Combinatorial Library Neural Network (CoLiNN) for Combinatorial Library Visualization Without Compound Enumeration

The development of vast virtual combinatorial spaces and libraries, built from commercially accessible building blocks, introduced a novel approach to identifying hits. Hence, there is a clear need for ‘big data’ compatible chemoinformatics methods to analyze such vast combinatorial compound collections. For example, a library can be characterized by its data distribution on a 2D map. Generative Topographic Mapping (GTM) is particularly well-suited for chemical library visualization and analysis due to its probabilistic nature and Big Data compatibility. Conventionally, preparation of a GTM of a chemical library implies (a) enumeration of compounds using available Building Block (BB) and a set of reaction rules, (b) molecular descriptors calculation, and (c) data projection on GTM. Aiming to bypass this time-consuming workflow, we propose a Combinatorial Library Neural Network (CoLiNN) model which uses a Graph Convolutional Network (GCN) to predict data projections on GTM, solely based on the information about the reaction and BB sets used for the library preparation. Ten DNA-Encoded Combinatorial Libraries (DELs), each containing one million compounds, were used to train and evaluate the model. The DEL compound projections inferred by CoLiNN closely resemble the true projections derived by conventional enumeration/projection protocol. Thus, the CoLiNN model enables the accurate mapping of large combinatorial libraries without their explicit enumeration, significantly speeding up the analysis of their chemical space.