Metal complexes, whether employed as prodrugs or as active pharmaceutical agents, have demonstrated considerable potential as drugs for the inhibition of biochemical targets. However, a substantial lack in modelling efforts can be observed, when compared to purely organic drugs. In particular, emerging computational tools based on machine learning predominantly focus on organic compounds, while they would be much needed for studying the inhibition of proteins by metal complexes. With this project we aim to bridge this gap and initiate the development of machine learning protocols for the design of metallodrugs.
During my visit I will perform a comparative study to determine the most suitable computational representation (descriptor) for metal complexes. Simultaneously, I will be delving into the exploration of regression and classification models. These models are designed to predict continuous values, such as IC50 values, or classify compounds as either active or inactive. The choice of descriptor plays a crucial role in influencing the performance of these models, underscoring the need for a concurrent approach. Additionally, we intend to use the active-learning protocol for machine learning potentials developed in the group of Prof. Dr. F. Duarte for metal-containing compounds. These potentials can be used to run dynamic simulations that seek to achieve a balance between conventional molecular mechanics and quantum mechanical/molecular mechanical methods, offering an intermediary solution in terms of both accuracy and speed.