Use the cluster method of the gmdistribution object to obtain cluster assignments and their posterior probabilities. Then pass these cluster indices to perfcurve as class labels and pass these posterior probabilities as classification scores. For example:
load fisheriris;
cv = cvpartition(species,'holdout',.5);
g = gmdistribution.fit(meas(cv.training,:),3);
[y,~,post] = cluster(g,meas(cv.test,:));
[fpr,tpr] = perfcurve(y,post(:,1),1,'negclass',[2 3]);
plot(fpr,tpr)
gives you a ROC curve of the 1st class vs 2nd and 3rd. (In this case there is perfect separation in the test data and you won't see a ROC curve, but that's the idea.)
