A structure-aware generative AI framework for revealing functional relationships in protein families

Proteins can be studied through their sequence statistics or structural properties. These represent complementary views that are useful but lack a quantitative framework to tell, family by family, which is most informative and how to combine them. We introduce a framework that builds both views in parallel: amino acid (AA) alignments are translated into parallel alignments over a 3D interaction (3Di) structure-informed alphabet. Variational autoencoders compress each into a two-dimensional map, and direct coupli…

By the Truvace deskOctober 14, 20262 min read1 sourcePeer-reviewedTRV-2026-1381· cite & verify

In brief

Researchers built a framework that translates amino acid alignments into parallel alignments over a structure-informed 3Di alphabet. Variational autoencoders compress each alignment into a two-dimensional map, and direct coupling analysis assigns a shared coevolutionary energy to both, creating generative landscapes.

The approach gives a quantitative way to judge, family by family, whether sequence or structure is more informative and how to combine them. The abstract describes the method as an aid to mechanistic insight and protein design, but it reports no experimental validation of newly designed proteins.

Main points

The framework defines information-theoretic distance metrics that quantify how sequence changes drive structural and functional variation.
In malate dehydrogenases, the 3Di landscape identifies a structurally conserved scaffold used to encode thermal adaptation.
In Flaviviridae E1 and E2 glycoproteins, structure shows evolutionary relationships not visible at the sequence level.
Because many sequences encode the same fold, the method separates family sequence variability from structural and functional variation.

The gain

Pairing amino acid and 3Di structural alignments in generative landscapes recovered known subfamilies and revealed evolutionary relationships in viral glycoproteins that sequence alone did not show.

The rundown

The study was tested on five protein families: malate dehydrogenases, globins, TRPM channels, and the Flaviviridae E1 and E2 glycoproteins. Each family illustrates a different way structure and sequence information can diverge or complement each other.

In malate dehydrogenases, the amino acid landscape revealed sequence variability linked to thermal adaptation, while the 3Di landscape picked out the conserved structural scaffold. In globins and TRPM, the structural landscape recovered subfamilies already known to be functionally distinct.

The authors say the landscapes can be sampled near functional regions, which they suggest may guide the design of new proteins and give insight into evolutionary forces.