If your protein has a close, experimentally characterised homolog, transfer the annotation from it with BLAST and check the identity against the level of detail you are claiming: three EC digits are safe well below the identity that four digits need (Tian & Skolnick, J Mol Biol 2003). If it does not, run the sequence against InterPro and read the domain architecture before you read any GO term (Blum et al., Nucleic Acids Research 2025). Reach for a learned predictor such as DeepFRI when both come back thin (Gligorijević et al., Nature Communications 2021): that is where it adds information, and also where it is least tested.
These three methods all return the same kind of object, a GO term or an EC number, and that sameness hides the difference that matters. A GO term from a homolog's experiment, a GO term from a curated family mapping and a GO term from a neural network are three different claims. The usual error is not choosing the wrong method. It is reporting a predicted function in the grammar of an observed one.
At a Glance#
| BLAST homology transfer | InterPro / InterProScan | DeepFRI and learned predictors | |
|---|---|---|---|
| Canonical reference | Altschul et al., 1997 (the search); CAFA's BLAST baseline (Radivojac et al., 2013) | Blum et al., 2025; Jones et al., 2014 (InterProScan 5) | Gligorijević et al., 2021 (DeepFRI) |
| Input | a sequence | a sequence, or a UniProt accession for a precomputed match | a structure or contact map, plus sequence |
| What it predicts | whatever the hit was annotated with: GO terms, EC number, name | families, domains, repeats and sites; GO terms through InterPro2GO | GO terms for each ontology; EC numbers; residue-level saliency |
| Evidence it rests on | one characterised neighbour | curated signature models from member databases; a curator-reviewed entry-to-term mapping | patterns learned from annotated structures and homology models |
| Where it is strong | high identity to an experimentally annotated protein | domain-level calls, with coverage of most of UniProtKB | low identity, where nearest-neighbour transfer runs out |
| Where it fails | short or single-domain alignments; propagated errors; multidomain proteins | family membership is not activity; a domain's GO term may not hold for your protein | unseen folds; under-represented enzyme families |
| GO evidence code if deposited | ISS-family codes when curator-reviewed, IEA when electronic | IEA | IEA |
| On .bloom* | the BLAST job (blastp against Swiss-Prot) | interpro_protein (by accession) | not offered |
The two rows that decide what you may write are evidence it rests on and
GO evidence code. Every column produces a prediction. Only an experiment on your protein, or on a protein close enough that the transfer is safe for the detail you are claiming, produces an observation.
The three methods, and the evidence behind each answer#
BLAST transfer: one neighbour, its label#
Homology transfer is the oldest method and still the workhorse. Search the query against a database of annotated proteins, take the best hit, and copy its annotation. CAFA formalised this as its BLAST baseline: the score for a GO term is the maximum sequence identity between the target and any protein experimentally annotated with that term (Radivojac et al., Nature Methods 2013).
Its strength is that the evidence is concrete. You can name the neighbour, the experiment and the paper. Its weakness is that the method hides a strong assumption: that two homologous proteins do the same thing. Homology is a statement about shared ancestry, and function drifts faster than ancestry. The BLAST statistics that tell you two proteins are related tell you nothing about whether they catalyse the same reaction. The trade-offs between pairwise, iterated and profile search are covered in BLAST vs PSI-BLAST vs HMMER. This article is about what you do with a hit once you have it.
InterPro: curated signatures, and a mapping to GO#
InterPro integrates signatures from its member databases into one classification. The 2025 update lists CATH-Gene3D, CDD, HAMAP, NCBIfam, PANTHER, PIRSF, Pfam, PRINTS, PROSITE (patterns and profiles), SFLD, SMART and SUPERFAMILY (Blum et al., 2025). Each is a model of a family, domain or motif: a profile HMM, a pattern, a fingerprint. InterProScan runs a sequence against all of them (Jones et al., Bioinformatics 2014).
