Prediction of Activity and Selectivity Profiles of Sigma Receptor Ligands Using Machine Learning Approaches

Lisa Lombardo, Verena Battisti, Thierry Langer, Rosaria Gitto, Laura De LucaJournal of Chemical Information and Modeling 2025 · DOI · OpenReview

TL;DR

This paper presents a machine learning framework to predict the activity and selectivity of compounds targeting sigma receptors (SRs), crucial proteins involved in neurological disorders and cancer. The authors developed and compared three distinct workflows, ultimately identifying a one-step multiclassification approach using Extra Trees and Mordred descriptors as the most effective for identifying selective SR ligands.

What Problem Does This Paper Address?

Sigma receptors (SRs), specifically subtypes S1R and S2R, are important therapeutic targets due to their roles in various diseases. While they share some similarities in their binding sites, they have distinct physiological functions, making the development of selective ligands crucial for targeted therapies. However, distinguishing between agonists and antagonists for these subtypes is challenging due to limited experimental data and underdeveloped functional assays, particularly for S2R. This research aims to address this gap by developing predictive models to accurately identify compounds with specific activity and selectivity profiles for S1R and S2R.

Methods

The study explored three machine learning (ML) workflows:

  • 2-Step Classification Workflow: This approach uses separate binary classification models to first predict activity (active/inactive) and then selectivity (selective/non-selective) for each receptor.
  • 2-Step Regression Workflow: This workflow uses regression models to predict the pActivity values for each receptor, and selectivity is derived from the difference in predicted pActivities.
  • 1-Step Multiclassification Workflow: This method uses a single multiclass classification model to simultaneously predict activity and selectivity (e.g., S1R-selective, non-selective, inactive).

The researchers employed a variety of ML algorithms, including Random Forest (RF), Extra Trees (ET), XGBoost, Support Vector Machines (SVM), and k-Nearest Neighbors (kNN). They also investigated five types of molecular representations: RDKit descriptors, Mordred descriptors, ECFP4, ECFP6, and MACCS keys. Rigorous nested and classical 5-fold cross-validation protocols were used for model selection and validation, with an external validation set used to compare the workflows.

Datasets

The study utilized curated datasets of bioactivity records for SRs. These data were collected from public databases such as ChEMBL, BindingDB, PubChem BioAssays, and S2RSLDB, supplemented with an in-house dataset. The data curation process involved standardizing molecular representations, removing salts, eliminating duplicates, and converting bioactivity measures into standardized pActivity values. The final curated dataset comprised 4994 compounds with 6110 high-quality activity records. For model development, six distinct datasets were created, categorized by the type of prediction (activity, selectivity, or multiclass) and receptor target. An external validation set, representing 10% of the data and specifically including "double-point" compounds (those with data for both receptors), was set aside to ensure an unbiased evaluation of the workflows.

Key Insights and Ideas

  • The 1-step multiclassification workflow is superior: The study found that a single model predicting all classes (selective, non-selective, inactive) simultaneously outperformed the two-step approaches for both activity and selectivity prediction.
  • Mordred descriptors are highly informative: Across various models and tasks, Mordred descriptors consistently emerged as the most effective molecular representation for capturing relevant chemical information.
  • Fingerprint choice matters for selectivity: While Mordred descriptors were generally strong, specific fingerprints like ECFP4 for S1R and ECFP6 for S2R showed better performance for selectivity prediction tasks.
  • Data imbalance is a significant challenge: The scarcity of data for S2R and particularly for selective compounds (S1R-selective and S2R-selective) posed a challenge, impacting model performance and requiring undersampling techniques.
  • Feature importance reveals subtype-specific drivers: Analysis of feature importance highlighted distinct structural features driving S1R vs. S2R selectivity. S1R selectivity was linked to basicity and electrostatic interactions, while S2R selectivity was more influenced by hydrophobicity and halogenated aromatic rings.

Key Results

The 1-step multiclassification workflow, specifically using Extra Trees (ET) combined with Mordred descriptors, demonstrated the best predictive performance in external validation.

  • For S1R activity and selectivity prediction, the ET multiclass model achieved an overall accuracy of 0.936 (CI = 0.88–0.97) and an MCC of 0.733 (CI = 0.56-0.88). It correctly predicted 94 out of 95 nonselective compounds and 4 out of 8 selective compounds.
  • For S2R activity and selectivity prediction, the ET multiclass model achieved an even higher ROC-AUC of 0.964 (CI = 0.93-0.99) and an MCC of 0.89 (CI = 0.77–0.98). It accurately classified 87 out of 90 nonselective compounds and 11 out of 12 S2R-selective compounds.

The 2-step classification workflow showed moderate success, with SVM models performing well for selectivity prediction. The 2-step regression workflow, however, struggled to accurately predict selectivity, likely due to data scarcity for S2R.

Limitations and Future Work

The authors acknowledge that the small size of the external validation set, particularly for selective compounds, limits the generalizability of their conclusions and may impact model performance. They suggest that larger validation sets are needed to confirm the observed trends and model performance more reliably. Future work could focus on expanding the datasets, exploring more advanced ML techniques, and integrating experimental validation to further refine these predictive models for drug discovery.

Summary updated

Abstract

Click to copy section linkSection link copied!
BibTeX
@article{lombardo2025prediction,
  author = {Lisa Lombardo and Verena Battisti and Thierry Langer and Rosaria Gitto and Laura De Luca},
  title = {{Prediction of Activity and Selectivity Profiles of Sigma Receptor Ligands Using Machine Learning Approaches}},
  journal = {Journal of Chemical Information and Modeling},
  year = {2025},
  doi = {10.1021/acs.jcim.5c01091},
}
Figure