Protein language models (PLMs) are often assumed to capture evolutionary information by training on large protein sequence datasets. Yet it remains unclear whether PLMs can reason about evolution--that is, infer evolutionary relationships between sequences. We test this capability by evaluating whether standard PLM usage, frozen or fine-tuned embeddings with distance-based comparison, supports evolutionary reasoning. Existing PLMs consistently fail to recover phylogenetic structure, despite strong performance on sequence-level tasks such as masked-token and contact prediction. We present PHYLA, a hybrid state-space and transformer model that jointly processes multiple sequences and is trained using a tree-based objective across 3,000 phylogenies spanning diverse protein families.

The enormous diversity of life forms thriving in drastically different environmental milieus involves a complex interplay among constituent proteins interacting with each other. However, the organizational principles characterizing the evolution of protein interaction networks (PINs) across the tree of life are largely unknown. Here we study 4,738 PINs belonging to 16 phyla to discover phyla-specific architectural features and examine if there are some evolutionary constraints imposed on the networks' topologies. We utilized positional information of a network's nodes by normalizing the frequencies of automorphism orbits appearing in graphlets of sizes 2-5. We report that orbit usage profiles (OUPs) of networks belonging to the three domains of life are contrastingly different not only at the domain level but also at the scale of phyla. Integrating the information related to protein families, domains, subcellular location, gene ontology, and pathways, our results indicate that wiring patterns of PINs in different phyla are not randomly generated rather they are shaped by evolutionary constraints imposed on them. There exist subtle but substantial variations in the wiring patterns of PINs that enable OUPs to differentiate among different superfamilies. A deep neural network was trained on differentially expressed orbits resulting in a prediction accuracy of 85%.