# How to Annotate Novel Protein NCBI CDD By the end of this guide you will be able to take a [protein sequence](/blog/guides/protein-sequence) of unknown function, run it against NCBI's Conserved Domain Database (CDD), decide which hits you can trust, record the annotation in a reusable format, and use the CDD alignment and linked structures as templates for downstream modeling. You will know how to annotate a single sequence, hundreds of sequences, or a whole proteome, and how to cross-check the result against other databases. Here is what you need on hand before starting: - A protein sequence in FASTA format, either a single sequence or a multi-FASTA file - A web browser for the interactive CD-Search service - Optional: Python 3 with the `requests` library, for batch submissions - Optional: NCBI EDirect utilities and the BLAST+ suite, for command-line work - Optional: the Cn3D structure viewer if you want to inspect domain models on 3D structures ## What the Conserved Domain Database Is The Conserved Domain Database is NCBI's public resource for annotating proteins with the location of conserved domain footprints and the functional sites inferred from those footprints [1]. A domain, in this context, is a segment of a protein that folds and often functions as a semi-independent unit, and that recurs across many proteins because it was conserved through molecular evolution. CDD represents each domain family as a [multiple sequence alignment](/blog/guides/multiple-sequence-alignment) and a derived profile search model [2]. Two properties of CDD matter most when you annotate a novel protein: 1. **CDD mixes curated and imported models.** NCBI curators build and refine models for well-characterized families, using protein 3D structures to fix domain boundaries and to annotate conserved catalytic and binding sites [1]. CDD also imports domain models from external providers such as Pfam, SMART, and COG, so the full collection is deliberately partially redundant [3]. 2. **Annotation is layered.** Redundant and homologous models are clustered into superfamilies, and every footprint is labeled at minimum with a superfamily designation. On top of that, a hit can be flagged as specific, meaning it identifies molecular function with high confidence, or non-specific, meaning it only establishes superfamily membership [2]. When you annotate a novel protein, this flag is the single most consequential field in the output, because it tells you how far you can safely push your functional claim. CDD has been publicly available for more than 20 years, and its curation effort now also assigns names to recurring domain architectures, meaning the ordered combination of domains along a protein, because a domain list alone often does not pin down function [4]. CDD maintains both an archive of pre-computed annotation for proteins tracked in Entrez and a live search system for sequences not yet in that archive [1]. ## An NCBI Databases Guide for This Workflow If you are still learning how to use NCBI, the key point is that CDD sits inside the Entrez system, NCBI's integrated query and retrieval network, and it cross-links to several neighboring databases [3]. For a domain annotation project you will touch five of them: | Entrez database | What it gives you in this workflow | |--|--| | Protein | The query sequence itself, plus pre-computed "Conserved Domain" links for proteins already tracked in Entrez [3] | | CDD | Domain footprints, alignments, functional sites, and superfamily hierarchy [1] | | Structure | 3D structures linked to domain models, viewable in Cn3D [5] | | PubMed | The primary literature cited on domain model pages [2] | | SPARCLE | Functional labels assigned to distinct subfamily domain architectures [6] | A practical habit: start every session from the CDD home page at https://www.ncbi.nlm.nih.gov/Structure/cdd/cdd.shtml [4]. From there, every result page links sideways to the aligned members, the 3D structures, and the literature, which is what turns a raw hit list into an evidence trail. ## Step 1. Prepare Your Protein Sequence CD-Search matches profile models against protein sequences [3]. Nucleotide queries are accepted and are translated before searching [5], but for a novel protein you will usually already have the translated product. If your protein came out of a [genome annotation pipeline](/blog/guides/genome-annotation-pipeline), keep the accession and version number. For bacterial genomes, the [NCBI Prokaryotic Genome Annotation Pipeline](/blog/guides/ncbi-prokaryotic-genome-annotation-pipeline) generates the protein set that Entrez tracks, and those accessions resolve cleanly to pre-computed CDD annotation when available [7]. Fetch a sequence from Entrez with EDirect if you work on the command line: ```bash efetch -db protein -id YP_498675.1 -format fasta > yp498675.fasta ``` Your FASTA file should look like this (residue lines abbreviated): ``` >YP_498675.1 hypothetical protein [Staphylococcus aureus] MKT... ``` Three preparation checks that prevent most downstream confusion: 1. **Confirm the sequence is a full-length protein, not a fragment.