At first sight, the QSAR issue might appear to be a mere pattern recognition problem; however, a purely "surface" approach to QSAR as a pattern recognition problem, not involving the profound plausibility of the solutions, has often been demonstrated to be devoid of scientific value and of predictive strength. The requirement for such a lateral validation should imply the recognition of the basic differences between the two terms of the QSAR issue: biology and chemistry. In particular, the difficulty to derive strong quantitative theories for the biological aspect of QSAR procedures should be taken into serious consideration. Within this conceptual framework, this paper examines the different families of mathematical models (classical regression, multivariate methods, neural networks) used in the QSAR research.