The TF-IDF is a measure of the discriminating ability of a term in a document. The TF-IDF of a specific term increases if the term is more frequent in a specific document but decreases if it is frequent in the whole corpus.
The problem with TF-IDF is that while it is good in distinguishing the document from the corpus, it isn't good in distinguishing one label from another!
What you could try to compute is a variation of the TF-IDF, where in the nominator you would count the term frequency, not a specific document, but in the set of documents under the same label. The denominator would stay the same. This way, instead of getting the top terms for each document you could get the top terms for each label. This metric could give you a rough estimation of what you want.
Note that while the top term for each label would be the most important in classifying examples to the specific label, it isn't necessarily the best for classification in general. The best term for classification, theoretically, would be one that could divide your data in two equal parts.