Coverage is the method's clearest advantage. In the 2025 release, 81.8% of UniProtKB sequences (201,228,950 of 245,896,766) had at least one InterPro match, rising to 96.8% for reviewed Swiss-Prot entries (Blum et al., 2025). A sequence with no close characterised homolog will usually still match some family or domain.
GO terms reach InterPro through InterPro2GO. Curators annotate each InterPro entry with GO terms, and those terms "can be assigned to all sequences matched by that entry." The curation paper describes InterPro2GO as the largest source of GO annotation predictions for proteins, and describes a protocol that assigns terms to a domain on the function of the domain itself, rather than the function of the families in which it occurs (Burge et al., Database 2012).
That design has one consequence that is easy to miss. The curator reviewed the
mapping, not your protein. If a domain carries the term "kinase activity", every protein containing that domain inherits it, including a pseudokinase that has lost its catalytic residues. The mapping is right about the domain and can still be wrong about your protein.
DeepFRI: function from structure and a language model#
DeepFRI is the clearest example of the learned, structure-aware approach. It has two stages (Gligorijević et al., 2021):
- A protein language model. An LSTM pre-trained on about 10 million Pfam sequences turns each residue into a feature vector, and is frozen afterwards.
- A graph convolutional network over the contact map. Residues are nodes, and two residues are joined if their Cα atoms lie within 10 Å. Graph convolutions pass the language-model features along those contacts, and separate models predict GO molecular function, biological process and cellular component terms, and EC numbers.
It was trained on about 30,000 non-redundant experimental structures from the PDB plus about 220,000 non-redundant SWISS-MODEL homology models. Adding the homology models is what let it predict a much larger set of functions.
Two results from the paper matter for use in practice:
- It tolerates predicted structures. Replacing experimental structures with predicted models cost only a minor drop in performance. For a lipid-binding protein, the correct GO terms appeared once the model passed a TM-score of about 0.58, and a more specific term above 0.73.
- It points at residues. Gradient-weighted class activation mapping assigns each residue a saliency for each predicted function. On 38 enzymes from the Catalytic Site Atlas, that saliency separated catalytic residues with an AUROC of 0.81.
Read the second result carefully. The authors note that for some binding terms the saliency highlights regions conserved among proteins with that function, not the binding residues themselves, and that DeepFRI was not trained to predict catalytic residues. A saliency peak says where the model looked. It is not a site assignment.
Embedding-based transfer: nearest neighbour, in a different space#
A second learned family keeps the logic of homology transfer and changes the distance. goPredSim embeds every protein with a protein language model and transfers GO terms from the nearest annotated protein in embedding space rather than sequence space. Under CAFA3 conditions it reached Fmax of 37 ± 2% for biological process, 50 ± 3% for molecular function and 57 ± 2% for cellular component, which the authors describe as numerically close to the top ten CAFA3 methods. Restricting transfer to neighbours with less than 20% pairwise sequence identity lowered those to 33%, 43% and 53%, still above naive sequence-based transfer (Littmann et al., Scientific Reports 2021).
CLEAN applies the same idea to EC numbers. It passes ESM-1b embeddings through a network trained with a contrastive loss, so that sequences sharing an EC number sit close together, and reports better accuracy, reliability and sensitivity than BLASTp, including on understudied and promiscuous enzymes (Yu et al., Science 2023). At the domain level, deep models trained on Pfam families extended Pfam coverage by more than 9.5% and assigned families to 360 human reference proteome proteins that previously had no Pfam annotation (Bileschi et al., Nature Biotechnology 2022).
What these methods share with DeepFRI is the kind of evidence they produce. The output is a score from a model trained on existing annotation. It can extend annotation into the twilight zone. It cannot be traced to an experiment on your protein or on any single neighbour.
