# Using AlphaFold Structures for Functional Annotation By the end of this guide you will be able to take a set of predicted proteins from a newly sequenced or poorly annotated genome, obtain AlphaFold structure models for them, search those models against the [Protein Data Bank](/blog/guides/protein-data-bank) (PDB) and the AlphaFold [Protein Structure](/blog/guides/protein-structure) Database with Foldseek, judge which structural hits are trustworthy, transfer function only where the evidence supports it, and record every decision in a reusable annotation template. The same pipeline scales from a handful of proteins to an entire proteome. To follow the hands-on steps you need a Linux or macOS terminal with Python 3, a local or browser-based installation of ColabFold, Foldseek installed, and optionally InterProScan for domain annotation and ChimeraX with the ANNOTEX plugin for manual curation. If you only want to try the approach before installing anything, the AlphaFold database website and the AlphaFold Server web interface let you run small parts of this workflow from a browser. ## Why Structure Reaches Where Sequence Similarity Stops Classical annotation pipelines transfer function by sequence similarity. Tools such as BLAST and HMMER find homologs by aligning amino acid sequences, and they work well when a clear relative exists in the database. They fail when a protein has diverged past the point where sequence signals survive. Metagenomic virology is a clear example, where highly divergent viral genomes routinely escape BLAST and HMMER detection, and where structure-based frameworks such as AlphaFold, ESMFold, and Foldseek recover relationships through conserved three-dimensional folds instead [1]. The same problem appears in whole genomes. Microsporidia are intracellular parasites with extremely streamlined, rapidly evolving genomes, so many of their proteins have no recognizable sequence relatives. A recent annotation effort for the microsporidian Vairimorpha necatrix combined sequence and structure similarity to tackle exactly this, using ColabFold for structure prediction and Foldseek for structural matching against the PDB and AlphaFold databases [2]. Protein structure evolves more slowly than sequence. Two proteins can share less than 20 percent sequence identity yet fold into near-identical shapes with conserved active sites, which is enough to propose a shared function if you check the details carefully. AlphaFold models make this signal available for every protein in a genome, not just the small fraction with experimental structures, and precomputed models for most catalogued proteins already sit in the AlphaFold Protein Structure Database, so for many proteomes you can skip prediction entirely. ## Concepts You Need Before Running Anything **pLDDT.** Each residue in an AlphaFold model carries a predicted local distance difference test (pLDDT) score from 0 to 100. Scores above 90 indicate very high confidence, 70 to 90 indicates a confident backbone, 50 to 70 indicates a low-confidence region, and below 50 often means disorder. Regions below 50 should never drive a functional claim. **PAE.** The predicted aligned error (PAE) matrix reports the expected position error between pairs of residues. Low PAE within a domain and high PAE between domains tell you that AlphaFold placed the domains confidently but is guessing about their relative orientation. Judge each domain independently. **TM-score.** The template modeling (TM) score measures global structural similarity between two proteins on a scale from 0 to 1, normalized so that it is length independent. As a rule of thumb used throughout structural biology, a TM-score above roughly 0.5 usually means the same global fold, while values below 0.5 leave the assignment open. Foldseek reports TM-scores for each match. **RMSD.** Root mean square deviation measures the average distance between aligned atoms. Unlike the TM-score, RMSD grows with protein length and only describes the aligned region, so a small RMSD over a tiny aligned fragment means far less than a moderate RMSD over a full-length match. Prefer TM-score plus coverage when comparing hits. **Coverage.** The fraction of each structure aligned to the other. A high TM-score over 15 percent of the query is a domain-level hint, not a whole-protein annotation. ## Tools in This Workflow | Task | Tool | What it gives you | Practical note | |--|--|--|--| | Structure prediction | ColabFold | PDB/CIF models with pLDDT and PAE | Runs locally, in Colab, or through the AlphaFold Server web interface for complexes | | Precomputed structures | AlphaFold Protein Structure Database | Ready-made models for most UniProt proteins | Download from the EBI-hosted site, no GPU needed | | Structure search | Foldseek | Alignments against PDB or AlphaFold DB with TM-scores | Fast enough for proteome-scale queries | | Sequence search baseline | MMseqs2 or BLASTP | Standard homologs to cross-check | Run before the structure search so you know which proteins it must rescue | | Domain and family evidence | InterProScan | Pfam, InterPro, GO terms | Confirms or contradicts structural hits | | Manual curation | ANNOTEX, a ChimeraX plugin | Visual inspection and annotation extension | Designed to combine sequence and structure evidence [2] | | Proteome-scale EC prediction | SAFA | Structure-based classification of Enzyme Commission numbers | Public code and data, built for whole predicted proteomes [3] | ## The Workflow at a Glance 1. Collect predicted proteins as a FASTA file and clean it. 2. Run a baseline sequence search so you know what sequence similarity alone can already explain. 