** A truncated query produces truncated domain footprints, and you will misread the domain architecture. Compare your query length against the expected size of the family you end up matching. 2. **Keep the FASTA header short and unique.** In batch output files, the first word of the header is your join key. Spaces in headers get truncated, so `>my_protein_phage_29_tail` becomes `my_protein_phage_29_tail` only if you avoid spaces. Use underscores. 3. **Mask nothing before the first run.** CD-Search applies its own internal filtering. Run your first pass on the raw sequence, then consider low-complexity handling only if the result looks noisy. ## Step 2. Annotate a Single Sequence With the CD-Search Web Service The interactive CD-Search service is the fastest path for one to a few sequences. It lives at http://www.ncbi.nlm.nih.gov/Structure/cdd/wrpsb.cgi and searches CDD profile models with BLAST heuristics [3] [2]. The walkthrough: 1. Open the CD-Search page. Paste your FASTA sequence into the query box, or upload the file. 2. Select the database. The default is the full CDD collection, which includes NCBI-curated models plus imported Pfam, SMART, and COG models [3]. If you want only the hand-curated NCBI models, restrict to those, because they carry the finest-grained functional descriptions [1]. 3. Set the expect value (E-value) threshold. The E-value estimates how many matches of this quality you would expect to find by chance in a database of this size. Values below roughly 0.001 are generally treated as strong evidence of homology, while anything above about 0.01 deserves skepticism. Keep the default for a first pass and tighten it when you write up results. 4. Submit and wait. The live search uses RPS-BLAST, a reverse version of position-specific BLAST that scans a query against pre-built profile databases, and returns within seconds to a few minutes for a single protein [5]. **What the output looks like.** The result page has three layers: - A graphical overview, with the query drawn as a horizontal bar and each domain footprint drawn as a colored block positioned along the sequence. Overlapping blocks of different colors at the same coordinates mean the same region matched several redundant models. - A hit list, ordered with specific hits first. Each row carries the model name, accession, E-value, bit score, and alignment coordinates. - A pairwise alignment view for each hit, where the query is aligned to the domain consensus and conserved motifs are visible. **How to read the hit class.** If the top hit is a specific hit, the model resolves to a family with a defined function, and you can cite that function with confidence [2]. If the best you get is a non-specific hit, its accession usually begins with `cl` for clade or superfamily, and your honest annotation is "member of superfamily X with undetermined specific function." Many genuinely novel proteins land here, and that is a real result, not a failure. The 2008 CDD release introduced this specific versus non-specific flagging precisely so that downstream users know which assignments carry high confidence [2]. ## Step 3. Annotate Many Sequences With Batch CD-Search Batch CD-Search computes and downloads annotation for large sets of protein queries [8]. This is the right service once your novel-protein list grows past a screenful, for example all hypothetical proteins from a genome or plasmid. Submit a multi-FASTA file from Python: ```python import requests with open("proteins.fasta") as handle fasta_data = handle.read() response = requests.post( "https://www.ncbi.nlm.nih.gov/Structure/bwrpsb/bwrpsb.cgi", data={"queries": fasta_data, "db": "cdd", "outfmt": "3"} ) with open("batch_results.tsv", "w") as out out.write(response.text) ``` Parameters that matter: - `db=cdd` searches the full combined collection. You can restrict to a single imported source if you want, for example, only Pfam-style families. - `outfmt=3` returns the standard tabular format, one line per query-hit pair. A verbose tabular option and machine-readable JSON formats are also available if you prefer to parse structure rather than text. - For very large batches, the server may return a Search ID instead of immediate results. Re-submit with the Search ID as the `queries` value to poll until the finished annotation appears. This asynchronous queueing is how the service handles proteome-sized input. The exact column layout of the tabular output is documented on the Batch CD-Search page, and the fields you will use most are the query identifier, the hit class (specific or non-specific), the CDD model accession, the superfamily assignment, the alignment start and end coordinates on the query, the E-value, and the bit score [8]. A realistic fragment of standard tabular output (values illustrative): ``` Query Hit type Accession Superfamily From To E-value Bits prot_001 specific cd06213 cl21446 12 270 3.1e-48 187.4 prot_002 non-specific cl00203 cl00203 41 190 2.8e-09 51.2 prot_003 - - - - - - - ``` The third line is the pattern you must plan for. Some novel proteins return no domain footprint at all. NCBI's curation staff actively surveys this unannotated fraction of sequence space, sometimes called protein dark matter, and adds models for new families as evidence