What CAFA actually measured#
The Critical Assessment of Functional Annotation is the benchmark that most claims in this field come back to, and its design is the reason its numbers are worth more than a held-out split. CAFA is time-delayed: targets are released, predictions are frozen, and methods are scored on the proteins that accumulate experimental annotations afterwards (Zhou et al., Genome Biology 2019). No method can have seen the answer.
The main metric is Fmax. For each score threshold, precision and recall are computed per protein and averaged, combined into an F-measure, and the maximum over thresholds is reported (Radivojac et al., 2013). CAFA also reports
Smin, which weights GO terms by their information content so that a method gains little for predicting broad, uninformative terms (Zhou et al., 2019).
Every method is scored against two baselines:
- Naive scores each GO term by its relative frequency in Swiss-Prot and gives every target the same predictions.
- BLAST scores each term by the maximum sequence identity to any protein experimentally annotated with it.
What three rounds found:
- CAFA1 evaluated 54 methods on 866 proteins from 11 organisms. The best methods substantially outperformed first-generation methods, and the authors concluded that they "universally outperform simple functional transfer." The BLAST baseline scored Fmax 0.38 for molecular function and 0.26 for biological process, and the authors attributed the weaker process results to process terms describing a more abstract level of function, less predictable from sequence alone (Radivojac et al., 2013).
- CAFA2 evaluated 126 methods from 56 groups on 3,681 proteins from 18 species. Its top methods beat CAFA1's, which the authors attributed to both more experimental annotation and better methods, and they found that which algorithm counts as top-performing depends on the ontology (Jiang et al., Genome Biology 2016).
- CAFA3 concluded that molecular function and biological process prediction had improved slightly over time and cellular component prediction had not. In molecular function, the top CAFA3 method, GOLabeler, beat all CAFA2 methods by a considerable margin; in biological process the top three beat their CAFA2 counterparts by very small margins; in cellular component the top method was still from CAFA2 (Zhou et al., 2019).
CAFA3 also did something no held-out benchmark can. It ran genome-wide mutation screens in Candida albicans and Pseudomonas aeruginosa for genes involved in biofilm formation and motility, and targeted assays in Drosophila for genes involved in long-term memory, producing new annotations for more than 1,000 genes. Top methods did significantly better than the baselines in C. albicans and D. melanogaster, but the paper reports that performance assessed by the screens was generally lower than in the protein-centric evaluations, and leaves "considerable room and need for improvement" (Zhou et al., 2019).
Two lessons carry into your own work. First, an Fmax around 0.5 is a strong result in this field, which is another way of saying that a predicted GO list is about half right at its best threshold. Second, CAFA's annotations are incomplete by construction, and CAFA's primary metric does not account for that open world: a prediction scored as a false positive may simply be a function nobody has tested yet (Dessimoz et al., Trends in Genetics 2013). Both effects run in the same direction. Benchmark numbers are a rough guide to ranking methods, not a probability that a given term is true of your protein.
How far homology transfer can be trusted#
Because BLAST transfer is the baseline everything else is measured against, it is worth knowing exactly where it breaks.
Enzyme function diverges faster than the folklore threshold. Earlier pairwise studies suggested function is conserved above about 40% identity. Rost argued that those data sets were biased or too small, and his unbiased analysis found that fewer than 30% of pair fragments above 50% identity had entirely identical EC numbers. He also found that even BLAST E-values below 10⁻⁵⁰ did not suffice to transfer enzyme function without errors, and that most misclassifications came from similarity over relatively short regions or from transferring the annotation of a different domain (Rost, J Mol Biol 200200016-5)).
A follow-up that grouped enzymes by both sequence and function reached a more usable rule: 40% identity is still a confident threshold for the first three EC digits, but all four digits need above 60% identity for at least 90% accuracy. The same study found PSI-BLAST E-values only weakly correlated with function conservation by the third iteration (Tian & Skolnick, 2003).
