{"id":"ProtNLM","title":"Automatic annotation using ProtNLM","lastModified":"2026-09-03","type":"help","content":"\nUniProt’s [Automatic Annotation pipeline](https://www.uniprot.org/help/automatic_annotation) enhances [unreviewed](https://www.uniprot.org/help/uniprotkb_sections)/TrEMBL [UniProtKB](https://www.uniprot.org/uniprotkb) entries with automatic classification and annotation. To do this, UniProt leverages machine learning and artificial intelligence (AI) using a **Protein Natural Language Model (ProtNLM)** developed by collaborators at [**Google DeepMind**](https://deepmind.google/science/). ProtNLM is trained on a snapshot of UniProtKB data to generate annotations for unreviewed UniProtKB entries (UniProtKB/TrEMBL).\n\nSimilar to how natural language models translate between human languages or generate text, ProtNLM “translates” amino acid sequences into biological annotations.\n\n![ProtNLM Pipeline Overview](https://github.com/ebi-uniprot/uniprot-manual/raw/main/images/ProtNLM_pipeline_overview.png)  \n*Overview of the ProtNLM pipeline translating raw amino acid sequences into natural-language biological annotations for unreviewed UniProtKB/TrEMBL entries.*\n\n## How to find and view ProtNLM annotations\n\nProtNLM predicted protein names have been provided for proteins previously named “uncharacterized protein” since UniProt release 2022\\_04. These have been integrated into UniProtKB in a similar way to names predicted by other automatic annotation methods. These annotations can be recognised by the “Automatic Annotation ([ECO:0008006](https://www.ebi.ac.uk/QuickGO/term/ECO:0008006)) Google:ProtNLM” evidence tag, for example in entries [A0A6E8VIU4](https://www.uniprot.org/uniprotkb/A0A6E8VIU4/entry#names_and_taxonomy) and [Q54TU2](https://www.uniprot.org/uniprotkb/Q54TU2/entry#names_and_taxonomy).\n\nTo search or filter for all entries annotated with a ProtNLM predicted protein name, use the search term `source:google` in the UniProt search bar.\n\n### Expansion of annotations beyond protein names\n\nWhile originally used only to predict protein names, an expanded **ProtNLM2** model now provides predictions for:\n\n* [Protein names](https://www.uniprot.org/help/protein_names)  \n* [Protein function comments](https://www.uniprot.org/help/function)  \n* [Subcellular locations](https://www.uniprot.org/help/subcellular_location)  \n* [Keywords](https://www.uniprot.org/help/keywords)  \n* [Gene Ontology (GO) terms](https://www.uniprot.org/help/gene_ontology)\n\n![ProtNLM Annotation Pipeline](https://github.com/ebi-uniprot/uniprot-manual/raw/main/images/Protnlm_annotation_pipeline.png)  \n*The functional annotation scope of ProtNLM2, covering protein names, function comments, subcellular locations, and GO terms, in addition to keywords.*\n\nA pilot release of ProtNLM2 annotations for approximately [26,000 TrEMBL entries](https://ftp.ebi.ac.uk/pub/contrib/UniProt/ProtNLM2/List_of_UniProt_accessions_that_have_ProtNLM2_annotations.tsv) from about 400 of the most searched species in UniProt has been made public in conjunction with the UniProt 2026\\_02 release.  ProtNLM2 predicted annotations are not integrated into the UniProtKB entries directly but are stored in parallel and can be viewed on the website by switching on the AI annotation toggle found at the top of the entry page next to the [entry name](https://www.uniprot.org/help/entry_name) of ProtNLM2 annotated entries, for example in entries [A0A7N5KEZ3](https://www.uniprot.org/uniprotkb/A0A7N5KEZ3/entry) and [C5XPJ5](https://www.uniprot.org/uniprotkb/C5XPJ5/entry). AI generated annotations are clearly differentiated by a purple background color and sparkle icon.\n\n![ProtNLM2 Example Entry](https://github.com/ebi-uniprot/uniprot-manual/raw/main/images/ProtNLM2_example_entry_image.jpg)  \n*Viewing ProtNLM2 predictions on a UniProtKB entry page via the AI annotations toggle.*\n\n**FTP download**: A complete list of TrEMBL entries annotated using ProtNLM2 is available on our [FTP site](https://ftp.ebi.ac.uk/pub/contrib/UniProt/ProtNLM2/List_of_UniProt_accessions_that_have_ProtNLM2_annotations.tsv). This file provides a list of approximately 26K protein entries with ProtNLM2 annotations. The protein names present in this file are those present in UniProt and are not protein names generated by ProtNLM2.\n\n\n## Quality control and corroboration\n\nPredicted annotations have been continually assessed throughout the development of the ProtNLM models, both by automated evaluation as well as manual evaluation by expert UniProt biocurators working in close collaboration with Google DeepMind. However, machine learning AI models do make errors. We invite users to carefully evaluate these predictions and flag any issues or errors they find via the “Entry feedback” link in the corresponding entry, or via the [UniProt help desk](https://www.uniprot.org/contact).\n\n### ProtNLM model score\n\nEach ProtNLM prediction is assigned a model score representing an estimate of the likelihood of the prediction being generated. This score reflects the model's probability of generating the prediction based on the input sequence. The score ranges from 0 to 1, with higher scores indicating a higher estimated likelihood.