Over the past thirty years, alignment-based methods have played a crucial role in elucidating protein evolutionary relationships. While alignment-based analysis has been successful for studying the conservation of folded proteins, it has fallen short in assessing the conservation of intrinsically disordered regions (IDRs). IDRs lack a fixed spatial structure, which reduces the need for absolute sequence conservation. This allows them to have different lengths and compositions while still conserving local or global sequence features. While traditional alignment tools struggle to capture these conserved features, recent advances in deep learning and physics-inspired bioinformatics allow us to encode sequences into information-rich vectorial representations, which in principle encode the conserved information. Unfortunately, these representations cannot be aligned using the traditional alignment tools. Here, we repurposed an approach from the signal processing literature called dynamic time warping, enabling direct alignment in embedded space. Our approach offers an alternative method to identify similar subregions. We show that even within disordered regions, local subregions can be highly conserved despite substantial sequence variation. In summary, our work offers a robust approach to assessing sequence similarity in arbitrary feature embeddings.
Razo et al. (Sun,) studied this question.