Transfer propagates errors. A study of 37 enzyme families with extensive experimental characterisation measured misannotation in the major databases. Swiss-Prot was close to 0% for most families; GenBank NR, TrEMBL and KEGG showed similar and surprisingly high error levels, averaging 5% to 63% by superfamily. In NR, misannotation rose over time, to about 40% of sequences submitted in 2005, and 85% of NR misannotations were overpredictions, a more specific function than the evidence supported (Schnoes et al., PLoS Computational Biology 2009). The authors point to error propagation through similarity-based annotation as a primary cause.
The practical conclusion is about the database, not the algorithm. Transfer from a curated Swiss-Prot entry whose function rests on an experiment, and check that the alignment covers the domain you are making a claim about. A hit to an unreviewed entry is often a transfer of a transfer.
When a predicted function is reportable#
The Gene Ontology already gives you the vocabulary for this. Every GO annotation carries an evidence code (GO Consortium evidence code guide):
| Category | Codes | What it means for you |
|---|---|---|
| Experimental | EXP, IDA, IPI, IMP, IGI, IEP (high-throughput: HTP, HDA, HMP, HGI, HEP) | someone ran an experiment on this gene product |
| Phylogenetic | IBA, IBD, IKR, IRD | inferred from a curated evolutionary model |
| Computational analysis | ISS, ISO, ISA, ISM, IGC, RCA | inferred by a curator from sequence or other analysis |
| Author / curator statement | TAS, NAS, IC, ND | asserted, not directly evidenced |
| Electronic | IEA | not manually reviewed |
The guide states plainly that IEA annotations are not manually reviewed and cannot generally be traced to an experimental source. InterPro2GO annotations are deposited with the IEA code under the reference GO_REF:0000002, and predictions from any automated pipeline are electronic in the same sense.
A working rule set for writing up function:
- An experimental code on your exact protein is an observation. Cite the reference attached to the annotation, not the database.
- Transfer from an experimentally annotated close homolog is reportable as "by homology to", with the identity, the coverage and the level of detail matched to it: four EC digits only above about 60% identity, three above about 40%.
- An InterPro match is reportable as family or domain membership. Its GO terms are reportable as "predicted from domain content". Before you write down an activity, check that the catalytic residues the family needs are present in your sequence.
- A learned predictor's term is a hypothesis with a score. Report the method, its version and the score, and treat it as a reason to run an assay, not as the result of one. If the protein's fold is new, the DeepFRI paper's own results on unseen CATH folds, with Fmax below 0.3, are the relevant prior.
- Agreement between independent methods is worth stating. A domain match and a homolog's experimental annotation that agree are stronger than either alone. A learned prediction that disagrees with both is the case worth an experiment.
When to use each#
| Situation | Use | Why |
|---|---|---|
| A Swiss-Prot homolog above about 60% identity with experimental annotation | BLAST transfer | Above that identity, all four EC digits transfer with at least 90% accuracy |
| Homologs only between about 40% and 60% identity | BLAST transfer, to three EC digits, plus InterPro | The fourth digit is where function diverges |
| No close characterised homolog | InterPro | Most UniProtKB sequences match some entry, and the call is at domain level |
| A multidomain protein | InterPro first | Transferring the annotation of a different domain is a leading cause of transfer errors |
| A structure or a confident predicted model, and both of the above thin | DeepFRI | Structure-based features add signal at low identity, with a modest drop on predicted models |
| You need a residue hypothesis for a predicted function | DeepFRI saliency, as a hypothesis | Catalytic-residue AUROC of 0.81 on 38 enzymes, not trained as a site predictor |
| EC numbers for an under-represented enzyme family | profile- or embedding-based methods, not DeepFRI alone | FunFams beat DeepFRI on EC prediction, especially for under-represented families |
| A function you will write as established | an experimental evidence code | Everything else is a prediction, whatever its score |
A reasonable default for an uncharacterised protein: BLAST against Swiss-Prot, read the identity and coverage of the best experimentally annotated hit, run InterPro to get the domain architecture, and only then look at GO terms, with their evidence codes. Use a learned predictor when the first two disagree or both come back thin, and write its output as a hypothesis.
