VariantLLM: An Open-Source Foundation Model Microservice for Zero-Shot Clinical Variant Effect Prediction

Accurate functional interpretation of missense single-nucleotide variants (SNVs) remains a foundational challenge in precision oncology and clinical genomics. Existing computational predictors predominantly rely on supervised architectures trained on historical clinical databases, rendering them susceptible to circularity bias and data leakage. Here, we introduce VariantLLM, a lightweight, production-grade genomic foundation artificial intelligence system engineered for zero-shot clinical variant effect prediction. Powered by Meta AI's evolutionary transformer foundation model (ESM-2), VariantLLM quantifies the disruption of intrinsic biophysical constraints across phylogenetic space via masked-marginal Log-Likelihood Ratio (ΔLLR) scoring without requiring task-specific supervised training. We provide a modular PyTorch package, high-throughput FastAPI REST microservices, automated CI/CD pipelines, and an interactive cloud web deployment. On a curated benchmark of validated human clinical mutations from NCBI ClinVar spanning oncology, neurodegeneration, and channelopathies, VariantLLM achieves exceptional discriminatory power (ROC-AUC = 1.0000, PR-AUC = 1.0000, Overall Accuracy = 94.44%, F1-Score = 0.9524, MCC = 0.8919). • Code Repository: https://github.com/hardiksood21/variantllm• Live Web Demo: https://huggingface.co/spaces/hardiksood21/variantllm