Discovering new drugs is a long and complex process that often starts with screening thousands of chemical compounds. In recent years, scientists have also explored how RNA sequencing data, which shows how genes are turned on or off in cells, can help identify drugs that reverse disease related changes in gene expression. Most of this work has focused on repurposing existing drugs, but less attention has been given to designing entirely new ones.
Researchers at Michigan State University have developed a new platform called GPS, gene expression profile predictor on chemical structures, to address this challenge. The system uses deep learning to predict how a chemical compound will affect gene expression, based only on its structure.
The key idea is to connect chemical structure with biological response. Instead of testing every compound experimentally, the model predicts how each molecule would change gene activity across the genome. These predicted changes can then be compared with disease associated gene expression patterns to identify compounds that may reverse harmful effects.
The GPS framework and performance
(A) An AI model for compound-induced transcriptomic signature prediction is developed using an RCL strategy. This model enables the prediction of transcriptomic signatures of a large library of drug-like compounds based solely on chemical structures. The predicted profiles are ranked by their ability to reverse the expression of a signature derived from disease or phenotype profiles. Selected candidates are tested in vitro to identify lead compounds, which are subsequently optimized for improved reversal scores and favorable medicinal chemistry properties through a Monte Carlo tree-search strategy. Final candidates are evaluated in preclinical models and SGARs are delineated for illustrating the MoA. In this study, the ZINC library and the Enamine HTS library are used to screen compounds for HCC and IPF, respectively. (B) The averaged F1 scores and balanced accuracies of 10 independent runs for HEPG2, MCF7, PC3, and VCAP under various comparison methods in an external validation. The tested drug-induced gene expression profiles are from LINCS 2020 at a concentration of 10 μM and treatment for 24 h. Only high-quality profiles and those compounds excluded from the training set are considered. Note that in the ablation study, DeepCop was trained on all profiles without quality improvement, whereas RCL was trained using the same architecture and dataset but with corrected data. DeepCop, deep gene compound profiler; RF, random forest; HQ, high-quality; MTL, multi-task learning; RCL, robust collaborative learning; and SG, single gene.
To improve performance, the team incorporated additional strategies, including refined scoring methods and a tree search optimization process. This allows the system not only to screen large libraries of compounds but also to optimize new molecules step by step, improving their potential effectiveness.
The platform was tested across multiple diseases. In hepatocellular carcinoma, a type of liver cancer, the researchers identified two new groups of compounds that showed strong selectivity and effectiveness in both cell based experiments and animal models. In idiopathic pulmonary fibrosis, a disease involving tissue scarring in the lungs, the system identified both a repurposed drug candidate and a newly designed compound with antifibrotic activity.
An important aspect of the approach is its use of single cell RNA sequencing data, which captures gene expression changes in specific cell types. This allows the model to account for complex biological environments where different cell populations respond differently to treatment.
By linking RNA sequencing data with chemical design, the GPS platform provides a new way to discover drugs that target the underlying molecular mechanisms of disease. It moves beyond simply finding existing compounds and opens the door to designing new therapeutics tailored to specific gene expression signatures.
As biological datasets continue to grow, combining artificial intelligence with RNA sequencing could help accelerate drug discovery and improve the precision of future treatments.
Availability – https://github.com/Bin-Chen-Lab/GPS
Xing J, Tan M, Leshchiner D, Sun M, Abdelgied M, Huang L, Paithankar S, Uhl K, Shankar R, Lisabeth E, Aleiwi B, Jager T, Lawson C, Chen R, Giletto M, Girgis R, Neubig R R, So S, Ellsworth E, Li X, Chua M S, Zhou J, Chen B. (2026) Deep-learning-based de novo discovery and design of therapeutics that reverse disease-associated transcriptional phenotypes. Cell [Epub ahead of print]. [article]
Related Posts
Discovering new drugs is a long and complex process that often starts with screening thousands of chemical compounds. In recent years, scientists have also explored how RNA sequencing data, which shows how genes are turned on or off in cells, can help identify drugs that reverse disease related changes in gene expression. Most of this work has focused on repurposing existing drugs, but less attention has been given to designing entirely new ones.