Limitations to keep in view#
- Every method inherits the annotation it learned from or copies. A biased annotation database biases all three. Well-studied model organisms and well-studied families dominate experimental GO annotation; goPredSim's paper notes that experimental GO annotation exists for fewer than 0.5% of known proteins.
- GO terms propagate up the ontology. Predicting a specific term implies all of its ancestors, so a long list of predicted terms is often a few predictions restated at several levels of generality. Count the leaf terms.
- Family membership is not activity. InterPro tells you what a protein is related to. Whether it still binds, catalyses or localises the same way is a separate question, and a homolog that has lost a catalytic residue still matches its family's model.
- Benchmarks lag methods. CAFA's time delay is its strength and also means that published rankings describe methods as they were when predictions were submitted.
- Biological process is harder than molecular function. In CAFA1 the BLAST baseline scored 0.38 for function and 0.26 for process, and the top methods showed the same gap. A predicted process term carries less weight than a predicted molecular function at the same score.
- None of these methods says anything about your experimental context. Expression, localisation and post-translational state decide what a protein does in a given cell, and no sequence-based predictor observes them.
How to run it on .bloom*#
On .bloom* you can run homology transfer and annotation lookup, and Ivy chains them: a BLAST search, then the annotation of the hit, then its domain record.
The BLAST job runs blastp against UniProtKB/Swiss-Prot. You give it one protein query, as a FASTA record, raw residues or a UniProt accession, and it returns a ranked hit list with E-values, percent identity and query coverage, plus the pairwise alignments, which you leave with as a hits JSON and an alignment text file. Swiss-Prot is the right database for transfer for the reason in the misannotation section above. Details are on the BLAST page.
uniprot_fetch takes a single UniProt accession and returns the curated record: the function statement, subcellular location, disease associations and sequence features. Use it on your best hit to read what the annotation actually says before you transfer it.
quickgo_annotations takes a UniProt accession and returns the GO annotations recorded for it, each with its GO ID, its aspect, its evidence code (in ECO form) and the reference it rests on. It is how you check whether a hit's function is experimental or electronic before you rely on it.
interpro_protein takes a UniProt accession and returns InterPro's record for that protein: its name, length, how many InterPro entries it matches and a link to its full domain architecture on InterPro. It looks up a precomputed match and does not take a raw sequence, so for a sequence with no accession, BLAST it first, or run it through InterProScan on InterPro's own site, which accepts batches of up to 100 sequences per run (Blum et al., 2025).
⚠ .bloom* does not run a sequence-only learned function predictor today. Ask ESM3 for function prediction and you get a message pointing you to InterPro lookup with a UniProt accession, or to uniprot_fetch for annotated function, instead of a job. DeepFRI is not offered either. If a learned prediction is part of your evidence, run it with the authors' published tools and report it as the hypothesis it is.
If you are not sure which of these fits your question, which tool? is the shorter route than guessing, and the UniProt page covers searching by GO term or keyword.
Run the BLAST-to-annotation chain on your own sequence on the .bloom* platform.
Key Takeaways#
- BLAST transfer is reportable when the hit is experimentally characterised and the identity matches the detail: three EC digits from about 40%, all four only above about 60% (Tian & Skolnick, 2003).
- A low E-value is not a function claim. Even E-values below 10⁻⁵⁰ did not transfer enzyme function without errors (Rost, 200200016-5)), and similarity-based annotation has propagated errors through the major databases (Schnoes et al., 2009).
- InterPro covers 81.8% of UniProtKB and 96.8% of Swiss-Prot (Blum et al., 2025), and its GO terms apply to every protein an entry matches. Read the domain architecture before the terms.
