Abstract
High-throughput screening and gene signature analyses frequently identify lead therapeutic compounds with unknown modes of action (MoAs), and the resulting uncertainties can lead to the failure of clinical trials. We developed an approach for uncovering MoAs through an interpretable machine learning model of transcriptomics, epigenomics, metabolomics, and proteomics. Examining compounds with beneficial effects in models of Huntington's Disease, we found common MoAs for compounds with unrelated structures, connectivity scores, and binding targets. The approach also predicted highly divergent MoAs for two FDA-approved antihistamines. We experimentally validated these effects, demonstrating that one antihistamine activates autophagy, while the other targets bioenergetics. The use of multiple omics was essential, as some MoAs were virtually undetectable in specific assays. Our approach does not require reference compounds or large databases of experimental data in related systems and thus can be applied to the study of agents with uncharacterized MoAs and to rare or understudied diseases.
Introduction
Developing new therapeutics is a high-risk endeavor, with approximately 86% of drug candidates failing to achieve FDA approval between 2000 and 2015. A primary obstacle is the "black box" problem: phenotypic screening often identifies compounds that produce a desired effect in a laboratory model, yet their molecular Mode of Action (MoA) remains unknown. Traditional computational tools, such as the Connectivity Map (CMap), rely on comparing gene signatures to libraries of known drugs. However, these methods fail when a compound acts through a novel pathway or a mechanism not represented in the database. This study introduces an unbiased, multi-omics framework designed to deduce MoAs from scratch, specifically targeting the complexities of Huntington’s Disease (HD).
Methods: An Integrated Multi-Omics Framework
The researchers developed a hierarchical approach that integrates four layers of biological data: transcriptomics (RNA-Seq), untargeted metabolomics (lipids and polar metabolites), epigenomics (ChIP-Seq), and proteomics/phosphoproteomics.
To interpret this vast amount of data, the team utilized specialized machine learning algorithms, including PIUMet and Omics Integrator (specifically the Prize-Collecting Steiner Forest algorithm). These tools map experimental observations onto a comprehensive "interactome"—a network of known physical interactions between proteins and metabolites. The model identifies the most parsimonious sub-network that connects the observed changes across all omics layers, effectively filtering biological noise and highlighting the most likely therapeutic pathways.
Results
The Limitations of Single-Omics Analysis
A key insight from the study is that transcriptomics alone can be a blunt instrument. While gene expression profiles for different compound groups appeared similar, metabolomic data revealed starkly different cellular effects.
As shown in Figure 6B, the number of significantly changed genes often dwarfs the number of metabolic changes, yet the metabolic data was essential for distinguishing between mechanisms. For instance, two FDA-approved antihistamines, Cyproheptadine and Meclizine, were found to have highly divergent MoAs despite similar primary targets.
Identifying Divergent Modes of Action
The model categorized compounds into two primary functional groups:
- Group A (e.g., Cyproheptadine, Loxapine): Predicted to activate autophagy (the cell's internal recycling system). This was validated by an increased LC3-II/LC3-I ratio in both mouse and human cell lines.
- Group B (e.g., Meclizine, DKI): Predicted to shift bioenergetics. Validation showed these compounds inhibit mitochondrial respiration while increasing glycolysis, maintaining stable ATP production through a metabolic shift.
Crucially, Group A compounds were found to rescue mitochondrial deficits, while Group B compounds addressed glycolysis deficits inherent in HD models.
Discussion and Conclusion
This multi-omics approach demonstrates that compounds with unrelated chemical structures can share identical MoAs, while structurally similar drugs can act through entirely different pathways. By forcing disparate data types to "talk to each other" through a physical interaction network, biologically relevant pathways like autophagy and mitochondrial respiration become clear.
While the approach is data-intensive and identifies pathways rather than specific primary protein targets, it offers a powerful tool for drug repurposing and the study of rare diseases where reference data is scarce. This framework provides a path forward for opening the "black box" of drug discovery, ensuring that promising leads are backed by a deep understanding of their molecular mechanisms.
Summary updated