Automated model selection in population pharmacokinetics using metaheuristic algorithms

Technical Summary: Automated Model Selection in Population Pharmacokinetics Using Metaheuristic Algorithms

Problem Statement
Population pharmacokinetics (PopPK) analysis relies on nonlinear mixed-effects (NLME) modeling, a process that traditionally requires extensive expert judgment and manual stepwise model development. This manual approach is resource-intensive, lacks standardization, and risks becoming trapped in local optima due to the sequential nature of component evaluation. While exhaustive search offers a theoretical global optimum, the exponential growth of candidate models with increasing dimensionality renders it computationally infeasible. Existing automated solutions generally fall into two categories: modular stepwise strategies (e.g., Pharmpy), which suffer from inter-module dependency and order bias, and global search strategies like Genetic Algorithms (GA), which can be hindered by epistasis (interactions between model components) and may favor overly simple models. Furthermore, current tools often rely on proprietary software (NONMEM) and lack rigorous performance benchmarks, such as comparison against exhaustive search ground truths across diverse datasets.

Methodology
This study developed a fully automated, end-to-end framework within the open-source R ecosystem (nlmixr2) to address these gaps. The framework integrates automated model building, initial estimate generation, parameter estimation, and model selection.

• Experimental Design: The performance of four algorithms was evaluated against 144 simulated datasets. These datasets were generated from 48 prespecified PK models (covering intravenous and oral routes, 1–3 compartments, various residual error types, and inter-individual variability [IIV] structures) combined with three sampling designs: optimal, rich, and sparse.
• Ground Truth: An exhaustive search was performed over the candidate model space (711 intravenous and 1,422 oral models) for each dataset. The model with the best fitness value (Bayesian Information Criterion with diagnostic penalties) was treated as the ground-truth reference.
• Algorithms Evaluated:
  1. Stepwise Algorithm: A conventional modular approach mimicking manual workflows, optimizing structural, IIV, correlation, and residual error modules sequentially.
  2. Genetic Algorithm (GA): Encodes models as binary strings, utilizing tournament selection, crossover, mutation, and a 1-bit local search.
  3. Ant Colony Optimization (ACO): Frames model selection as a path-selection problem on a graph. It employs a rank-based pheromone updating strategy, elitism, and a minimum-probability path mechanism to maintain diversity.
  4. Tabu Search (TS): A neighborhood search algorithm using a tabu list to prevent cycling and a perturbation mechanism to escape local optima when no improving neighbors are found.
• Evaluation Metrics: Performance was measured by retrieval rates (structural model, IIV, correlations, residual error, and overall model concordance) and computational efficiency (runtime and number of unique models visited).

Key Contributions

1. Development of nlmixr2auto: The authors introduced a new open-source R package that provides a generalized modeling pipeline for automated PopPK development, lowering the barrier to entry for automated modeling outside proprietary environments.
2. Introduction of ACO and TS to PopPK: While GA has been explored previously, this study systematically introduced and evaluated Ant Colony Optimization and Tabu Search for automated NLME model selection.
3. Rigorous Benchmarking: Unlike prior studies that often compared algorithms only to original designs or fitness function decreases, this study used exhaustive search as a ground truth across a large, diverse set of simulated scenarios to objectively assess algorithmic robustness.
4. Comprehensive Sensitivity Analysis: The study conducted detailed parameter sensitivity analyses for GA and ACO to establish optimal configurations for PopPK applications.

Results

• Overall Performance: Metaheuristic algorithms significantly outperformed the traditional stepwise approach. Tabu Search (TS) and Ant Colony Optimization (ACO) achieved the highest overall model retrieval rates (91% and 90%, respectively), followed by GA (87%). The stepwise algorithm achieved a retrieval rate of approximately 60%.
• Component-Level Accuracy: For specific model components (structural, IIV, correlations, residual error), the metaheuristic algorithms consistently achieved retrieval rates between 92% and 98%. In contrast, the stepwise algorithm ranged from 73% to 92%, with notably lower performance in structural model identification (≤75%).
• Impact of Data Richness:
  • Under optimal and rich data conditions, all metaheuristics maintained high component retrieval rates (90–100%).
  • Under sparse data conditions, performance diverged. TS maintained high accuracy (88–100%), followed by ACO (88–97%), while GA dropped to 72–90%. The authors attribute GA's lower performance in sparse scenarios to its random initialization, whereas ACO and TS could utilize predefined starting points to converge more stably.
  • Oral administration scenarios generally yielded lower full-model retrieval rates than intravenous scenarios, particularly for the stepwise algorithm and GA under sparse conditions.
• Efficiency: While metaheuristic algorithms required approximately 4–4.6 hours per dataset (evaluating ~110–120 unique models, covering ~15–17% of the search space), the stepwise algorithm was significantly faster (<30 minutes) but evaluated far fewer models (<1% of the search space).

Significance and Claims
The paper claims that metaheuristic algorithms, specifically TS and ACO, offer a superior alternative to traditional stepwise methods for automated PopPK model selection. The findings suggest that these algorithms provide higher accuracy, greater robustness in low-information (sparse data) scenarios, and broader coverage of the model search space. The study positions the nlmixr2auto package as a practical, open-source solution that promotes objectivity and standardization in pharmacometric workflows. The authors conclude that integrating these metaheuristic strategies into next-generation pharmacometric toolchains is a viable path forward, though they acknowledge limitations regarding the use of simulated data and the computational cost of exploring larger model spaces, suggesting future work should focus on refining search strategies (e.g., multi-bit neighborhood searches, niching) to maintain efficiency without sacrificing accuracy.