A team of researchers has developed an evolutionary algorithm that uses machine learning to identify relevant structures in molecules and predict their properties. This method can be applied to any molecular data set and does not require expert knowledge.