De Novo Design of Molecules with Multiaction Potential from Differential Gene Expression using Variational Autoencoder

Abstract

The modulating effect of chemical compounds and therapeutics on gene transcription is well-reported and has been intensively studied for both clinical and research purposes. Emerging research points toward the utility of drug-induced transcriptional alterations in de novo molecular design and highlights the idea of phenotype-matching an expression signature of interest to the structures being designed. In this work, we build an autoencoder-based generative model, BiCEV, around this concept. Our generative autoencoder has demonstrably generated a set of new molecules from gene expression input with notable validity (96%), uniqueness (98%), and internal diversity (0.77). Further, we attempted to validate BiCEV by testing the model on gene-knockdown profiles and combined signatures of synergistic drug pairs. From these investigations, we found the designed structures to be consistently high in collective quality. However, when their similarities to the supposed functional equivalents as determined by shared targets were considered, the findings were somewhat mixed. In spite of this, we believe the generative model merits further development in conjunction with in vitro corroboration to lend itself to being an assistive tool for drug discovery experts, particularly to support the initial stages of hit identification and lead optimization.

TransGEM: a molecule generation model based on Transformer with gene expression data

“…The LINCS1000 database comprises data of five levels, with level 3 data being utilized for model training and application, as opposed to the previously common usage of level 5 data in prior research. This is mainly because previous studies have shown that level 5 data exhibits high levels of noise when used to characterize gene expression profiles ( Pham et al 2022 , Das et al 2023 , Pravalphruekul et al 2023 ). The level 3 data of the LINCS1000 database represent the standardized and normalized results of experimentally measured fluorescence intensity values ( Subramanian et al 2017 ).…”

“…However, the proportion of valid molecules generated by this model was relatively low, only 8.2% ( Méndez-Lucio et al 2020 ). Furthermore, Adversarial Autoencoder (AAE) ( Shayakhmetov et al 2020 ), Variational Autoencoder (VAE) ( Das et al 2023 , Pravalphruekul et al 2023 ), and fragment-based generative model ( Pham et al 2022 ) have been used to generate new molecules from gene expression data. These above models generated a greater number of valid molecules than GAN, and these methods designed molecules based on gene knocked-out transcriptomic profiles.…”

“…These above models generated a greater number of valid molecules than GAN, and these methods designed molecules based on gene knocked-out transcriptomic profiles. Nevertheless, molecules generated from these knock-out transcriptomic profiles exhibited a notable lack of correlation with known inhibitors of the targeted genes ( Pham et al 2022 , Das et al 2023 , Pravalphruekul et al 2023 ). PaccMann RL ( Born et al 2021 ), a hybrid model combined VAE with reinforcement learning, was designed to generate molecules based on target-specific transcriptomic data.…”

“…The cost-effective gene-expression experiments led to development of various open-source transcriptomics databases, such as the Gene Expression Omnibus ( Clough and Barrett 2016 ), The Cancer Genome Atlas ( Tomczak et al 2015 ), Connectivity Map ( Lamb et al 2006 ), and Library of Integrated Network-based Cellular Signatures 1000 landmark genes (LINCS1000) ( Subramanian et al 2017 ). Among them, the LINCS1000 database has been used in multiple studies for molecular generation based on gene expression data ( Pham et al 2022 , Das et al 2023 , Pravalphruekul et al 2023 ). LINCS1000 database comprises gene expression data of various cell lines subjected to different perturbation conditions, including small molecule compound addition, gene knockout, gene overexpression, and other treatments ( Subramanian et al 2017 ).…”

“…LINCS1000 database comprises gene expression data of various cell lines subjected to different perturbation conditions, including small molecule compound addition, gene knockout, gene overexpression, and other treatments ( Subramanian et al 2017 ). The LINCS1000 database comprises gene expression data at five levels of various perturbed cell lines, but only level 5 data have been used in the existing studies ( Pham et al 2022 , Das et al 2023 , Pravalphruekul et al 2023 ). However, level 5 data have been demonstrated to exhibit noise when considered as gene expression data ( Pham et al 2021 ).…”