The idea of combining models instead of simply selecting the ``best'' one, in order to improve performance, has a long theoretical background in statistics. However, theoretical results are ordinarily based on strong hypotheses, seldom satisfied in practice. When dealing with real-world problems, overfitting is often the main limitation, which cannot be overcome but with a strict complexity control of the combiner selected. SVMs should thus be well suited for these difficult situations. Investigating this idea, we introduce a new family of multi-class SVMs, and assess them as ensemble methods for protein secondary structure prediction. Experimental evidence highlights the gain in prediction accuracy resulting from combining some of the current best prediction methods with our SVMs rather than with the combiners traditionally used in the field.