Exploring the Predictive Efficacy of the K-Nearest Neighbors Algorithm Using Genetic and Pharmacological Data From Cancer Cell Lines

Although heterogeneity in cancer cell response to drugs continues to represent a significant barrier to anti-cancer agent development and testing, machine learning, particularly KNN models, has shown promise in prediction using pharmacological and genetic profiles. However, optimization of KNN parameters, notably neighborhood size or k, and unforeseen structural trends and limitations remain as key questions in application to cancer cell lines. This study aims to examine patterns, optimizations, and relationships to biological intuition and relevance in the context of drug sensitivity and cancer cell heterogeneity in this limited investigation of KNN considering only the drug Camptothecin as a reference with a blindly chosen cell line, EW-11. It was found that smaller k values leading to disproportionately inaccurate results could likely be attributed to molecularly similar yet pharmacologically different cell lines. A degree of robustness in ranking, insensitive to even doubling intuitive parameters such as mutations, was observed, and an increased Euclidean distance to distinguish similar lines had a positive effect. A k of 14 was found to be optimized for cell line EW-11. Yet, this was not generalizable to other randomly chosen cell lines, likely due to different biological cluster sizes. However, a k of 14 yielded a relatively accurate prediction for the nearest neighbor of EW-11, supporting the notion that an optimized k value accounts for the size of a local subgroup. Further investigations are necessary to expand these implications and continue to refine the capacities of a KNN model.