Mobile Menu

DeepCE: A deep-learning framework for screening novel chemicals

Researchers have proposed a new mechanism – DeepCE – driven by deep learning, that provides a robust predictive model for screening novel chemicals.


The drug discovery process is dominated by target-based high-throughput screening. It has been the focus of computer-aided drug discovery for decades. However, the readout from the modulation of a single protein by a chemical poorly correlates with therapeutic effects or side effects. This leads to high failure rates. Phenotype-based screening has generated interest for identifying cell-active compounds. However, this approach has suffered from low throughput and difficulty in target deconvolution.

Systematic analysis of genome-wide gene expression of chemical perturbations has led to considerable improvements in drug discovery. In particular, its applications in drug repurposing, discovering drug mechanisms, identifying lead compounds and predicting side effects for preclinical compounds. Nonetheless, the use of such data is limited by their sparseness, unreliability and relatively low throughput. In addition, few methods can perform phenotype-based de novo chemical compound screening.


In this paper, published in Nature Machine Intelligence, researchers designed a mechanism-driven neural network-based model – DeepCE. This model captures high-dimensional associations among biological features and also non-linear relationships between biological features and outputs. This enables the model to predict gene expression profiles when given new chemical compounds.

The team found that DeepCE outperformed state-of-the-art models for predicting gene expression profiles in the L1000 dataset. This dataset consists of ~1,400,000 gene-expression profiles on the responses of ~50 human cell lines to one of ~20,000 compounds across a range of concentrations. The team saw this response not only in a de novo chemical setting but also in a traditional imputation setting.

To demonstrate the value of this model, the team then applied it to drug repurposing of COVID-19. They did this by in silico screening of all chemical compounds in DrugBank against COVID-19 patient clinical phenotypes. From this, they generated novel lead compounds consistent with clinical evidence.

These findings suggest that DeepCE could be a powerful tool for phenotype-based compound screening, including in cases of urgency.

Image credit: By ilkaydede –

More on these topics

Deep Learning / drug discovery / Screening

Share this article