Rewriting Protein Alphabets with Language Models: An Agent-Based Approach to Homology

ABSTRACT

Detecting remote protein homology across billions of sequences using high-dimensional embeddings is computationally prohibitive. We introduce an agentic approach that utilizes contrastive learning to rewrite protein language model embeddings into a novel, optimized 20-letter alphabet (TEA). A dedicated 'Search Agent' then leverages this compressed alphabet to perform highly efficient, large-scale homology searches using standard tools like MMseqs2. Our autonomous system retains the exceptional remote homology detection of large pLMs while accelerating search speeds by three orders of magnitude, making whole-biosphere comparisons feasible on standard lab hardware.