ProtEx

A protein whose nearest characterized relative shares only a quarter of its residues defeats both standard routes to functional annotation. BLAST finds that relative and copies its Enzyme Commission number across, with no learned sense of whether the alignment justifies the label. A classifier trained over a fixed label vocabulary has few examples of a rare Pfam family, and no output unit at all for a family added after training. ProtEx keeps the homology search but changes the network's job: rather than memorizing which sequences map to which class, it compares a query against labeled examples of a class and decides whether it belongs.

The mechanism is per-label and semiparametric. Similarity search returns the query's top 100 annotated training sequences, and the union of their labels becomes the candidate set. For each candidate, ProtEx assembles an input holding the query, the label string, retrieved sequences that carry that label (positive exemplars) and others that do not (negative exemplars); a T5 encoder-decoder then emits a single character, p or n. Because the class boundary rides in the exemplars rather than the weights, extending the retrieval database with unseen classes lets the same checkpoint annotate them.

Developed at Google DeepMind with contributors from Google and the University of Cambridge, ProtEx was posted to bioRxiv in June 2024. It takes the opposite bet from parametric models such as ESM-2 and structure-aware classifiers such as GearNet and ESM-GearNet: a modest model beside a database, not knowledge in parameters.

#Key Features

• Multi-sequence pretraining: ProtEx pretrains on pairs of UniRef90 sequences, predicting the masked residues and a bucketed similarity score between them — teaching alignment and comparison, not reconstruction alone.
• Per-label binary decisions: Each candidate annotation is scored independently against its own exemplars, turning a 17,929-class Pfam assignment into many focused two-way comparisons instead of one enormous output layer.
• Generalization to unseen labels: With 10% of EC and GO classes withheld from fine-tuning but still retrievable, the model reaches 0.970 and 0.839 max F1 on them, above BLAST's 0.964 and 0.816.
• Retriever matched to the task: BLAST supplies exemplars for EC and GO; for the much larger Pfam corpus the authors substitute their own parallelizable per-class local-alignment retriever.
• Sequence in, annotation out: No structural information is used, so the method covers proteins with no experimental structure and disordered proteins that structure-based pipelines handle poorly.

#Technical Details

The neural component is a T5 encoder-decoder with a character-level vocabulary, evaluated at Small (60M) and Base (220M) sizes, with Base used for the headline results. Pretraining samples UniRef90 sequence pairs spread evenly across similarity buckets, masks roughly 10% of residues, and asks the decoder for the masked spans plus a length-normalized Levenshtein similarity. One pretrained checkpoint is shared across every task, then fine-tuned per task; EC and GO inference uses up to two positive and two negative exemplars, Pfam four positive and none.

On the clustered Swiss-Prot splits of ProteInfer, ProtEx reaches 0.958 max micro-F1 for EC and 0.854 for GO, against 0.950 and 0.824 for thresholded BLAST. On the CLEAN evaluations it records 0.932 weighted AUC on NEW-392 — enzymes added to Swiss-Prot after the April 2022 cutoff — and 0.842 on the experimentally characterized Price-149, against CLEAN's 0.740 and 0.733. It reports 0.909 Fmax on the PDB-based EC split at the 95% similarity stratum, ahead of a 15B-parameter ESM-2 with a classifier head (0.892) and ProtST (0.878). On the Pfam clustered split — 1.3M training sequences, 17,929 families, test sequences under 25% identity to training — it reaches 92.6% family accuracy against 89.7% for an ensembled ProtTNN; ablating exemplars drops that to 76.3%.

#Applications

The immediate use is annotation at database scale: assigning EC numbers to enzymes in newly sequenced genomes and metagenomes, propagating GO terms to uncharacterized proteins, and sorting sequences into Pfam families. The gains concentrate where annotation pipelines are least trustworthy — rare classes and "dark matter" sequences with no close characterized relative — which makes it most useful for triaging wet-lab candidates and for curators judging which automated annotations warrant review. New classes arrive by extending the retrieval database rather than retraining.

#Impact

ProtEx makes a pointed argument about where knowledge should live in a function predictor: a 220M-parameter model reading four retrieved neighbors outperforms a 15B-parameter protein language model with a classification head, and structure-aware graph models, on the same splits. It is equally pointed about baselines — a properly thresholded BLAST beats most published neural methods here, a comparison much prior work omitted. The costs are real: inference needs a forward pass per candidate label, and the retrieval database must be maintained. The released repository carries dataset-conversion and evaluation scripts plus the model's predictions on every test split; the trained checkpoints and training code are not distributed. The work remains a preprint, submitted to ICLR 2025 and not peer reviewed.