Identifying Molecular Properties of Ataxin-2 Inhibitors for Spinocerebellar Ataxia Type 2 Utilizing High-Throughput Screening and Machine Learning

Spinocerebellar ataxia type 2 (SCA2) is an autosomal dominant neurodegenerative disorder marked by cerebellar dysfunction, ataxic gait, and progressive motor impairments. SCA2 is caused by the pathologic expansion of CAG repeats in the ataxin-2 (<i>ATXN2</i>) gene, leading to a toxic gain-of-function mutation of the ataxin-2 protein. Currently, SCA2 therapeutic efforts are expanding beyond symptomatic relief to include disease-modifying approaches such as antisense oligonucleotides (ASOs), high-throughput screening (HTS) for small molecule inhibitors, and gene therapy aimed at reducing <i>ATXN2</i> expression. In the present study, data mining and machine learning techniques were employed to analyze HTS data and identify robust molecular properties of potential inhibitors of <i>ATXN2</i>. Three HTS datasets were selected for analysis: <i>ATXN2</i> gene expression, CMV promoter expression, and biochemical control (luciferase) gene expression. Compounds displaying significant <i>ATXN2</i> inhibition with minimal impact on control assays were deciphered based on effectiveness (E) values (<i>n</i> = 1321). Molecular descriptors associated with these compounds were calculated using MarvinSketch (<i>n</i> = 82). The molecular descriptor data (MD model) was analyzed separately from the experimentally determined screening data (S model) as well as together (MD-S model). Compounds were clustered based on structural similarity independently for the three models using the SimpleKMeans algorithm into the optimal number of clusters (<i>n</i> = 26). For each model, the maximum response assay values were analyzed, and E values and total rank values were applied. The S clusters were further subclustered, and the molecular properties of compounds in the top candidate subcluster were compared to those from the bottom candidate subcluster. Six compounds with high <i>ATXN2</i> inhibiting potential and 16 molecular descriptors were identified as significantly unique to those compounds (<i>p</i> < 0.05). These results are consistent with a quantitative HTS study that identified and validated similar small-molecule compounds, like cardiac glycosides, that reduce endogenous ATXN2 in a dose-dependent manner. Overall, these findings demonstrate that the integration of HTS analysis with data mining and machine learning is a promising approach for discovering chemical properties of candidate drugs for SCA2.