Generally, PLS-DA is carried out when discrimination was not apparent neither by using visual inspection nor principal component analysis (PCA). Although as you are saying PCA provides a good separation among different samples, they are not the most appropriate tool for developing models that allow previous designations of specific sets of samples and determine precise thresholds for classification. For this aim, supervised classification multivariate methods are required, such as PLS-DA. PLS-DA provides high rates of sensitivity and specificity. A PLS-DA model can also substantially reduce the number of discriminatory variables by creating significant VIP (variable importance in projection) scores.

So I would suggest incorporating both PCA and PLS-DA.