accumulates, but a truly novel domain can outrun the database [4]. ## Step 4. Run RPS-BLAST Locally Against CDD For reproducible pipelines, high-volume annotation, or work behind a firewall, download the pre-built CDD profile databases from the NCBI FTP site at ftp.ncbi.nlm.nih.gov/pub/mmdb/cdd/ and run RPS-BLAST from the BLAST+ suite on your own hardware [5]. Install BLAST+, download and unpack the little-endian archives for the full CDD collection (or the curated-only collection if you prefer), and run: ```bash rpsblast -query proteins.fasta -db Cdd -evalue 0.01 -out cdd_hits.tab -outfmt "6 qseqid sseqid evalue bitscore qstart qend" ``` Expected output is standard BLAST tabular, one line per hit (values illustrative): ``` prot_001 cd06213 3.1e-48 187.4 12 270 prot_002 cl00203 2.8e-09 51.2 41 190 ``` Parameters worth controlling: - `-db Cdd` points at the full collection. Substituting the curated-only database when you downloaded it keeps specific-hit claims conservative. - `-evalue 0.01` is a reasonable screening cutoff. Record whatever you use, because your threshold is part of your method. - `-outfmt 6` gives you tabular output that parses cleanly into pandas, R, or a spreadsheet. Checking that the local run is right takes two minutes. Pick one protein with a known domain, run it through both the web CD-Search and your local RPS-BLAST, and confirm the same model appears at the same coordinates with a comparable E-value. If the two disagree, the usual cause is a stale local database, so re-download before debugging anything else. ## An Annotate Novel Protein NCBI CDD Example A published example shows the full decision path end to end. Chakma and colleagues took the hypothetical protein YP_498675.1 from Staphylococcus aureus, a 281-residue protein of no assigned function, and characterized it computationally [9]. Their sequence of operations is the template worth copying: 1. **Basic sequence characterization first.** They established the protein's physicochemical profile, secondary structure content (predominantly alpha-helical), stability, and cytoplasmic subcellular localization using multiple predictors. This context frames everything that follows. 2. **Domain annotation with CDD plus independent databases.** They ran the sequence through NCBI-CDD, Pfam, and InterProScan, and all three pointed to the same answer: the protein carries a pyridoxal phosphate (PLP)-dependent 2,3-diaminopropionate biosynthesis protein SbnA domain, placing it in the tryptophan synthase beta superfamily [9]. The convergence of three independent systems is what turns a computational hit into a defensible annotation. 3. **Homology modeling on a structure-based template.** Using the HHpred server, they built a 3D model of the protein on the template structure of the SbnA enzyme bound to PLP, PDB entry 5D84, an X-ray diffraction model [9]. The CDD hit named the family, and the family pointed to the modeling template. Notice the logic of the citation chain. CDD's curated models are built so that model boundaries agree with domain boundaries observed in 3D structures, and the models represent the structurally conserved cores of families [1]. That structural grounding is exactly why a confident CDD specific hit is such a strong starting point for choosing a modeling template. If your novel protein returns a specific hit, follow the same chain: CDD model, then the linked structures, then a modeling server. A second example shows the scale of the same approach. During the COVID-19 pandemic, researchers annotated the SARS-CoV-2 proteome through CDD to map specific functional features such as ATP and ADP binding domains, visualizing the domain models on 3D structures in Cn3D to prioritize drug design targets [10]. The lesson transfers to any novel protein set: CDD gives you the specific feature locations, and a structure viewer turns them into testable hypotheses. ## Reading the Annotate Novel Protein NCBI CDD Format The "format" question has two answers, the input you submit and the output you record. Input is plain FASTA, which every tool in this workflow accepts. Output is where precision pays off, so here is the field-by-field breakdown of a domain annotation record. | Field | Meaning | How to use it | |--|--|--| | Query ID | First word of your FASTA header | Join key for merging results across tools | | Hit type | Specific or non-specific [2] | Specific hits support named function claims, non-specific hits support superfamily claims only | | Model accession | CDD model identifier, for example a `cd` number for curated models, `pf` or `smart` or `COG` style identifiers for imported models, `cl` for superfamilies [3] | Stable key for reporting and for retrieving the alignment | | Coordinates (from, to) | Residue range of the footprint on your query | Draw your own domain architecture, check whether domains overlap or abut | | E-value | Chance-corrected measure of match significance | Tighten thresholds before publication | | Bit score | Normalized alignment score, higher is better | Comparable across models of different lengths | | Superfamily | Cluster designation covering redundant and homologous models [2] | Your floor-level annotation