This work describes and discusses an innovative approach for the prediction of physical, chemical and biological properties of compounds, ranging from small molecules to large polymers. It is based on the direct and adaptive treatment of molecular structure by means of a Recursive Neural Network (RNN) to derive Quantitative Structure-Property/Activity Relationships (QSPR/QSARs). Chemical compounds are represented through appropriate graphical tools that bypass the need for numerical descriptors. The capabilities of this methodology are investigated by applying it to different predictive problems: the melting point of ionic liquids, the glass transition temperature of polymers and the toxicity of organic molecules. The results show that the graphical molecular representation was able to effectively model each case, providing accurate predictions using practically no background knowledge. The proposed structure-based RNN approach, it is argued, can provide a simple and general prediction method with great potential in molecular design, toxicology and evaluation of other complex properties.