Proteins are the molecular machines that turn genetic information into cellular function. But we discover new sequences far faster than we can describe what they do. Only about 0.25% of known proteins carry manually curated annotations. The rest rely on computational inference. The most trusted method is to transfer function by sequence homology. This works well when a protein closely resembles a characterized one. It breaks down where discovery matters most: proteins whose function has drifted, proteins with no detectable homology, and the vast "functionally dark" regions of sequence space. By design, reference-based methods cannot discover any function unlike the reference collection.

Language-model-based methods offer a reference-free path forward. We developed ProtTale to read a protein sequence and write a Swiss-Prot-style description of its molecular function. The output is natural-language text, like that of a human curator, rather than a fixed set of ontology terms. ProtTale couples an ESM-C protein encoder to a Galactica language decoder. It is trained with Machine-Inferred Protein Similarity, which turns the functional geometry learned by protein language models into an explicit training signal. Unlike other protein-to-text models, ProtTale also emits a reliability score for every prediction, analogous to AlphaFold's pLDDT. Users can then keep only the annotations the model deems trustworthy.

Citing ProtTale. If you use ProtTale in published research, please cite:

Citation coming soon...