Fig. 1: System illustration of the proposed QMO framework.
From: Optimizing molecules using efficient queries from property evaluations
The QMO system progressively optimizes an input lead molecule (for example, remdesivir) according to a set of user-specified properties (for example, binding affinity and Tanimoto similarity) by leveraging the learned molecule embeddings from a pre-trained encoder and decoder pair (that is, an autoencoder) and by evaluating the properties of the generated molecules. Given a candidate embedding zt at optimization step t, QMO randomly samples the neighbouring vectors of zt in the embedding space, evaluates the properties of the corresponding decoded molecules, and uses the evaluations for gradient estimation (equation (4)) and query-based gradient descent (equation (3)) to find the next candidate embedding vector zt+1.