\n\nFor ProtNLM protein name ([Recommended Name](https://www.uniprot.org/help/protein_names)) predictions, we apply a model score threshold of 0.2, removing any predictions that fall below this.\n\nFor ProtNLM2 predictions, we have reduced the model score threshold to 0.05 as many predictions with low model scores were judged to be accurate. For predictions above this threshold, Google DeepMind have implemented an additional post-processing strategy (described below) to filter out unreliable predictions. Based on ongoing curator evaluation and user feedback on the ProtNLM2 pilot release, this model score threshold may be altered (increased or removed) for future ProtNLM2 releases. Model scores for all predictions are provided for the user to make their own assessment on prediction reliability.\n\n### The Evidencer: prediction evaluation\n\nTo ensure the reliability of the data presented to users, ProtNLM2 predictions above the model score threshold are further evaluated using a post-processing system known as the *Evidencer*. This pipeline attempts to emulate expert [biocurator evaluation](https://www.uniprot.org/help/biocuration) by filtering out unreliable predictions and requiring corroborating evidence before a prediction is displayed. It consists of two parts:\n\n*1. Exclusion Evidencer*\n\nPredictions are rejected if they match exclusion criteria such as GO annotations that are restricted by [GO taxon constraints](https://geneontology.org/docs/taxon-constraints/#:~:text=Taxon%20constraints%20are%20used%20to,specificity%20of%20the%20GO%20terms), or inaccurate protein names that do not conform to the [International Protein Nomenclature Guidelines](https://www.uniprot.org/help/international_protein_nomenclature_guidelines).\n\n*2. Corroboration Evidencer*\n\nPredicted annotations are corroborated by:\n\n* **Exact/Partial Matches:** Predicted annotations are corroborated by looking for exact or partial (substring) matches with other annotations in the entry being annotated. The corroboration process also queries several external databases such as [InterPro](https://www.ebi.ac.uk/interpro/), and checks for an exact or partial (substring) match (designated as a hydrated string/substring match) within descriptions of signatures cross-referenced within the entries being annotated.\n* **Sequence Similarity:** If no string/substring match is found, the corroboration process checks for sequence similarity (determined by the [phmmer tool](https://www.ebi.ac.uk/Tools/hmmer/search/phmmer)) to other entries in UniProtKB containing annotations identical to the predicted annotation. Predicted annotations are considered reliable if they have matching annotations in entries with a sequence similarity phmmer bit score greater than 25.\n* **Structural Similarity:** If there is no corroboration by sequence similarity, the process checks structural similarity (determined by the [tm-align tool](https://www.rcsb.org/alignment)) to other entries in UniProtKB containing annotations identical to the predicted annotation. Corroboration requires the TM-align score normalized to the protein with the predicted annotation to be greater than 0.5 and the TM-align score normalized to the corroborating protein possessing a matching prediction to be greater than 0.1. Additionally, structural similarity corroboration requires a minimum sequence identity of 5%, a minimum of 3 matching secondary structure elements and a minimum pLDDT score for aligned residues of 70:\n\n  |                                            |        |\n  |:-------------------------------------------|--------|\n  | TM-align score (protein with prediction) | >= 0.5 |\n  | TM-align score (protein with matching annotation) | >= 0.1  |\n  | Sequence identity | >= 5% |\n  | Matching secondary structure elements | >= 3 |\n  | pLDDT of aligned residues | >= 70  |\n\nThe *Evidencer* workflow produces three possible outcomes:\n\n* **Rejected**: Predictions rejected by the *Exclusion Evidencer* are not displayed in UniProtKB.\n* **Approved**: Predictions corroborated by the *Corroboration Evidencer* are displayed in UniProtKB.\n* **Uncertain**: Predictions that are neither rejected nor corroborated have uncertain quality and are not displayed in UniProtKB.\n\n![ProtNLM Evidencer Workflow](https://github.com/ebi-uniprot/uniprot-manual/raw/main/images/ProtNLM_evidencer_flow.png)  \n*Post-processing corroboration workflow evaluating predicted annotations through string matches, sequence homology (phmmer), and structural alignment (TM-align).*\n\n### Summary\n\nPredictions are excluded if their ProtNLM model score is below 0.05, or if they are excluded by the *Exclusion Evidencer* or if they cannot be corroborated by the *Corroboration Evidencer*.