Proteomic Learning of Gamma-Aminobutyric Acid (GABA) Receptor-Mediated Anesthesia

Guo-Wei Wei, Jian Jiang, Long Chen, Yueying Zhu, Yazhou Shi, Huahai Qiu, Bengong Zhang, Tianshou Zhou, Guo-Wei Wei 0001Journal of Chemical Information and Modeling 2025 · DOI

TL;DR

This paper introduces a novel proteomic learning framework to discover new anesthetic agents targeting GABA receptors. By integrating protein-protein interaction networks with machine learning models and extensive compound databases, the study identifies promising lead compounds with improved efficacy and reduced side effects, paving the way for safer anesthesia.

What Problem Does This Paper Address?

Anesthetics are vital for surgical procedures and therapeutic interventions, but current agents often come with significant limitations. These include undesirable side effects (like respiratory depression and cognitive dysfunction), variable patient responses, and challenges in precise control. Developing new anesthetic agents with enhanced safety, fewer side effects, and greater controllability is a critical unmet need in medicine. GABA receptors, the primary inhibitory neurotransmitters in the central nervous system, are key targets for many anesthetics. However, understanding the complex interactions between drugs and the numerous GABA receptor subtypes, as well as predicting potential off-target effects, remains a significant challenge. Traditional drug development methods are often slow, costly, and inefficient, necessitating more advanced approaches.

Methods

The study employs a multi-faceted approach combining proteomics, machine learning, and virtual screening.

  • Proteomic Learning Framework: The core of the method involves building a proteomic learning system. This starts by constructing protein-protein interaction (PPI) networks for 24 GABA receptor subtypes. These networks help identify potential therapeutic targets and sites that might cause side effects.
  • Machine Learning Models: Three machine learning algorithms (Gradient Boosting Decision Tree, Support Vector Machine, and Random Forest) were used. Specifically, Support Vector Machines (SVM) were found to be the most effective for the datasets.
  • Molecular Embeddings: To represent molecules, two types of embeddings were generated:
  • Bidirectional Encoder Transformer (BET-FP): A pre-trained NLP model that encodes SMILES strings into latent embedding vectors.
  • Sequence-to-Sequence Auto-Encoder (AE-FP): Another deep learning approach that translates molecular representations into latent space representations.
  • Drug-Target Interaction (DTI) Network: A DTI network was constructed to identify potential lead compounds.
  • Virtual Screening: The framework integrates ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity) property prediction, synthetic accessibility (SAS) assessment, and hERG channel activity prediction to screen for safe and effective compounds.
  • Molecular Optimization: The OptADMET online server was used to optimize the molecular structures of existing anesthetics to improve their properties and reduce side effects.

Datasets

  • ChEMBL Database: This database was the primary source for compound information, including SMILES strings and experimental binding affinity (BA) data.
  • GABA Receptor PPI Networks: PPI networks for 24 GABA receptor subtypes were extracted from the STRING database.
  • hERG Inhibitor Datasets: Datasets for hERG inhibitors were collected to assess cardiac safety.
  • DrugBank Database: Used for machine-learning repurposing of existing drugs.

The study curated a dataset of 136 targets from 980 proteins within the PPI networks, with a minimum of 250 data points per dataset, resulting in 183,250 inhibitor compounds for analysis.

Key Insights and Ideas

  • GABA Receptor Complexity: The study highlights the complexity of GABA receptors, with 24 subtypes requiring a comprehensive approach to target anesthetic development.
  • Proteomics-Guided ML: Integrating proteomic data (PPI networks) with machine learning offers a powerful strategy for identifying not only effective anesthetic targets but also potential side effect pathways.
  • Cross-Target Prediction for Safety: Predicting binding affinities across multiple targets is crucial for identifying compounds that are potent against the desired receptor while minimizing interactions with off-target proteins, thereby reducing side effects.
  • Drug Repurposing Potential: The framework effectively identifies existing drugs that could be repurposed as anesthetics, offering a faster and more cost-effective route to new therapies.
  • ADMET and hERG as Essential Filters: Rigorous screening of ADMET properties and hERG activity is vital for ensuring the safety and clinical viability of potential anesthetic candidates.

Key Results

  • High Predictive Accuracy: The developed machine learning models demonstrated high predictive accuracy for binding affinity, with an average R value of 0.791 and RMSD values within a reasonable range.
  • Identification of Side Effects: The cross-target prediction analysis identified numerous instances of potential mutual side effects between proteins within the same families (e.g., somatostatin receptors, muscarinic acetylcholine receptors).
  • Repurposing Candidates: The study identified two nearly optimal lead compounds (ChEMBL1372447 and ChEMBL200482) for repurposing as anesthetics, showing strong binding to GABRA5 and no predicted side effects on hERG.
  • Optimization of Existing Anesthetics: Molecular optimization of propofol and cinchocaine led to improved ADMET properties and reduced hERG side effects, demonstrating the utility of the optimization approach.
  • Repurposing of FDA-Approved Drugs: Four FDA-approved drugs (Chloramphenicol, Darolutamide, Maribavir, and Upadacitinib) were identified as potential candidates for anesthetic repurposing based on their predicted binding affinities to GABRA5 and lack of hERG side effects.

Limitations and Future Work

  • Dataset Scarcity: The authors acknowledge that the scarcity of specialized datasets tailored for anesthetic drug research and the incomplete understanding of anesthesia mechanisms pose significant challenges.
  • In Vitro/In Vivo Validation: While the study provides strong computational predictions, the identified lead compounds and repurposed drugs require experimental validation through animal studies to confirm their agonist/antagonist properties, toxicological characteristics, and blood-brain barrier permeability.
  • Generative Network for Continuous Discovery: The paper mentions the use of a generative network module to continuously generate new candidate drugs and predict their side effects, suggesting this as a future direction for accelerating drug development.
  • Personalized Anesthesia: The authors envision integrating LLMs with proteomic technologies for personalized anesthesia, intensive care, and pain management, indicating a future research trajectory.

Summary updated

BibTeX
@article{wei2025proteomic,
  author = {Guo-Wei Wei and Jian Jiang and Long Chen and Yueying Zhu and Yazhou Shi and Huahai Qiu and Bengong Zhang and Tianshou Zhou and Guo-Wei Wei},
  title = {{Proteomic Learning of Gamma-Aminobutyric Acid (GABA) Receptor-Mediated Anesthesia}},
  journal = {Journal of Chemical Information and Modeling},
  year = {2025},
  doi = {10.1021/acs.jcim.5c00114},
}
Figure