Artificial neural networks are presented from the perspective of their potential use as modeling tools in quantitative structure-activity relationships (QSAR) research. First, general merits and drawbacks of the neural network modeling approach are discussed, and the relationship between neural networks, statistics and expert systems is clarified. A separate section is devoted exclusively to the subject of validating neural networks models. Next, the review focuses on presenting the most commonly used artificial neural networks in QSAR: backpropagation neural networks, probabilistic neural networks, Bayesian regularized neural networks, and Kohonen SOM. For each of them, both merits and shortcomings are revealed, and references are made to publications presenting their QSAR applications. Another section is devoted to genetic algorithms, their merits and shortcomings, and their potential use for model variables dimensionality reduction in QSAR studies. The last section is devoted to software resources.