Loading…
Multimodal protein function annotation that scores a sequence against free-text descriptions, including GO and EC labels unseen during training.
No providers recorded yet. Browse all providers
The Gene Ontology gains roughly 298 terms and retires 591 every year. A predictor built the usual way — a fixed output layer with one neuron per GO term — cannot represent a function coined after its training snapshot, and keeps scoring a vocabulary the curators have moved on from. The same wall stands between it and any annotation scheme it was never trained on, such as Enzyme Commission numbers.
ProtNote replaces the fixed vocabulary with text. Instead of classifying a sequence into N known functions, it takes a protein and a free-form description of a function and predicts the probability that the two go together — a binary match over an open label space. Because the label enters as a sentence rather than an index, any function a scientist can describe in words is a valid query, whether or not it existed at training time. That reframing lets one fixed checkpoint serve both supervised annotation over familiar GO terms and zero-shot prediction of functions no training protein was ever annotated with.
ProtNote was developed at Microsoft Research by Samir Char, Nathaniel Corley, Sarah Alamdari, Kevin K. Yang and Ava P. Amini, posted as a preprint in October 2024 and published in Bioinformatics in April 2025. It competes with sequence-only predictors such as ProteinBERT and with homology search, but is closer in spirit to vision-language retrieval: two encoders, one per modality, joined by a head that scores their compatibility.
Training uses SwissProt sequences under ProteInfer's random split — 418,015 training sequences carrying 31,365 distinct GO labels — with term descriptions from the 1 July 2019 GO release. Each encoder's output passes through its own MLP with three hidden layers of 3,072 units and a 1,024-dimensional output; the two projections are concatenated and fed to a fusion MLP of three 3,072-unit layers ending in a single output neuron, using ReLU activations, batch normalization and no bias terms. Training ran 46 epochs on eight 32 GB V100 GPUs at an effective batch size of 256 (Adam, learning rate 3e-4, gradient clipping at 1, mixed precision), with BLOSUM62 substitutions corrupting sequences as augmentation.
On supervised GO prediction it reaches 0.9042 mAP micro against ProteInfer's 0.9032 — statistically indistinguishable — while trailing on mAP macro (0.6018 versus 0.6417), the price of the more general architecture. The zero-shot settings are where it separates: on GO terms introduced between July 2019 and May 2024, over sequences added in the same window, ProtNote scores 0.2139 mAP macro against 0.1556 and 0.1236 for E5 and BioGPT description-similarity baselines. On the 2,534 unseen EC labels it reaches 0.3062 mAP macro and 0.3456 mAP micro, against 0.2058/0.1107 for E5 and 0.0961/0.0714 for BioGPT. All figures average five seeds.
Fewer than 1% of UniProt entries carry human-verified functions, and the gap is widest where predefined label sets are least useful — metagenomic sequences, engineered variants, newly described activities. ProtNote lets a researcher pose the question in their own words: score candidates against a described catalytic activity, screen for enzyme classes with no trained classifier available, or re-annotate a collection against a current ontology release. The released checkpoints are intended for research use, and perform best on descriptions written in the style of functional annotations rather than arbitrary text.
ProtNote shows that protein function prediction can be posed as text-conditioned matching rather than fixed-vocabulary classification, and that doing so costs little on the labels a conventional classifier already handles. The tradeoffs are real: inference requires one forward pass per protein–text pair, expensive when scoring many sequences against many functions; it does not exploit the ontology's hierarchy; and its judgments inherit the vocabulary and style of the Gene Ontology text it was trained on. Code is MIT-licensed, and weights for all five seeds, the datasets and the computed metrics are archived on Zenodo, making the benchmarks straightforward to reproduce.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.