Alignment of each motif concept peaks at the position of that motif in the protein. We also detect a few motifs absent from individual databases, though these are typically annotated in other databases. 4/4

Building on the idea of Concept Activation Vectors from arxiv.org/pdf/1711.11279 3/4

We take embeddings of protein fragments w/ and w/o a motif, train a simple linear classifier, and use the normal vector to the decision boundary as the “motif direction.” So for motif detection, all you need is a dictionary of learned motif concept vectors, and a PLM to embed the protein with. 2/4