3. Obtain structures, either by downloading AlphaFold database models or predicting them with ColabFold. 4. Search structures against the PDB and AlphaFold databases with Foldseek. 5. Filter and interpret hits using TM-score, coverage, pLDDT, and active-site evidence. 6. Transfer function conservatively and record the evidence for each call. 7. Curate borderline cases visually in ChimeraX with ANNOTEX. 8. Scale to the full proteome with a batch pipeline such as SAFA when needed. ## An AlphaFold Functional Annotation Workflow You Can Run Today This section is the step-by-step tutorial. Every command assumes a working directory containing your proteome file, `proteins.fasta`. ### Step 1: Prepare and Check Your Protein Set Start by counting records and looking for obvious problems. ```bash grep -c ">" proteins.fasta head -2 proteins.fasta ``` Expected output for a small set: ``` 12 >locus_0001 hypothetical protein [Vairimorpha necatrix] ``` Remove sequences shorter than about 30 amino acids, since short peptides produce noisy structures and unstable alignments. If your assembly pipeline produced isoforms, keep one canonical transcript per gene, because duplicated queries waste prediction time and inflate your hit lists. ### Step 2: Run the Baseline Sequence Annotation First Always run sequence-based annotation before structure-based annotation. The baseline tells you which proteins need structural rescue and prevents you from re-annotating proteins that sequence evidence already settles. A quick MMseqs2 search against Swiss-Prot, the manually reviewed subset of UniProt: ```bash wget https://ftp.uniprot.org/pub/databases/uniprot/current_release/knowledgebase/complete/uniprot_sprot.fasta.gz gunzip uniprot_sprot.fasta.gz mmseqs easy-search proteins.fasta uniprot_sprot.fasta seq_hits.m8 tmp ``` Then annotate domains and families: ```bash interproscan -i proteins.fasta -f TSV -goterms -pa ``` Sort the result by bitscore and mark every protein with a strong, full-length sequence hit as already annotated. Build a candidate list of the rest. These uncharacterized proteins are the input to the structure stage. ### Step 3: Obtain AlphaFold Structures ![Comparison table of downloading precomputed AlphaFold models, predicting locally with ColabFold, and using the AlphaFold Server web interface.](/article-images/illustrations/alphafold-functional-annotation-workflow-prediction-routes.webp) You have three routes, and the best choice depends on your proteins. **Route A, download precomputed models.** The AlphaFold Protein Structure Database, hosted by EMBL-EBI at https://alphafold.ebi.ac.uk, holds predicted models for proteins with UniProt accessions. Files follow the naming pattern `AF--F1-model_v4.cif`, where the version number increases as models are refreshed. ```bash wget https://alphafold.ebi.ac.uk/files/AF-P04637-F1-model_v4.cif ``` **Route B, predict locally with ColabFold.** For novel proteins with no UniProt entry, run ColabFold on a machine with a GPU or in the free Google Colab notebook environment. ```bash colabfold_batch proteins.fasta structures/ ``` ColabFold writes one model file per sequence plus per-residue pLDDT values embedded in the output. Compute a mean pLDDT for each model so you can set a quality gate later. ```python from Bio.PDB import MMCIFParser import numpy as np parser = MMCIFParser(QUIET=True) s = parser.get_structure("m", "structures/locus_0001_unrelaxed_rank_001.cif") plddt = [r["B"] for r in s.get_residues() if "B" in r] print(f"{np.mean(plddt):.1f}") ``` **Route C, AlphaFold Server.** The AlphaFold Server is a free web interface from Google DeepMind for predicting structures, including complexes of proteins, [nucleic acids](/blog/guides/nucleic-acids), and small molecules. It is the fastest way to test a single protein or a candidate complex without installing anything. Use it for exploratory questions and switch to local ColabFold for batches. ### Step 4: Search Structures with Foldseek ![Calculation panel showing the TM-score gate applied to the example Foldseek m8 line with 24.2 percent identity and TM-scores 0.61 and 0.58.](/article-images/illustrations/alphafold-functional-annotation-workflow-tm-score-call.webp) Foldseek aligns protein structures directly, so it can find proteins that share a fold but not a sequence. First decide your target database. The PDB gives you experimentally determined structures with rich functional information, and the AlphaFold database gives you near-complete coverage of catalogued proteomes. Download ready-made Foldseek databases: ```bash foldseek databases PDB pdb tmp download ``` If the module names differ in your Foldseek release, list what is available and use the current names. You can also build a database from your own downloaded CIF files, which always works regardless of release changes: ```bash foldseek createdb