Cognitive algorithms need to adapt complex abstract structures encoded from previous behavioural patterns and generalize the computations to flexibly support goal-oriented behaviour in novel situations. However, the neural algorithms that implement the mapping of these abstract structures from behavioural patterns remain unknown. Now, El-Gaby et al. outline an algorithm implemented by neurons in medial frontal cortex (mFC) in mice that composes such abstract structures, and demonstrate how that enables the adaptation of computations during goal-oriented behaviour to novel situations.

The authors trained mice to solve a series of goal-oriented ‘ABCD’ behavioural tasks on a 3 × 3 grid maze. Each task had two essential ingredients, an unchanging abstract structure (the sequential ABCD series of four goals to find) and a unique spatial structure (the rewarded locations corresponding to each goal’s grid placement). Mice performed trials within each ABCD task in a loop-like fashion, in which finding goal D immediately started the next trial (involving finding goal A). Earlier tasks trained mice to extract regularities from the grid-like environment and merge them with information from their behavioural patterns across repeated ABCD sequences to compose the unchanging abstract task structure. The first trial in later tasks assessed how well mice could generalize that composition to solve novel spatial structures, by examining whether mice that reached goal D could infer the correct location of goal A to immediately begin a second trial in a new task (in one shot).