Capsid-specialized protein language models reveal higher-order viral architecture from sequence

Current methods for identifying viral capsid proteins rely heavily on similarity to known sequences, but these approaches struggle with detecting capsids in highly divergent environmental sequences. To address this limitation, researchers have developed ESMCapsid, a specialized protein language model tailored for remote capsid detection and understanding their complex architecture.

Testing ESMCapsid across 343 million metagenomic protein clusters uncovered a substantial number of "homology-dark" capsids – capsid proteins that don't have clear matches to existing reference databases. About 62% of these candidates lacked any related sequences in current databases, indicating a vast reservoir of viral diversity that has been overlooked by traditional methods.

By applying a technique called sparse autoencoder decomposition, the researchers identified recurring patterns or "motifs" that linked these homology-dark proteins to established structural lineages. These motifs provided a way to infer the higher-order architectural information contained within the capsid sequences, even when their sequence similarity to known proteins was minimal.

When the spatial distribution of these conserved motif cores was mapped onto resolved viral shells, it became clear that ESMCapsid is capturing more than just sequence similarity. The motifs appeared to be associated with specific geometric positions within the viral capsid, suggesting that the model is uncovering essential architectural constraints that go beyond simple sequence patterns.

In summary, ESMCapsid offers a powerful approach to organizing the vast array of homology-dark viral diversity by revealing the underlying architectural principles that govern these proteins. This method has the potential to greatly expand our understanding and discovery of viral capsids, even in cases where traditional sequence-based techniques fall short.