afdb_cifs/ afdb foldseek easy-search structures/ afdb afdb_hits.m8 tmp -e 0.001 -c 0.5 ``` Each parameter matters. `-e 0.001` sets the E-value cutoff, keeping only alignments unlikely to occur by chance. `-c 0.5` requires that at least half of the shorter protein is covered by the alignment, which filters out fragmentary matches that look significant but explain little. Repeat the search against the PDB database for experimentally grounded hits: ```bash foldseek easy-search structures/ pdb pdb_hits.m8 tmp -e 0.001 -c 0.5 ``` Typical output in the extended m8 format, with three TM-score columns appended to the standard twelve: ``` locus_0001 AF-P12345-F1 0.242 312 236 14 4 318 11 330 3.1e-11 68.4 0.61 0.58 0.59 ``` Reading this line, locus_0001 matches AF-P12345-F1 with 24.2 percent sequence identity over a 312-residue alignment, an E-value of 3.1e-11, and query and target TM-scores of 0.61 and 0.58. That pattern, low sequence identity with a solid TM-score, is exactly the case the structure stage exists to catch [2][1]. ### Step 5: Interpret Hits and Transfer Function Conservatively Apply these gates to every candidate hit, in order. 1. **Model quality.** The aligned region of your query model should sit mostly above pLDDT 70. A confident fold cannot come from a confidently wrong model. 2. **TM-score.** Treat TM-scores of about 0.5 or higher as evidence of a shared fold and lower values as weak hints needing other support. 3. **Coverage.** Both query and target coverage should be substantial for a whole-protein annotation. Partial matches define domain-level annotations only. 4. **Active-site agreement.** If the target is an enzyme, align the sequences implied by the structural superposition and check whether catalytic residues occupy equivalent positions. A conserved fold without conserved catalytic residues frequently means a lost or altered function, which is common in parasites. 5. **Reciprocal consistency.** Search the target back against your proteome. Hits that are not reciprocal are often artifacts. 6. **Independent confirmation.** A Pfam domain, a genomic neighborhood with related genes, or an RNA-seq expression profile consistent with the proposed function should agree. The rice-to-human study discussed below added exactly this kind of supplemental domain evidence before finalizing its calls [4]. A useful decision rule for evidence tagging. If a structure hit passes gates 1 through 4 and the target has experimental functional data, annotate with a phrase such as "function inferred from structural similarity to PDB entry X, TM-score Y". If only gate 5 or 6 remains uncertain, annotate as "putative" and describe the limitation in a note field. Never copy an Enzyme Commission number from a TM-score alone. ### Step 6: Curate Borderline Cases Visually Automated scores miss things a trained eye catches in seconds, such as an unmodeled loop sitting exactly where an active site should be. The ANNOTEX plugin for ChimeraX was built for this step. It lets an annotator open predicted structures alongside sequence and structure-based hits, inspect the superposition, and record curated annotations with proper evidence codes directly from the viewing session [2]. Open your model and the top hit, superpose them, and check the aligned region, the pLDDT coloring, and the geometry of any conserved residues before writing a final call. ### Step 7: Scale to a Whole Proteome For proteome-scale work, the SAFA pipeline automates the pattern above. It was developed and demonstrated on the predicted structural proteome of the peat moss Sphagnum divinum, covering 25,134 primary transcripts, and it aligns every high-confidence model against a nonredundant database of over 90,000 crystallographic structures to assign putative Enzyme Commission numbers across the proteome [3]. All code and data for that project are in public repositories, which makes SAFA a practical template for your own proteome-scale run rather than a method you must rebuild from scratch. Its architecture has one lesson worth copying regardless of whether you use the code itself: predict structures in parallel, filter models by confidence before searching, align against a curated nonredundant structural target set, and classify function only after those filters. ## A Real-World Example: Three Studies, Three Shapes of the Same Workflow **Divergent parasite genome, manual curation.** The Vairimorpha necatrix study is the closest published template for the full pipeline described above. The authors started from a telomere-to-telomere tetraploid genome assembly, extracted 3,080 predicted protein-coding sequences, confirmed 89 percent of them with [RNA sequencing](/blog/guides/rna-sequencing), predicted structures with ColabFold, searched those structures with Foldseek against the PDB and AlphaFold databases alongside sequence-based searches, and curated the combined evidence in ChimeraX using ANNOTEX. Manual curation of the combined sequence and structure hits increased annotation accuracy relative to either evidence type alone [2]. The general lesson holds for any divergent genome: structure search supplies candidates, sequence evidence and expression data corroborate them, and a human closes the remaining gaps. **Cross-species transfer with transcriptome support.