when nothing specific matches | | Functional sites | Catalytic or binding residues annotated on the model and transferred to your query [8] | Direct hypotheses for mutagenesis or active-site confirmation | Two format conventions save hours of cleanup later. First, archive the raw service output, never just your spreadsheet summary, because reviewers sometimes ask for the untouched table. Second, when you merge CDD output with Pfam or InterProScan output, join on coordinates as well as protein ID, since one protein can carry two partial copies of the same superfamily and naive ID-only joins collapse them. For a reusable annotation template, keep a single line per protein in this shape: ``` protein_id length_aa hit_type cdd_accession model_name superfamily evalue bitscore qstart qend conserved_sites evidence_level notes ``` The `evidence_level` column is where you encode your own confidence calls, for example `specific_CDD`, `superfamily_only`, `no_hit`, or `conflict_between_databases`. ## Using CDD Models as Templates "Template" carries three working meanings in this workflow, and all three are useful. **Template as a reusable record-keeping scaffold.** The single-line template above, applied to every protein in your set, becomes your annotation table for a manuscript, a thesis chapter, or a database submission. Fill it from the tabular output of Steps 2 through 4. **Template as a domain alignment.** Each CDD model is a multiple sequence alignment of representative family members, built to cover the structurally conserved core and annotated with conserved features [1]. When your query hits a model, the alignment view shows exactly which residues of your protein correspond to conserved columns. Catalytic residues marked on the model transfer to your query if the alignment is sound, which is how CDD infers functional sites on proteins with no experimental data [8]. **Template as a 3D structure.** Because CDD curation anchors models to known domain boundaries in protein structures, and most known structures are represented by tracked models [5], a specific hit connects your novel protein to experimentally determined homologs. The SbnA example followed this chain to PDB 5D84 as the modeling template [9]. For a novel protein with a strong specific hit, this is the highest-yield next step available, and it costs one visit to the linked structure records. CDTree, NCBI's helper application for examining curated domain hierarchies, deserves mention if your project is about classifying many related proteins rather than annotating one. CDD and CDTree used together let you place sequences within a domain family hierarchy, which is a powerful approach for protein classification [3] [5]. ## Key Options at a Glance | Decision | Web CD-Search | Batch CD-Search | Local RPS-BLAST | |--|--|--|--| | Input scale | One sequence at a time [5] | Large sets, proteome scale [8] | Any size, hardware-limited | | Reproducibility | Manual, screenshot the result | Recorded Search ID and downloaded table | Fully scripted, pinned database version | | Best for | First look at a single novel protein | Systematic annotation of a protein set [8] | Pipeline integration and large repeated runs [5] | | Output | Graphical and tabular, interactive | Tabular files for parsing [8] | Standard BLAST tabular | | Database choice | Full CDD or curated subset | `db=cdd` or restricted collections | Whatever archive you downloaded | ## Common Mistakes and Limitations **Treating a non-specific hit as a function call.** This is the most consequential error in the workflow. A non-specific hit establishes superfamily membership only [2]. Writing "the protein is a kinase" when the evidence supports "the protein belongs to the [protein kinase](/knowledge/molecular-biology/protein-kinase) superfamily" overstates your result and invites reviewer correction. **Trusting a single database.** CDD imports models from multiple providers and the collections partially overlap [3]. A hit that appears in CDD, Pfam, and InterProScan simultaneously, as in the SbnA example, is far stronger than a hit in one system [9]. Cross-check before you commit. **Over-interpreting domain boundaries.** Domain boundary assignment is a prediction problem in its own right, and dedicated machine-learning tools for residue-level domain partitioning are benchmarked against CDD and InterPro rather than treated as equivalent to them [11]. Treat the coordinate range of a footprint as an approximation, and expect boundary differences of a few residues between systems. **Forgetting that pre-computed and live annotation can differ.** The archive of pre-computed annotation is large but updating it has slowed the cadence of CDD releases [4]. A protein you fetched from Entrez may show stale annotation compared with a live CD-Search run on the same sequence. When precision matters, run the live search. **Chasing low E-values on short alignments.** A tiny E-value over a 30-residue overlap can describe a fragment match to one helix of a domain, not the domain. Always check that the footprint covers a substantial fraction of the model before calling the family. **Ignoring the no-hit case.