Searching the chemical space for effective magnesium dissolution modulators: a deep learning approach using sparse features

Small organic molecules can alter the degradation rates of the magnesium alloy ZE41. However, identifying suitable candidate compounds from the vast chemical space requires sophisticated tools. The information contained in only a few molecular descriptors derived from recursive feature elimination was previously shown to hold the potential for determining such candidates using deep neural networks. We evaluate the capability of these networks to generalise by blind testing them on 15 randomly selected, completely unseen compounds. We find that their generalisation ability is still somewhat limited, most likely due to the relatively small amount of available training data. However, we demonstrate that our approach is scalable; meaning deficiencies caused by data limitations can presumably be overcome as the data availability increases. Finally, we illustrate the influence and importance of well-chosen descriptors towards the predictive power of deep neural networks.