Adversarial Threshold Neural Computer for Molecular de Novo Design

Abstract

In this article, we propose the deep neural network Adversarial Threshold Neural Computer (ATNC). The ATNC model is intended for the de novo design of novel small-molecule organic structures. The model is based on generative adversarial network architecture and reinforcement learning. ATNC uses a Differentiable Neural Computer as a generator and has a new specific block, called adversarial threshold (AT). AT acts as a filter between the agent (generator) and the environment (discriminator + objective reward functions). Furthermore, to generate more diverse molecules we introduce a new objective reward function named Internal Diversity Clustering (IDC). In this work, ATNC is tested and compared with the ORGANIC model. Both models were trained on the SMILES string representation of the molecules, using four objective functions (internal similarity, Muegge druglikeness filter, presence or absence of sp-rich fragments, and IDC). The SMILES representations of 15K druglike molecules from the ChemDiv collection were used as a training data set. For the different functions, ATNC outperforms ORGANIC. Combined with the IDC, ATNC generates 72% of valid and 77% of unique SMILES strings, while ORGANIC generates only 7% of valid and 86% of unique SMILES strings. For each set of molecules generated by ATNC and ORGANIC, we analyzed distributions of four molecular descriptors (number of atoms, molecular weight, logP, and tpsa) and calculated five chemical statistical features (internal diversity, number of unique heterocycles, number of clusters, number of singletons, and number of compounds that have not been passed through medicinal chemistry filters). Analysis of key molecular descriptors and chemical statistical features demonstrated that the molecules generated by ATNC elicited better druglikeness properties. We also performed in vitro validation of the molecules generated by ATNC; results indicated that ATNC is an effective method for producing hit compounds.

KekuleScope: prediction of cancer cell line sensitivity and compound potency using convolutional neural networks trained on compound images

“…Whereas existing computational tools to model in vitro compound activity mostly rely on established algorithms (e.g., Random Forest or Support Vector Machines), the utilization of deep learning in drug discovery is gaining momentum, a trend that is only expected to increase in the coming years [21]. Deep learning techniques have been already applied in numerous drug discovery tasks, including toxicity modelling [22,23], bioactivity prediction [24][25][26][27][28][29][30], and de novo drug design [31][32][33][34], among others. Most of these studies have utilized feedforward neural networks consisting of multiple fully-connected layers trained on one of the many compound descriptors developed over the last >30 years in the chemoinformatics field [27,35].…”

Human microbiome aging clocks based on deep learning and tandem of permutation feature importance and accumulated local effects

“…The renaissance of deep learning that started in 2015 resulted in unprecedented machine learning performance in image, voice, and text recognition, as well as a range of biomedical applications 29 such as drug repurposing 30 and target identification 31 . One of the most impactful applications of DL in biomedicine was in the applications of generative models to de novo molecular design [32][33][34][35][36] . In the context of aging research, these new methods can be combined for geroprotector discovery [37][38][39][40][41] .…”

Inverse‐QSPR for de novo Design: A Review

“…For a drug data set, the ratio of valid molecules decreased from 92 % to 76 % but the output SMILES's uniqueness ratio rose up to 76 % instead of 24 %. The subsequent addition of an AT to the system allowed to increase the validity of the output SMILES …”

“…The subsequent addition of an AT to the system allowed to increase the validity of the output SMILES. [142] As for VAE, other molecules' representations were used with GAN. For example, DruGAN [146] uses MACCS molecular fingerprints as binary vectors and Tanimoto distances as the similarity measure.…”