During the last years, new connectionist approaches for adaptive structure processing called folding architecture networks or recursive neural networks have been developed. The main objective of this paper is to explore the applicability of these networks in the field of chemistry. Experimental results on a benchmark for the prediction of quantitative structure activity relationships are presented and compared to results achieved with other machine learning techniques. Though the results achieved with the new networks are slightly better, it has to be stated that the statistical evidence is rather weak. For future comparisons, bigger benchmarks are needed.