DiscoTope-3.0 is a novel B-cell epitope prediction tool that improves upon existing methods by using inverse folding latent representations and a positive-unlabelled learning strategy. It can be applied to both experimentally solved and predicted protein structures, significantly enhancing the accuracy of epitope prediction across multiple datasets. The tool maintains high performance on both solved and predicted structures, reducing the reliance on experimental data and extending the applicability of accurate B-cell epitope prediction by three orders of magnitude. DiscoTope-3.0 is available as a web server and downloadable package, processing over 100 structures per submission and interfacing with RCSB and AlphaFoldDB for large-scale prediction across over 200 million proteins. The tool demonstrates improved performance compared to existing methods, particularly in predicting linear and conformational epitopes. It is robust to structural variations, including relaxation and predicted structures, and shows strong performance on exposed and non-linear epitopes. DiscoTope-3.0 also calibrates scores for antigen length and surface area, improving the reliability of epitope predictions. The tool is accessible via a web server and is designed for academic use, allowing batch processing of up to 50 structures at a time. DiscoTope-3.0 represents a significant advancement in structure-based B-cell epitope prediction, offering state-of-the-art performance across solved, relaxed, and predicted structures.DiscoTope-3.0 is a novel B-cell epitope prediction tool that improves upon existing methods by using inverse folding latent representations and a positive-unlabelled learning strategy. It can be applied to both experimentally solved and predicted protein structures, significantly enhancing the accuracy of epitope prediction across multiple datasets. The tool maintains high performance on both solved and predicted structures, reducing the reliance on experimental data and extending the applicability of accurate B-cell epitope prediction by three orders of magnitude. DiscoTope-3.0 is available as a web server and downloadable package, processing over 100 structures per submission and interfacing with RCSB and AlphaFoldDB for large-scale prediction across over 200 million proteins. The tool demonstrates improved performance compared to existing methods, particularly in predicting linear and conformational epitopes. It is robust to structural variations, including relaxation and predicted structures, and shows strong performance on exposed and non-linear epitopes. DiscoTope-3.0 also calibrates scores for antigen length and surface area, improving the reliability of epitope predictions. The tool is accessible via a web server and is designed for academic use, allowing batch processing of up to 50 structures at a time. DiscoTope-3.0 represents a significant advancement in structure-based B-cell epitope prediction, offering state-of-the-art performance across solved, relaxed, and predicted structures.