Protein Language Model Embeddings Recover Evolutionary and Functional Organization Across the Ras Superfamily.

Abstract

The Ras protein superfamily comprises small GTPases that share a conserved G-domain but differ in flanking regions, regulation, and cellular roles. Because existing classifications rely mainly on G-domain phylogeny, this superfamily provides a useful test case for assessing whether protein language model embeddings recover biologically meaningful sequence organization consistent with established evolutionary classifications. Here, we analyzed a curated Ras superfamily dataset using three classification schemes: the classical five-family view, a G-domain phylogeny-based classification, and UniProtKB family annotation classification. We compared embeddings from multiple protein language models using supervised classification, unsupervised clustering, and residue-level ablation. Sequence-derived embeddings recovered known Ras superfamily organization across analyses. In supervised analyses, simple linear classifiers achieved high performance, indicating that Ras family and subfamily information is linearly accessible from sequence-derived embeddings. In unsupervised analyses, ESM-C-300 layer 12 gave the strongest full-protein recovery of the G-domain phylogeny-based classification, whereas ProstT5 performed best for G-domain embeddings. Residue ablation identified recurrent candidate subfamily-informative regions both within and outside the G-domain, including signals mapping to structurally coherent regions associated with subfamily-specific regulatory or interaction-related functions. Together, these results indicate that protein language model embeddings provide an effective alignment-free representation of Ras family and subfamily organization, and can highlight candidate sequence regions associated with functional specialization from sequence alone.