Enhancing ERα-targeted compound efficacy in breast cancer threapy with ExplainableAI and GeneticAlgorithm

Abstract

Breast cancer remains a major cause of mortality among women globally, driving the need for advanced therapeutic solutions. This study presents a novel, comprehensive methodology integrating explainable artificial intelligence (AI), machine learning models, and genetic algorithms to enhance the bioactivity and ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity) properties of compounds targeting estrogen receptor alpha (ERα). By employing SHAP (SHapley Additive exPlanations) and LassoNet, we identified and refined 50 critical molecular descriptors from an initial set of 729, significantly influencing the prediction of bioactivity. The selected descriptors were systematically validated, bolstering the predictive robustness of our models, which demonstrated a mean coefficient of determination of 77% for bioactivity and high accuracy scores of 90.2%, 93.7%, 89.5%, 87.3%, and 95.8% for absorption, distribution, metabolism, excretion, and toxicity, respectively. Further optimization through genetic algorithms identified candidate compounds with superior bioactivity, achieving pIC50 values as high as 10.05, surpassing the previously observed peak values in the dataset. These results underscore the potential of leveraging advanced machine learning and optimization techniques to accelerate the discovery of effective cancer therapies.