We show the improvement provided by data-dependent embedding methods in preserving the evolutionary distance between sequences. In particular, the hyperbolic space is able to capture the hierarchical relationship between the sequences, significantly reducing the distortion.

Finally, we propose a series of ways of adapting the framework to perform the combinatorial intractable tasks of hierarchical clustering and multiple sequence alignment, all of which show significant runtime improvements over baseline methods.

You can find the paper at and the code at . Work done under the amazing supervision of @RexYing0923 @Mpmisko @PetarV_93 @jure and @pl219_Cambridge

@GabriCorso @PetarV_93 @jure @Mpmisko @RexYing0923 @pl219_Cambridge Indeed great work. I enjoy to read it a lot. Just my concern is the claim that edit distance is the best way to measure evolutionary distance between biological sequence. I guess even small edit distance might change function a lot, sometimes vice versa.

@GabriCorso @PetarV_93 @jure @Mpmisko @RexYing0923 @pl219_Cambridge @ixxmael_freitas mais uma aplicação em ciências biológicas do meu tema de estudo ó que legal

@GabriCorso @PetarV_93 @jure @Mpmisko @RexYing0923 @pl219_Cambridge Really enjoyed this paper - great observations regarding performance of hyperbolic vs. standard distances. However, I gotta say that labeling these performance tables as "Figures" is rather... generous :)