Researchers at Michigan State University have developed a new platform called GPS, gene expression profile predictor on chemical structures, to address this challenge. The system uses deep learning to predict how a chemical compound will affect gene expression, based only on its structure.
The key idea is to connect chemical structure with biological response. Instead of testing every compound experimentally, the model predicts how each molecule would change gene activity across the genome. These predicted changes can then be compared with disease associated gene expression patterns to identify compounds that may reverse harmful effects.
The GPS framework and performance
(A) An AI model for compound-induced transcriptomic signature prediction is developed using an RCL strategy. This model enables the prediction of transcriptomic signatures of a large library of drug-like compounds based solely on chemical structures. The predicted profiles are ranked by their ability to reverse the expression of a signature derived from disease or phenotype profiles. Selected candidates are tested in vitro to identify lead compounds, which are subsequently optimized for improved reversal scores and favorable medicinal chemistry properties through a Monte Carlo tree-search strategy. Final candidates are evaluated in preclinical models and SGARs are delineated for illustrating the MoA. In this study, the ZINC library and the Enamine HTS library are used to screen compounds for HCC and IPF, respectively. (B) The averaged F1 scores and balanced accuracies of 10 independent runs for HEPG2, MCF7, PC3, and VCAP under various comparison methods in an external validation. The tested drug-induced gene expression profiles are from LINCS 2020 at a concentration of 10 μM and treatment for 24 h. Only high-quality profiles and those compounds excluded from the training set are considered. Note that in the ablation study, DeepCop was trained on all profiles without quality improvement, whereas RCL was trained using the same architecture and dataset but with corrected data. DeepCop, deep gene compound profiler; RF, random forest; HQ, high-quality; MTL, multi-task learning; RCL, robust collaborative learning; and SG, single gene.
To improve performance, the team incorporated additional strategies, including refined scoring methods and a tree search optimization process. This allows the system not only to screen large libraries of compounds but also to optimize new molecules step by step, improving their potential effectiveness.
The platform was tested across multiple diseases. In hepatocellular carcinoma, a type of liver cancer, the researchers identified two new groups of compounds that showed strong selectivity and effectiveness in both cell based experiments and animal models. In idiopathic pulmonary fibrosis, a disease involving tissue scarring in the lungs, the system identified both a repurposed drug candidate and a newly designed compound with antifibrotic activity.
An important aspect of the approach is its use of single cell RNA sequencing data, which captures gene expression changes in specific cell types. This allows the model to account for complex biological environments where different cell populations respond differently to treatment.
By linking RNA sequencing data with chemical design, the GPS platform provides a new way to discover drugs that target the underlying molecular mechanisms of disease. It moves beyond simply finding existing compounds and opens the door to designing new therapeutics tailored to specific gene expression signatures.
As biological datasets continue to grow, combining artificial intelligence with RNA sequencing could help accelerate drug discovery and improve the precision of future treatments.
Availability – https://github.com/Bin-Chen-Lab/GPS
Xing J, Tan M, Leshchiner D, Sun M, Abdelgied M, Huang L, Paithankar S, Uhl K, Shankar R, Lisabeth E, Aleiwi B, Jager T, Lawson C, Chen R, Giletto M, Girgis R, Neubig R R, So S, Ellsworth E, Li X, Chua M S, Zhou J, Chen B. (2026) Deep-learning-based de novo discovery and design of therapeutics that reverse disease-associated transcriptional phenotypes. Cell [Epub ahead of print]. [article]
Related Posts
Recent Posts
Scientists identify the neuroblastoma cells that signal poor outcomes
GARAGE improves rare cell detection in single-cell RNA sequencing analysis using AI
Machine learning reveals how RNA splicing is regulated in individual cell types
Illumina releases SpliceAI2 to help advance rare disease research
Postdoctoral Fellow – Long-Read RNA Methods Development for Neurodegeneration
Novel biosensor uses DNA ‘barcodes’ to uncover disease signals in cells
RNA-Seq Products & Services
Stay Connected