Abstract

Background: Single nucleotide polymorphisms (SNPs) are tiny DNA changes that can play a big role in health, especially for rare diseases where catching issues early is tough. Sorting through millions of variants to find the harmful ones is no easy task and calls for smart computational tools. Objective: We set out to build and test machine learning models to predict whether SNPs linked to rare diseases are harmful, using genomic data available to everyone. Methods: We looked at 25,000 SNPs from databases like ClinVar and dbSNP, labeling them based on their clinical importance. Our dataset had 60% harmful SNPs and 40% benign ones. We trained several models—Random Forest, Gradient Boosting, Support Vector Machines, and Deep Neural Networks—using factors like allele frequency, conservation scores, and protein effects. To gauge how well they worked, we checked accuracy, precision, recall, F1-score, and the area under the ROC curve (AUC-ROC). We also used cross-validation and analyzed which features mattered most to make sure our results held up. Results: The Random Forest model came out on top, hitting 87% accuracy, 85% precision, 88% recall, and an AUC-ROC of 0.92. It was great at spotting patterns in harmful SNPs, especially those messing with protein function. The biggest drivers of predictions were conservation scores, protein domain changes, and allele frequency. Compared to older methods, our model cut false positives by 15%. Conclusion: Machine learning, especially Random Forest, is a strong tool for pinpointing harmful SNPs in rare diseases. Bringing these models into clinical practice could speed up diagnoses, improve genetic screening, and help tailor treatments to patients