Transparent AI methods for drug design based on group equivariant non expansive operators

Equivariant operators are proving to be increasingly important in deep learning, in order to make neural networks more transparent and interpretable. The use of such operators corresponds to the rising interest in the so called “explainable artificial intelligence”, which looks for methods and techniques whose functioning can be understood by humans. In accordance with this line of research, Group Equivariant Non-Expansive Operators (GENEOs) have been recently proposed as elementary components for building new kinds of networks. Their use is grounded in Topological Data Analysis (TDA) and guarantees good mathematical properties to the involved spaces, such as compactness, convexity, and finite approximability, under suitable assumptions on the space of data and by choosing appropriate topologies. In this talk we will show promising results obtained by applying GENEOs to protein pocket detection. We will also discuss the robustness and trasnparency of the method developed for this driving application, and the possibility to generalize it to other applications, from computer vision to agriculture, or to other types of AI architectures.