- DeepFRI adds structure-aware signal at low identity and tolerates predicted models, but struggled on unseen folds and lost to FunFams on EC numbers (Gligorijević et al., 2021).
- CAFA's time-delayed design is why its results are trusted: methods have improved, cellular component has not, and experimental screens graded methods lower than protein-centric benchmarks did (Zhou et al., 2019).
- The evidence code is the reportability test. IEA means not manually reviewed (GO Consortium).
- On .bloom*: the BLAST job,
uniprot_fetch,quickgo_annotationsandinterpro_protein; no sequence-only learned predictor. - Related reading: BLAST vs PSI-BLAST vs HMMER for getting the hit in the first place, and GSEA vs over-representation analysis vs WGCNA for what happens downstream when those GO annotations become gene sets.
References#
- Gligorijević, V., Renfrew, P.D., Kosciolek, T. et al. "Structure-based protein function prediction using graph convolutional networks." Nature Communications (2021). doi:10.1038/s41467-021-23303-9
- Blum, M. et al. "InterPro: the protein sequence classification resource in 2025." Nucleic Acids Research (2025). academic.oup.com/nar/article/53/D1/D444
- Jones, P., Binns, D., Chang, H.Y. et al. "InterProScan 5: genome-scale protein function classification." Bioinformatics (2014). doi:10.1093/bioinformatics/btu031
- Burge, S., Kelly, E., Lonsdale, D. et al. "Manual GO annotation of predictive protein signatures: the InterPro approach to GO curation." Database (2012). doi:10.1093/database/bar068
- Altschul, S.F., Madden, T.L., Schäffer, A.A. et al. "Gapped BLAST and PSI-BLAST: a new generation of protein database search programs." Nucleic Acids Research (1997). doi:10.1093/nar/25.17.3389
- Radivojac, P., Clark, W.T., Oron, T.R. et al. "A large-scale evaluation of computational protein function prediction." Nature Methods (2013). doi:10.1038/nmeth.2340
- Jiang, Y., Oron, T.R., Clark, W.T. et al. "An expanded evaluation of protein function prediction methods shows an improvement in accuracy." Genome Biology (2016). doi:10.1186/s13059-016-1037-6
- Zhou, N., Jiang, Y., Bergquist, T.R. et al. "The CAFA challenge reports improved protein function prediction and new functional annotations for hundreds of genes through experimental screens." Genome Biology (2019). doi:10.1186/s13059-019-1835-8
- Dessimoz, C., Škunca, N., Thomas, P.D. "CAFA and the open world of protein function predictions." Trends in Genetics (2013). doi:10.1016/j.tig.2013.09.005
- Littmann, M., Heinzinger, M., Dallago, C. et al. "Embeddings from deep learning transfer GO annotations beyond homology." Scientific Reports (2021). doi:10.1038/s41598-020-80786-0
- Yu, T., Cui, H., Li, J.C. et al. "Enzyme function prediction using contrastive learning." Science (2023). doi:10.1126/science.adf2465
- Bileschi, M.L., Belanger, D., Bryant, D.H. et al. "Using deep learning to annotate the protein universe." Nature Biotechnology (2022). doi:10.1038/s41587-021-01179-w
- Rost, B. "Enzyme function less conserved than anticipated." Journal of Molecular Biology (2002). doi:10.1016/s0022-2836(02)00016-500016-5)
- Tian, W., Skolnick, J. "How well is enzyme function conserved as a function of pairwise sequence identity?" Journal of Molecular Biology (2003). doi:10.1016/j.jmb.2003.08.057
- Schnoes, A.M., Brown, S.D., Dodevski, I., Babbitt, P.C. "Annotation error in public databases: misannotation of molecular function in enzyme superfamilies." PLoS Computational Biology (2009). doi:10.1371/journal.pcbi.1000605
- Gene Ontology Consortium. "Guide to GO evidence codes." geneontology.org/docs/guide-go-evidence-codes