16 covariate relationships not included by manual stepwise models. The model parameter values between the final manual stepwise and best single-objective, hybrid genetic algorithm models differed by a median of 26.7 % (q1 = 4.9 % and q3 = 57.1 %). Finally, the single-objective, hybrid genetic algorithm approach was able to identify models capable of estimating absorption rate pa...

Covariate selection in pharmacometric analyses: a review of methods

“…Bies et al [50] introduced the genetic algorithm (GA) into the pharmacometric literature, which was followed by Sherer et al [51] more recently. The authors use the algorithm for overall model selection, including the structural model (e.g.…”

“…Additional features, such as 'nicheing' and 'downhill search' were included by the authors to improve the robustness and/or convergence of the GA algorithm. Bies et al [50] and Sherer et al [51] added the following components to the penalty of the GIC other than those based on sample size or dimensionality: a 400 unit penalty if the covariance matrix (COV) was nonpositive semidefinite; and a 300 unit penalty if any estimated pairwise correlations of the estimates were >0.95 in absolute value. We do not find these penalties necessary for pure variable selection.…”

Meta-Heurísticas Na Configuração De Parâmetros Em Simulação Computacional

“…Nos casos onde ela não é encontrada os possíveis motivos são, entre outros, número insuficiente de partículas ou de troca de informações, complexidade do problema e tamanho do espaço de busca. Sherer et al (2012) introduz Algoritmos Genéticos (Genetic Algorithms, GA) como uma classe de técnicas de pesquisa global normalmente utilizada em engenharia e ciência para estimação de parâmetros multi-variáveis; esta ferramenta tem sido aplicada com sucesso em problemas como distribuição elétrica, redes biológicas, em processo de agendamento. O método, como seu nome sugere, simula o processo biológico de seleção natural.…”

“…Técnicas avançadas como "niching" ("de nicho") e elitism ("de elite") evitam que os indivíduos procurem em apenas uma parte do espaço e que os melhores indivíduos fiquem estagnados entre gerações, respectivamente (SHERER et al, 2012).…”

Machine Learning in Drug Discovery and Development Part 1: A Primer

“…For example, an AI framework in drug discovery may optimize drug candidates through a combination of ML models that predict favorable physicochemical characteristics (e.g., solubility and permeability), pharmacokinetics (PK), safety, and possibly efficacy . An AI framework in drug development may use ML methods to prescreen covariates in PK‐pharmacodynamic data, identifying patient subpopulations, predicting clinical outcomes, informing clinical trial design, and investigating novel therapeutic purpose for existing drugs (i.e., drug repositioning or drug repurposing) . However, ML methods have been utilized more often in drug discovery than in development.…”