Machine Learning Study of the Extended Drug-target Interaction Network informed by Pain Related Voltage-Gated Sodium Channels

Guo-Wei Wei, Long Chen, Jian Jiang, Bozheng Dou, Hongsong Feng, Jie Liu, Yueying Zhu, Ben-gong Zhang, Tianshou Zhou, Guo-Wei Wei 0001arXiv 2023 · DOI · arXiv

Abstract

Chronic pain management remains a critical challenge in modern medicine, primarily due to the addictive potential of opioids and the systemic side effects of non-steroidal anti-inflammatory drugs (NSAIDs). This study presents a machine learning framework designed to identify non-addictive pain relief candidates by targeting voltage-gated sodium channels (Navs). By constructing an extended drug-target interaction (DTI) network and utilizing natural language processing (NLP) for molecular representation, we screened over 150,000 compounds. Our methodology identified two promising lead compounds with high specificity for peripheral Nav channels and low predicted cardiotoxicity, offering a computationally driven path toward safer analgesics.

Introduction

Chronic pain affects approximately 35% of the U.S. population. Current treatments are often inadequate; opioids carry significant risks of addiction and respiratory depression, while NSAIDs are associated with gastrointestinal and cardiovascular complications. Voltage-gated sodium channels (specifically subtypes Nav1.3, 1.7, 1.8, and 1.9) represent a high-value therapeutic target because they are primarily expressed in the peripheral nervous system. Inhibiting these channels can theoretically block pain signals before they reach the central nervous system, thereby avoiding addictive side effects.

However, the structural similarity between Nav subtypes and other critical proteins, such as the hERG potassium channel (essential for heart rhythm), makes the discovery of selective inhibitors a "needle in a haystack" problem. Traditional high-throughput screening is often too slow and costly to navigate this complex landscape effectively.

Methodology

Network Construction and Molecular Embedding

To address the complexity of the target space, we moved beyond single-target analysis to construct an extended Drug-Target Interaction (DTI) network. Using the String Database (v11), we built Protein-Protein Interaction (PPI) networks for the four key Nav channels, identifying over 1,000 functionally related proteins. This network-centric approach allows for the detection of potential off-target effects early in the discovery process.

Overview of the methodology, including PPI network construction and the resulting drug-target interaction mapping.

To represent chemical structures, we employed advanced NLP-based embedding techniques, treating SMILES strings as a chemical language. Two primary models were used:

  1. Bidirectional Encoder Transformer (BET-FP): A self-supervised model pre-trained on millions of molecules to capture deep structural features.
  2. Sequence-to-Sequence Autoencoder (AE-FP): A model that compresses molecular data into a 512-dimension latent vector.

Machine Learning Ensemble

These embeddings served as inputs for an ensemble of machine learning algorithms, including Gradient Boosting Decision Trees (GBDT), Support Vector Machines (SVM), and Random Forests (RF). The final binding affinity predictions were derived from a consensus of these models, which consistently outperformed individual algorithms. Promising candidates were further filtered using ADMETlab 2.0 for pharmacokinetic properties and virtually docked using AutoDock Vina to confirm physical binding modes.

Results and Discussion

Predictive Accuracy and Network Insights

The ensemble models demonstrated high reliability, achieving a mean Pearson correlation ($R$) of 0.77 across the 111-target dataset, with top-performing models reaching $R = 0.93$. The heatmap of the extended DTI network reveals the broad scope of the predictive framework, highlighting how specific ligands interact across a wide array of protein targets.

Heatmap showing the predicted binding affinities (ML-BA) and model performance (Pearson R) across the extended drug-target interaction network.

A key insight from this network analysis is the correlation of cross-target activity. If a model predicts similar binding patterns for two different proteins, it often indicates shared structural motifs, which can be used to predict "mutual side effects" or identify drug repurposing opportunities. For instance, the study suggests that inhibitors originally designed for MAPK8 or CASP3 might be repurposed as potent Nav1.7/1.8 blockers.

Lead Discovery

The framework identified two "nearly optimal" lead compounds: CHEMBL1767278 and CHEMBL1453498. These molecules exhibited strong predicted binding to Nav1.7 and Nav1.9 while maintaining favorable ADMET profiles. Crucially, they showed low predicted affinity for the hERG channel, suggesting a reduced risk of cardiotoxicity compared to other candidates.

Conclusion

By integrating NLP-based molecular embeddings with extended protein interaction networks, this study provides a robust computational pipeline for analgesic discovery. While these in silico results require in vitro and in vivo validation, the identification of high-specificity leads and the mapping of potential side-effect pathways represent a significant step toward non-addictive pain management. Future work will focus on refining models for targets with sparse experimental data and expanding the network to include broader signaling pathways.

Summary updated

BibTeX
@misc{wei2023machine,
  author = {Guo-Wei Wei and Long Chen and Jian Jiang and Bozheng Dou and Hongsong Feng and Jie Liu and Yueying Zhu and Ben-gong Zhang and Tianshou Zhou and Guo-Wei Wei},
  title = {{Machine Learning Study of the Extended Drug-target Interaction Network informed by Pain Related Voltage-Gated Sodium Channels}},
  year = {2023},
  eprint = {2307.05794},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2307.05794},
}
Figure