The prediction of properties of molecules just on the basis of their chemical structures is desirable to selectively make molecules that have the wanted properties, like biological activity, viscosity, or toxicity. Here, we present an example of a new way to predict a property from the chemical structure of a chemically heterogeneous class of compounds. The clearing temperatures of nematic liquid-crystalline phases of 17383 compounds were used to train neural networks to derive this material property directly from their chemical structure. The trained neural networks were subsequently tested with 4345 structural patterns of molecules unknown to the networks to assess their predictive value. The clearing temperatures were predicted by the best network with a standard deviation of 13 degrees.