Large-scale sequencing has uncovered many protein-coding variants, yet clinical interpretation remains limited by variants of uncertain significance. We evaluate a two-stage, frozen-encoder approach that pairs ESM-1b embeddings with a lightweight classifier and compare it with a conventional Transformer baseline on a ClinVar-UniProtKB dataset of 98,626 variants spanning more than 5500 genes. The ESM-1b route improves overall accuracy and AUC relative to the baseline, with the largest gains for benign variants; for pathogenic variants, precision increases with a small recall trade-off, yielding a net F1 improvement. Keeping the encoder fixed makes training simple, reproducible, and compute-efficient while preserving long-range, evolution-informed sequence features. Importantly, the improved classification of pathogenic and benign variants suggests that this approach could assist clinical geneticists in interpreting VUS, thereby supporting early diagnosis, risk assessment, and precision medicine applications. These results support frozen protein-language-model embeddings as a practical basis for pathogenicity prediction and motivate follow-up work on operating-point selection and the addition of complementary evolutionary or structural context to further improve sensitivity.