An in silico protocol for predicting genetic biomarkers in rare diseases: a case study in sporadic amyotrophic lateral sclerosis

Studying the genetics of rare diseases is challenging because small sample sizes limit the statistical power of standard methods like Genome-wide association studies (GWAS). We created a new machine-learning approach to find candidate Single Nucleotide Polymorphisms (SNPs) when data is scarce. Our method trains a Random Forest model to spot similarities between SNPs. We used 189 known Sporadic Amyotrophic Lateral Sclerosis (sALS)-linked SNPs as positive examples and 938,544 unrelated SNPs as negatives. The model learns from genomic location, significance levels, nearby genes, and other features. When we tested it on sALS, it performed exceptionally well, with 93.8% accuracy and near-perfect AUC scores. The method uncovered 1,890 new SNP candidates for sALS. Among these, 209 reached genome-wide significance, and 50 appeared repeatedly in our analyses, making them strong candidates. Key genes like <i>SARM1</i>, <i>OPHN1</i>, and <i>BPTF</i> emerged from the results, all connected to neural health and survival pathways. Our examination revealed a notable excess of SNPs on chromosome 18 compared to expectations. This non-random distribution underscores the region's particular interest. Here, our approach demonstrates its ability to extract meaningful signals from a restricted sample. The results generated by this approach enable early diagnosis of the disease under study, explanation of its mechanism, and identification of therapeutic targets.