Evaluating Protein Language Model Embeddings for Viral Clade Assignment

Protein language models (PLMs) provide powerful sequence representations, yet their effectiveness for unsupervised viral clade assignment remains uncertain. In this study, we evaluated embeddings from ProtT5, ProtBert, CARP, and several ESM-2 variants on influenza A/H3N2 hemagglutinin sequences. Using dimensionality reduction (t-SNE, UMAP, PCA, MDS) and clustering with HDBSCAN, we compared PLM embeddings against baseline Hamming distance approaches. Our results show that t-SNE combined with PLM embeddings can recover clade structure, with ProtBert yielding the most stable performance and larger ESM-2 models occasionally achieving lower normalized variation of information scores but with greater variability. These findings suggest that while PLM embeddings capture clade-relevant signals, they also suffer from instability and the loss of site- or nucleotide-specific detail. Future improvements in pooling strategies may enhance their utility for viral surveillance.