A Ligand-Based Approach to the Discovery of Lead-Like Potassium Channel KV1.3 Inhibitors

Posted by
Florbela Pereira
on 2019-09-12

A Ligand-Based Approach to the Discovery of Lead-Like Potassium Channel KV1.3 Inhibitors

Voltage-gated ion channels are key molecular targets for autoimmune diseases such as multiple sclerosis, rheumatoid arthritis and psoriasis. In silico models, using 340 molecules whose IC50 towards Kv1.3 was determined by reported assays, were developed through exploration of four machine learning (ML) techniques. ML techniques explored included Random Forest, Support Vector Machine, Multilayer Perceptron, and K-Nearest Neighbors. Two QSAR classification approaches were developed. In the first approach, the compounds were classified into either moderate-active-to-very-active or inactive-to-moderate-active categories. Only the compounds predicted to be moderate-active-to-very-active in the first classification model were submitted to the second model, a classification model that predicted two more categories, very-active and not- very-active. The performances of the models were successfully evaluated by internal validation and external test set validation, with an overall predictability (Q) of 0.83 and 0.69 for the test set in the first and second approaches of the best models, respectively. The best models for the two approaches were employed for the virtual screening of KV1.3 inhibitors from ZINC natural products and approved drugs databases. A list of the most promising lead-like KV1.3 inhibitors was proposed, from which an approved drug and a natural product were experimentally evaluated with whole-cell voltage-clamp assays at 1000 nM.

Visit the publication

 

Voltage-gated ion channels are key molecular targets for autoimmune diseases such as multiple sclerosis, rheumatoid arthritis and psoriasis. In silico models, using 340 molecules whose IC50 towards Kv1.3 was determined by reported assays, were developed through exploration of four machine learning (ML) techniques. ML techniques explored included Random Forest, Support Vector Machine, Multilayer Perceptron, and K-Nearest Neighbors. Two QSAR classification approaches were developed. In the first approach, the compounds were classified into either moderate-active-to-very-active or inactive-to-moderate-active categories. Only the compounds predicted to be moderate-active-to-very-active in the first classification model were submitted to the second model, a classification model that predicted two more categories, very-active and not- very-active. The performances of the models were successfully evaluated by internal validation and external test set validation, with an overall predictability (Q) of 0.83 and 0.69 for the test set in the first and second approaches of the best models, respectively. The best models for the two approaches were employed for the virtual screening of KV1.3 inhibitors from ZINC natural products and approved drugs databases. A list of the most promising lead-like KV1.3 inhibitors was proposed, from which an approved drug and a natural product were experimentally evaluated with whole-cell voltage-clamp assays at 1000 nM.

Visit the publication