The activities of a series of azidothymidine derivatives, compounds with anti-HIV potency, were computationally modelled using multivariate image analysis applied to quantitative structure–activity relationships (MIA-QSAR). Two regression methods were tested in order to find the best correlation between actual and predicted activities: bilinear (traditional) partial least squares (PLS), applied to the unfolded dataset, and multilinear PLS (N-PLS), applied to the three-way array. The predictive abilities of the PLS- and N-PLS-based models were found to be nearly equivalent, and both the methods derived QSAR models that are statistically superior to conventional QSAR, in which physicochemical descriptors and multiple linear regression were applied.