Proteins, like people, have family trees. Two proteins performing the same job in different organisms descend from a common ancestor. By comparing their sequences, you can reconstruct that history. But when proteins diverged long ago, standard methods lose the trail: so many differences pile up that the signal fades, like old photographs yellowing.
The new approach, HyperEvoGen, builds a relatedness map in curved space — one where parallel lines can diverge. This space is saddle-like: the farther from the center, the faster distances grow. In it, descendant proteins sit on branches radiating from the ancestor, and the distances between them honestly show how many evolutionary steps separate them. The neural network learns to arrange proteins so the map preserves the hierarchy of relatedness, even for deeply diverged families.
Such a map is useful not only for history. It helps create new proteins with desired properties, since you can move across the map toward the traits you want. For large protein families, where conventional computations would take weeks, the new method works in hours. This opens the way to studying the evolution of thousands of proteins at once and to accelerated development of enzymes and drugs.
Lobachevsky space, where HyperEvoGen operates, is infinite: it holds far more 'places' than a flat map. So even very distant proteins remain distinguishable, and relatedness is not lost.