** A study of heat stress response in rice integrated public transcriptome data with structures from the AlphaFold Protein Structure Database to annotate understudied plant genes. After identifying heat-responsive gene groups by meta-analysis and confirming stress-related terms by enrichment analysis, the authors performed structural and sequence alignment between rice and human proteins, deliberately focusing on candidates with low sequence similarity but high structural similarity, then folded in shared domain information. This surfaced candidates connected to metal homeostasis, including iron and copper metabolism. The plant2human workflow for the analysis is publicly available, so it doubles as a reusable template for cross-species structural annotation [4]. **Viromics at scale.** In viral discovery, metagenomic sequencing has exposed the limits of BLAST and HMMER on highly divergent viral genomes. Structure-based frameworks such as AlphaFold, ESMFold, and Foldseek enable annotation of these divergent viral proteins through conserved folds, and current proposals combine AI pattern recognition with classical phylogenetics and functional annotation to control the false discovery rates that pure AI approaches carry [1]. If your input is metagenomic rather than a genome assembly, this hybrid framing applies directly. Run your structure search, then phylogenetically place the top hits before claiming function. ## Workflow Format and a Reusable Template Structure your output so that every annotation carries its evidence trail. Two formats matter. **The search output format.** Foldseek writes the tabular m8 format. Columns 1 through 12 match standard BLAST tabular fields, query, target, percent identity, alignment length, mismatches, gap opens, query start, query end, target start, target end, E-value, and bitscore. Columns 13 through 15 hold the query TM-score, target TM-score, and their average. Knowing this format means you can parse any Foldseek result with `cut` or pandas and sort directly on TM-score. ```bash sort -t $'\t' -k15,15gr afdb_hits.m8 | head -20 ``` **The per-protein annotation record.** A minimal YAML template that captures the full evidence chain for one protein. Keep one record per protein in a directory of files, or one record per row of a spreadsheet, whichever your team will actually maintain. ```yaml locus_tag: locus_0001 sequence_length_aa: 318 structure: source: colabfold file: structures/locus_0001_unrelaxed_rank_001.cif mean_plddt: 84.2 low_confidence_regions: "310-318" baseline_sequence_search: tool: mmseqs2 best_hit: "none under e-value 1e-5" structure_search: tool: foldseek target_db: pdb top_hit: "6XYZ" tm_score_query: 0.63 tm_score_target: 0.59 query_coverage: 0.91 target_coverage: 0.88 functional_call: proposed_function: "putative metalloprotease" ec_number: "none assigned" go_terms: ["GO:0008236"] evidence: "structural similarity to PDB 6XYZ, conserved catalytic residues H41 and C145" confidence: medium curator: initials date: 2026-11-11 notes: "N-terminal region below pLDDT 70 excluded from alignment" ``` If you keep records in this shape, a collaborator, reviewer, or automated pipeline can retrace every call without repeating the analysis, which is the single biggest difference between an annotation that survives peer review and one that does not. ## Checking That Your Results Are Right Three checks catch most problems. First, verify a random sample of 20 to 50 annotations by hand in ChimeraX, including all of your highest-impact calls such as virulence factors or enzyme assignments. Second, compare your structure-based annotations against the sequence-only baseline. Proteins that both methods agree on are nearly always safe, and the disagreement set is where your structure stage earns its keep, so review every disagreement individually. Third, look for internal consistency signals such as expression data or genomic context agreeing with the structural call, which is the strategy the rice study used when it layered transcriptome evidence and shared domains on top of structural alignment [4]. If a structural call contradicts all other evidence, downgrade it and record why. Sensible next steps after a first pass are to submit your curated annotation set to a community resource where appropriate, deposit predicted models in a repository such as ModelArchive the way the Sphagnum project did [3], and rerun the pipeline when Foldseek, ColabFold, or database versions change, since both the models and the search indices improve over time. ## Common Mistakes and Limitations **Treating a high TM-score as a function.** A shared fold constrains what a protein might do, it does not prove what it does. Many folds support multiple functions, and convergent evolution produces look-alike folds with unrelated mechanisms. Anchor functional transfer to experimental data on the target whenever possible. **Using low-confidence model regions.** Alignments that depend on residues below pLDDT 50, which usually indicates disorder, are built on regions AlphaFold itself does not commit to. Intrinsically disordered regions have no fixed structure to align in the first place, so this workflow will systematically miss proteins whose function lives entirely in disorder. **Skipping the sequence baseline.