GeneGuard AI: Real-Time Pathogenicity Prediction of Single Nucleotide Variants using Genomic Foundation Models and Transformer Fine-Tuning

Single nucleotide variants (SNVs) are among the most clinically significant mutation types in humans, linked to inherited diseases, rare Mendelian conditions, and cancer predisposition. Functional evaluation of the millions of novel SNPs discovered annually is prohibitively expensive, while existing machine learning approaches depend on annotated data that may not generalize to novel variants. We present GeneGuard AI, an integrated genomics platform that combines the theoretical foundations of Evo2 (trained on 9.3 trillion base pairs) with a fine-tuned Nucleotide Transformer (500M parameters). From ?350,000 ClinVar variants, a rigorous filtering pipeline retains 345,909 high-confidence SNVs. Chromosomestratified splitting (chromosomes 1–18 for training; 19–22, X, Y held out) prevents genomic data leakage. After three epochs of mixed-precision training on a Tesla T4, GeneGuard AI achieves 71.5% accuracy on 55,784 unseen variants (precision 71.8%, recall 73.2%, F1 72.5%), with 690 ms end-to-end inference latency confirming suitability for interactive clinical use.