\n\n\n### Interpreting prediction evidence information\n\nWe provide *Evidencer* information for ProtNLM2 predictions to help users evaluate their accuracy. These can be viewed by clicking the “Automatic Annotation” tag next to the predicted information. This will open a text box containing [evidence](https://www.uniprot.org/help/evidences) in support of the prediction. Several types of evidence statements are given:\n\n* **Exact/Partial matches**: An exact string or substring match of the prediction is found in the entry. Currently, predictions that are an exact match to an annotation in an entry (e.g. a predicted function statement matching the function statement already in the entry) are not shown. Alternatively, the match is found in the description of a term in an external database cross-referenced in the entry. The cross-referenced databases queried using this approach include [InterPro/Pfam](https://www.ebi.ac.uk/interpro/), the [Gene Ontology (GO) knowledgebase](https://geneontology.org/), and the [Enzyme nomenclature database](https://enzyme.expasy.org/) for [EC (Enzyme Commission) numbers](https://en.wikipedia.org/wiki/Enzyme_Commission_number). A link is provided to the matching database entry.  \n* **Sequence similarity**: The predicted annotation is found in other entries with sequence homology to the entry. A link is provided to the entry with the highest similarity as determined by phmmer alignment. The higher the provided similarity bit score, the more significant the sequence similarity.  \n* **Structure similarity**: The predicted annotation is found in other entries that share structural homology with the entry. A link is provided to the entry with the highest similarity as determined by TM-align. The TM-align score ranges from 0 to 1, with a score of 1 indicating a perfect match and structures with scores of \\<0.2 considered unrelated while those with scores \\>0.5 are considered to have a highly similar protein fold.\n\nIn the case of sequence and structure evidence, links are included to perform sequence alignments using the UniProt [Align tool](https://www.uniprot.org/align) or structural alignments using the FoldSeek [FoldMason tool](https://search.foldseek.com/foldmason). This allows users to evaluate the prediction evidence themselves.\n\n## Technical methodology\n\nProtNLM uses the amino acid sequence as input and processes it through a [transformer-based sequence-to-sequence AI model](https://storage.googleapis.com/brain-genomics-public/research/proteins/protnlm/uniprot_2022_04/protnlm_preprint_draft.pdf) for whole-protein functional annotation in an approach similar to AI natural language models that generate titles or captions for images. Each prediction is assessed by an independent bioinformatics pipeline that provides corroborating evidence found in entries with sequence and/or structural similarity\n\nThe current ProtNLM2 model generates predictions using a single model trained entirely on amino acid sequences. It was trained on 240 million protein entries from UniProt release 2023\\_04, using both curator-reviewed/Swiss-Prot entries and unreviewed/TrEMBL entries, including those annotated by the UniProt [Automatic Annotation pipeline](https://www.uniprot.org/help/automatic_annotation).\n\nThe original iteration of ProtNLM, trained on protein name data from UniProt release 2021\\_02 (including all recommended names (RecNames), alternative names (AltNames) and submitted names (SubNames)), has provided predicted protein names for proteins with the name “Uncharacterized protein”. From UniProt release 2022\\_04 onwards, these predicted protein names have been included in unreviewed/TrEMBL entry pages with the “Automatic Annotation ([ECO:0008006](https://www.ebi.ac.uk/QuickGO/term/ECO:0008006)) Google:ProtNLM” evidence tag.\n\nProtNLM has undergone continuous evaluation and refinement since its introduction:\n\n* **Release 2022\\_04**: The original iteration of ProtNLM was used to predict protein names for proteins labeled “Uncharacterized protein,” using a single model.  \n* **Release 2022\\_05**: Predictions were updated using an ensemble of 3 models that take only the amino acid sequence as input, and 3 models that take both the amino acid sequence and the organism taxonomy identifier (TaxID) as input.  \n* **Release 2023\\_01**: A model score threshold was introduced, along with improved post-processing, to enhance prediction accuracy.  \n* **Release 2023\\_02**: A new ensemble of 7 models was trained, including 1 model that used AlphaFoldDB secondary structure predictions as input.  \n* **Ongoing Evaluation**: Continuous assessment revealed that performance gains could be maximized using an ensemble of models trained on sequence and sequence \\+ organism TaxID, so models using secondary structure predictions were deprecated.\n* **Release 2026\\_02 (ProtNLM2)**: An initial ProtNLM2 pilot release, expanding annotations beyond protein names for roughly 26,000 TrEMBL entries, was made public.\n\n## Links\n\n- [ProtNLM preprint](https://storage.googleapis.com/brain-genomics-public/research/proteins/protnlm/uniprot_2022_04/protnlm_preprint_draft.pdf)\n- [YouTube video](https://www.youtube.com/watch?v=FLkoaDJBC54)\n- [Explore all entries with ProtNLM protein name predictions](https://www.uniprot.org/uniprotkb?query=%28source:google%29)\n- [List of all ProtNLM2 pilot release annotated entries](https://ftp.ebi.ac.uk/pub/contrib/UniProt/ProtNLM2/List_of_UniProt_accessions_that_have_ProtNLM2_annotations.tsv)","categories":["UniProtKB","Automatic annotation","help"]}