** Structure search is slower and noisier than sequence search. Run sequence methods first and reserve structural effort for proteins sequence methods leave unexplained, which is the ordering used in the published workflows [4][2]. **Ignoring false discovery rates.** AI-driven and structure-search approaches trade sensitivity for precision. Reviews of AI in viromics flag elevated false discovery rates and limited explainability as standing problems, and recommend pairing AI pattern recognition with classical phylogenetic and annotation checks [1]. Budget time for that pairing. **Forgetting that predicted complexes can mislead.** AlphaFold interfaces are predictions. An interface with moderate pTM and high PAE across the interface may be a docking artifact, and any functional claim that depends on a predicted interaction needs a flag saying so. **Letting database staleness creep in.** The AlphaFold database refreshes its models as the underlying network improves, and Foldseek database modules change between releases. Pin versions in your methods section and rerun searches after major updates, because a result reproducible today may shift with the next model refresh. **Over-annotating by automation alone.** In the Vairimorpha study, adding manual curation on top of automated sequence and structure hits increased accuracy [2]. Automated lists are candidate lists. The last mile is human. **One honest scope limit.** This workflow proposes functions, it does not measure them. Any single functional claim, especially one that will guide wet-lab experiments or clinical or agricultural decisions, needs direct experimental validation by the appropriate specialists before it is treated as fact. ## Frequently Asked Questions ### Can I use AlphaFold structures for annotation without predicting structures myself? Yes, for any protein with a UniProt accession. The AlphaFold Protein Structure Database already contains precomputed models that you can download and search directly. You only need ColabFold for novel proteins absent from the database. ### What TM-score should I trust for transferring function? A TM-score around 0.5 or higher usually indicates a shared global fold, which is the starting point for a functional hypothesis rather than proof of function. Below that, treat hits as weak hints that need domain, active-site, or contextual evidence to support them. ### Is there a difference between ColabFold and the AlphaFold Server? Yes. ColabFold is an open-source implementation you can run in a browser notebook or on your own hardware, which suits batch prediction. The AlphaFold Server is a free web service from Google DeepMind for predicting individual structures and complexes interactively. Use the server for quick questions and ColabFold for pipelines. ### How is this workflow different from BLAST or HMMER annotation? BLAST and HMMER detect similarity in [amino acid sequence](/blog/guides/amino-acid-sequence), and they miss proteins whose sequences have diverged too far. Structure-based search detects conserved three-dimensional folds, so it recovers distant relationships sequence methods cannot see, which is why divergent genomes and viruses are its best use cases [2][1]. ### What do the extra columns in Foldseek m8 output mean? The standard twelve columns mirror BLAST tabular output. Foldseek appends three more, the query TM-score, the target TM-score, and their average, so you can sort directly on structural similarity without running a separate TM-score calculation. ### Can I annotate an entire proteome at once? Yes. The SAFA approach did this for the 25,134-protein predicted structural proteome of Sphagnum divinum, aligning every high-confidence model against over 90,000 crystallographic structures to assign putative EC numbers [3]. The same design, predict, filter by confidence, search, then classify, scales to other proteomes. ### How do I know if a predicted structure is good enough to use? Check per-residue pLDDT and the PAE matrix. Restrict any functional argument to regions above pLDDT 70, treat regions below 50 as likely disordered, and read the PAE matrix to confirm that the domains your argument depends on were each modeled confidently. ### Is there an alphafold tutorial I can follow for this workflow? Yes. ColabFold maintains official notebooks that walk through prediction step by step, Foldseek documents its search commands on its GitHub wiki, and the plant2human workflow from the rice heat stress study is publicly available as a complete worked example of structure-based cross-species annotation [4]. ## Sources 1. [Charting the virosphere: computational synergies of AI and bioinformatics in viral discovery and evolution.](https://pubmed.ncbi.nlm.nih.gov/41222234/) 2. [Functional annotation of a divergent genome using sequence and structure-based similarity.](https://pubmed.ncbi.nlm.nih.gov/38166563/) 3. [Predicted structural proteome of Sphagnum divinum and proteome-scale annotation.](https://pubmed.ncbi.nlm.nih.gov/37589594/) 4. [Functional annotation of novel heat stress-responsive genes in rice utilizing public transcriptomes and structurome.](https://pubmed.ncbi.nlm.nih.gov/41675695/)