Abstract
Substance addiction, particularly involving opioids and cocaine, represents a significant public health crisis with high relapse rates and limited therapeutic options. This study introduces a novel computational framework that bridges transcriptomic analysis and drug discovery through multiscale topological data analysis. By applying Persistent Spectral Theory (PST) to protein-protein interaction (PPI) networks, the researchers identified pivotal genetic targets that maintain functional network integrity. These targets were then used to screen the DrugBank database using advanced machine learning models. The pipeline identified several promising drug candidates, such as Sirolimus and Ozanimod, which demonstrate high binding affinity and favorable safety profiles, offering a robust path for anti-addiction drug repurposing.
Introduction
The escalating drug addiction crisis necessitates the development of new therapeutic strategies. Traditional drug target identification often relies on Differentially Expressed Genes (DEGs) and simple network centrality measures. However, these methods frequently overlook the complex, high-dimensional "shape" of the interactome and the multiscale nature of gene interactions. This paper addresses these limitations by utilizing multiscale topology to identify nodes that are structurally essential to the network's functional integrity, rather than just those with the highest number of connections.
Methodology
The study employs a multi-stage framework designed to move from raw gene expression data to validated drug candidates.
Topological Differentiation
The core innovation lies in the use of Persistent Spectral Theory (PST). Instead of static network analysis, the authors performed "topological perturbation analysis." This involves systematically deleting genes from the PPI network and measuring the resulting changes in the network's topology and geometry across multiple scales using Persistent Laplacians. This approach captures how a gene's absence affects the multiscale "holes" (voids) and connectivity of the interactome.
Target Identification and ML Predictors
By intersecting topologically significant genes across various interaction thresholds, the authors identified three primary downstream targets: mTOR, mGluR5, and NMDAR. To find drugs for these targets, Gradient Boosted Decision Tree (GBDT) models were trained on bioactivity data from the ChEMBL database. These models utilized a consensus of three molecular fingerprints: Transformer-based embeddings (NLP for chemical structures), Autoencoder-based latent vectors, and traditional 2D ECFP fingerprints.
Results and Key Insights
The transcriptomic analysis of heroin and cocaine users revealed distinct genetic signatures that nonetheless converged on shared downstream pathways, such as the MAPK signaling cascade.
Predictive Performance and Candidates
The machine learning models achieved high predictive accuracy, with Pearson correlation coefficients ($R$) ranging from 0.754 to 0.876 across the three targets. Screening the DrugBank database yielded several high-priority candidates:
- mTOR: Beyond known inhibitors like Sirolimus, the study identified Copanlisib and Delafloxacin.
- mGluR5: Doravirine (an HIV drug) and Ozanimod (an MS treatment) emerged as potential inhibitors.
- NMDAR: Gavestinel and Perzinfotel were highlighted as strong antagonists with favorable drug-like properties.
Subsequent ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity) analysis and molecular docking confirmed that many of these candidates, including Dipraglurant and Lersivirine, possess well-balanced safety profiles and strong binding potential.
Discussion and Future Work
This study demonstrates that topology can serve as a powerful proxy for biological importance in complex networks. By looking at the interactome at multiple resolutions, the authors captured structural features that traditional centrality measures miss.
While the computational results are rigorous, the identified candidates require in vivo validation through animal or clinical trials to confirm their efficacy in reducing addictive behaviors. Future research could extend this topological framework to single-cell RNA sequencing (scRNA-seq) data, providing higher resolution for cell-type-specific drug targeting in addiction research.
Summary updated