DiscoTope-3.0 is an advanced tool for predicting B-cell epitopes, leveraging inverse folding latent representations and a positive-unlabeled learning strategy. This tool significantly improves upon existing methods by accurately predicting linear and conformational epitopes across multiple datasets, both for solved and predicted structures. The performance of DiscoTope-3.0 is maintained across solved, relaxed, and predicted structures, reducing the need for experimental structures and expanding the applicability of B-cell epitope prediction by three orders of magnitude. The tool is accessible via web servers and as a downloadable package, interfacing with RCSB and AlphaFoldDB to process over 100 structures per submission. It demonstrates superior performance compared to other state-of-the-art methods, particularly in predicting non-linear epitopes. The positive-unlabeled learning approach and robustness to structural quality make DiscoTope-3.0 a valuable resource for vaccine and immunotherapy development.DiscoTope-3.0 is an advanced tool for predicting B-cell epitopes, leveraging inverse folding latent representations and a positive-unlabeled learning strategy. This tool significantly improves upon existing methods by accurately predicting linear and conformational epitopes across multiple datasets, both for solved and predicted structures. The performance of DiscoTope-3.0 is maintained across solved, relaxed, and predicted structures, reducing the need for experimental structures and expanding the applicability of B-cell epitope prediction by three orders of magnitude. The tool is accessible via web servers and as a downloadable package, interfacing with RCSB and AlphaFoldDB to process over 100 structures per submission. It demonstrates superior performance compared to other state-of-the-art methods, particularly in predicting non-linear epitopes. The positive-unlabeled learning approach and robustness to structural quality make DiscoTope-3.0 a valuable resource for vaccine and immunotherapy development.