An Introduction to Recursive Neural Networks and Kernel Methods for Cheminformatics

Abstract

The aim of this paper is to introduce the reader to new developments in Neural Networks and Kernel Machines concerning the treatment of structured domains. Specifically, we discuss the research on these relatively new models to introduce a novel and more general approach to QSPR/QSAR analysis. The focus is on the computational side and not on the experimental one.

A Deep Graph Network–Enhanced Sampling Approach to Efficiently Explore the Space of Reduced Representations of Proteins

“…An essential element of the proposed method is thus a graph-based representation of our object of interest, namely a protein. With their long and successful story both in the field of coarse-graining ( Gfeller and Rios, 2007 ; Webb et al, 2019 ; Li et al, 2020 ) and in the prediction of protein properties ( Borgwardt et al, 2005 ; Ralaivola et al, 2005 ; Micheli et al, 2007 ; Fout et al, 2017 ; Gilmer et al, 2017 ; Torng and Altman, 2019 ), graph-based learning models represent a rather natural and common choice to encode the (static) features of a molecular structure; here, we show that a graph-based machine learning approach can reproduce the results of mapping entropy estimate obtained by means of a much more time-consuming algorithmic workflow. To this end, we rely on Deep Graph Networks (DGNs) ( Bacciu et al, 2020 ), a family of machine learning models that learn from graph-structured data, where the graph has a variable size and topology; by training the model on a set of tuples (protein, CG mapping, and S map ), we can infer the S map values of unseen mappings associated with the same protein making use of a tiny fraction of the extensive amount of information employed in the original method, i.e., the molecular structure viewed as a graph.…”

Generative Kernels for Tree-Structured Data

“…For example, parse trees arise in natural language processing tasks where a parse tree or a semantic related tree structure is generated starting from a sentence [1], [2]; moreover, tree-like representations/patterns can be naturally derived, for example, from documents (e.g. [3]) and HTML/XML documents in information retrieval [4], [5], [6], structured network data in computer security [7], molecule structures in computational chemistry [8], [9], and image analysis. In all these application domains, learning plays a crucial role since very often the user is interested in automatic classification/regression tasks where, starting from a set of labeled instances, a classifier/regressor is pursued.…”

Recursive Neural Networks for Undirected Graphs for Learning Molecular Endpoints

“…The algorithms we describe are based on recursive neural networks and they deal with molecules directly as graphs, in that no features are manually extracted from the structure, and the networks automatically identify regions and substructures of the molecules that are relevant for the property in question. The basic structural processing cell we use is similar to those described in [15,16,17,18], and adopted in essentially the same form in applications including molecule regression/classification [19,20,21], image classification [22], natural language processing [23], face recognition [24]. In the case of molecules, there are numerous disadvantages in these earlier models: they can only deal with trees, thus molecules (that are more naturally described as Undirected Graphs (UG)) have to be preprocessed before being input; the preprocessing is generally task-dependent; special nodes ("super-sources") have to be defined for each molecule; application domains are generally limited, thus the effectiveness of the models is hard to gauge.…”