Noise-Resistant Algorithms for the Optimization via Simulation of In Vitro Pharmacology Protocols in Cancer Treatment

Optimization via Simulation applied to biological systems is a powerful tool for conducting large-scale, automated in silico experiments aimed at improving real-world processes. This field presents significant challenges, as the complexity of the simulated biological systems results in complex, high-dimensional, black-box functions that are difficult to optimize algorithmically, alongside the substantial computational cost of repeated simulations. A preliminary phase of this work addresses the algorithm decision problem for non-convex function optimization, proposing a model that estimates the performance of different optimization meta-heuristics on arbitrary functions using statistical Fitness Landscape Analysis measures. The study then applies a multi-scale model of tumor growth and cell resistance to treatments to optimize the delivery strategy of tumor necrosis factors (TNF). The specific challenges of the multi-scale model are identified and then addressed by developing three custom algorithms. A hybrid approach combines a population-based algorithm for broad exploration of the search space with a noise-resistant single-state algorithm for refining promising solutions. Two population-based algorithms are adapted to address the specific challenges posed by the model, resulting in two noise-resistant methods that proved able to efficiently optimize the problem even with a limited computational budget.