Finding methods to increase the complexity of the Boolean discriminant functions and to stay within the limits of tractability set by combinatorics is an important task in the field of symbolic machine learning. The original formalism based on meta-features is introduced. Meta-features are predicates that describe relations between the features of the investigated objects and the subclasses (clusters inside classes) of the training set. The formalism facilitates finding Boolean discriminant functions of three variables. These are more complecated than simple conjunctions if the partition of the original training set into subclasses is given. The structure of meta-feature predicates is close to the structure of statements used by domain experts to describe their knowledge. Consequently, the formalism can be applied in hybrid learning systems, which incorporate information obtained from domain experts.