ZPS (Zero-shot Protein Segmentation)

Proteins are modular: one chain typically contains independently folding domains, smaller sub-domain units, and intrinsically disordered regions (IDRs) that function without a stable structure. Finding where one unit ends and the next begins has long been the job of profile-matching tools like Pfam and PROSITE, and of composition-based methods such as fLPS2 and Chi-Score Analysis for low-complexity stretches. Both inherit their coverage from what has been catalogued: a region no one has annotated is a region no profile can match.

ZPS (Zero-shot Protein Segmentation), from Alan Moses's group at the University of Toronto with Julie Forman-Kay's lab at the Hospital for Sick Children, takes a different route. It applies a change point analysis directly to the per-residue embedding matrix produced by the frozen ProtT5 encoder from ProtTrans, treating the embedding as a signal along the sequence and marking where that signal shifts. Nothing is trained or fine-tuned — the segmentation is a latent property of the language model's representation, read out with a classical method from time-series analysis.

Run across the reviewed human proteome from UniProtKB/Swiss-Prot, ZPS reproduces curated boundary annotations better than the tools it was benchmarked against, and the averaged embedding of each segment proves a strong descriptor of what kind of region it is. Because nothing was fit to existing annotations, the same machinery applies to regions never annotated at all. Published in PLOS Computational Biology in 2025.

#Key Features

• No trained parameters: Segmentation is a change point analysis over frozen embeddings, so there is no model to fit, no labels to supply, and no bias toward the annotation set it is later evaluated on.
• Domains and disorder in one pass: One procedure delivers boundaries for folded domains, sub-domains, and IDRs simultaneously, rather than one tool for profiles and another for low-complexity regions.
• Segment embeddings as descriptors: Averaging per-residue embeddings across a segment yields a vector whose nearest neighbours share the same UniProt annotation, letting one lookup categorize 224 annotation types.
• Discovery of unannotated regions: Clustering segment embeddings isolated unannotated N-terminal segments whose proteins localize to mitochondria, and nearest-neighbour search from FUS recovered SYGQ-rich prion-like domains that BLAST does not find.
• Domain-diagram visualization: Segment embeddings are projected from 1,024 dimensions into RGB and drawn as coloured blocks along the sequence, so segments sharing an annotation share a colour.

#Technical Details

ZPS embeds each protein under 8,000 residues with the ProtT5 encoder, producing an L x 1,024 matrix. A sliding-window change point search with window size 30 compares each 15-residue block to the next under a radial basis function kernel cost; large enough changes become boundaries. On the human proteome this yields 253,000 segments, or 164,000 after correcting over-segmentation.

Against 92,000 annotated UniProt segments, ZPS reaches an average IoU of 0.525, ahead of Pfam (0.364), PROSITE Scan (0.363), and filtered Chi-Score Analysis (0.375). On the 23,000 MobiDB disorder and compositional-bias annotations the margin widens to 0.580 against a best baseline of 0.429. On the 22,000 ProRule domain annotations ZPS scores 0.534, below PROSITE Scan (0.879) and Pfam (0.711): supervised profile tools remain stronger on the annotation type they were built for. By boundary distance, 41.4% of UniProt boundaries have a ZPS boundary within 10 residues, rising to 48.5% on MobiDB. For categorization, one-nearest-neighbour precision separating ProRule domains from MobiDB IDRs is 0.983 and 0.981, averaging 0.986 across the 20 most common domain types; over 150,000 segments carrying 224 labels, average precision is 0.772, against 0.573 for 3-mer composition embeddings.

#Applications

ZPS suits anyone needing region-level structure for proteins that annotation databases cover poorly: IDR biologists mapping the architecture of disordered regions, groups characterizing proteins no Pfam profile matches, and proteome-scale analyses needing one consistent segment definition rather than a patchwork of tool-specific outputs. The segment embeddings support a practical loop — segment a protein, retrieve its nearest neighbours, inspect which annotations are enriched — that yields testable hypotheses about an uncharacterized region. Because only embeddings are required, the approach extends to any proteome ProtT5 can process, and the reference implementation runs from a Google Colab notebook.

#Impact

ZPS shows that protein language model embeddings encode segment structure explicitly enough for a generic change point detector to recover it, without supervision and without exposure to the annotation vocabulary being predicted. That framing, borrowed from zero-shot object discovery in vision, inverts the usual dependency in protein annotation, where coverage is bounded by curation, and yields a method whose blind spots differ from those of Pfam and PROSITE rather than duplicating them. Its limits are equally clear: precision is low because ZPS over-segments relative to curated annotations, supervised profile tools remain better for cataloguing known folded domains, and published results cover only the human proteome. Code and the full set of human segment boundaries and embeddings are released, though the repository carries an academic-use-only notice rather than an open-source licence.