Functional annotation of protein conformations by Self-Organizing Maps

We previously developed a strategy to compare conformational ensembles of proteins using Self-Organizing Maps (SOMs) combined with hierarchical clustering. This strategy includes three steps: a) Cα’s Cartesian coordinates of the conformations are encoded as input vectors for the SOM; b) the SOM is trained and its neurons are clustered; c) the conformations belonging to each cluster are extracted and compared. We present two applications of this protocol in order to discuss both the interpretability and the topological properties of the clustered SOM. First, we investigated the role of flexibility in protein binding with a focus on the differences between bound and unbound forms of transient complexes including members of the RAS “superfamily”. We show the possibility of using conformations sampled by two different methods (Molecular Dynamics and tCONCOORD) as input for the SOM, highlighting how different samplings are reflected by the SOM topology. Second, we identified the specific mutations that more efficiently convert the flexibility of a psychrophilic enzyme (AHA) to that of its mesophilic counterpart (PPA). This was achieved by annotation of the PPA conformations with the map trained on some AHA variants. The projection of local features, not used as input information, to the clustered SOM allowed a clear annotation of the functional differences among the structural clusters. The applications here presented support the use of this approach as a general and exploitable protocol for multiple comparison of protein conformational ensembles and highlight its potentiality for protein engineering.