** A protein with no domain footprint anywhere is not a failed analysis. CDD curators monitor unannotated sequence space and add new family models as structural and literature evidence appears [4], so a no-hit result today may resolve in a future release. Re-run periodically and, in the meantime, lean on fold-recognition tools rather than profile searches. **Not recording your thresholds.** The E-value you accept is a methodological choice. Report it alongside results, because a hit at 1e-50 and a hit at 0.009 do not carry the same weight. One general limitation applies to every claim in this workflow. Computational annotation generates hypotheses about function, and individual cases require experimental confirmation before functional statements enter the literature. ## Next Steps After Annotation Once the domain architecture of your novel protein is settled, several productive directions open up: - **Subcellular localization.** Conserved functional domains carry real predictive signal for where a protein acts in the cell, and domain features are used directly as inputs to subcellular localization models [12]. Combine your CDD superfamily assignments with a localization predictor. - **Network and structural context.** If you work with protein interaction networks, the cddApp extension for Cytoscape loads CDD domain and site information onto network nodes identified by NCBI or UniProt identifiers and visualizes that information together with 3D structures through the structureViz2 integration [13]. - **Architecture-based function assignment.** Search SPARCLE, the Subfamily Protein Architecture Labeling Engine, with your domain architecture. CDD's curation effort assigns functional characterizations to distinct subfamily architectures, so a match there can upgrade a superfamily-only annotation [6]. - **Formal classification of families.** For large sets of related proteins, CDTree within the CDD framework supports hierarchical classification [3]. - **Submission.** If your novel protein comes from your own sequencing, submit the annotated sequence with your domain assignments to GenBank, where the NCBI annotation pipeline will process it and, in time, attach the pre-computed CDD annotation that Entrez serves to future users [7]. ## Frequently Asked Questions ### What does a non-specific CDD hit actually mean for my novel protein? It means the sequence region matches models from a cluster of homologous domain families, so you know the superfamily but not the specific family or function [2]. Report it as superfamily membership and look for additional evidence, such as key residues in the alignment or a domain architecture match in SPARCLE, before claiming a molecular function. ### What E-value cutoff should I use when annotating with CDD? ![Worked calculation showing the 0.01 screening cutoff and 0.001 strong homology threshold applied to an illustrative hit.](/article-images/illustrations/annotate-novel-protein-ncbi-cdd-evalue-cutoff.webp) A common screening cutoff is 0.01, with anything below roughly 0.001 treated as strong homology evidence. Record the cutoff you use, inspect the alignment extent for each accepted hit, and tighten the threshold for publication-level claims, because a specific-hit designation already reflects high-confidence family assignment [2]. ### Can CD-Search annotate a nucleotide sequence directly? Yes, the interactive CD-Search service accepts nucleotide queries as well as protein queries and handles the translation [5]. For multi-kilobase nucleotide input you will generally get cleaner results by predicting open reading frames first and submitting the proteins. ### How many sequences can I submit to Batch CD-Search at once? Batch CD-Search is built for large sets and handles proteome-scale input by queueing [8]. For very large submissions the service returns a Search ID that you re-submit to retrieve the finished annotation, so structure your script to poll rather than expect a synchronous response. ### How should I cite CDD in a manuscript? Cite the current CDD database paper, the 2023 [Nucleic Acids](/blog/guides/nucleic-acids) Research update [4], plus the specific CDD accession numbers for each domain you report, for example in a table column. If you used the web service, batch service, or SPARCLE, the corresponding methods papers [8] [6] are the standard citations. ### Why do CDD and Pfam sometimes disagree about where a domain starts and ends? CDD's curated models use protein 3D structures to refine boundaries toward structurally observed domain cores [1], while Pfam models are built from sequence alignments alone. Imported Pfam models sit inside CDD alongside curated ones [3], so the two can legitimately disagree by a few residues. Report the coordinates from the system you cite and note the disagreement when it changes your interpretation. ### What should I do if my novel protein has no CDD hit at all? ![Flowchart for a protein with no CDD hit: confirm full length, cross-check other databases, try fold recognition, re-check later.](/article-images/illustrations/annotate-novel-protein-ncbi-cdd-no-hit-path.webp) Confirm the sequence is full-length, run it against other profile systems such as Pfam and InterProScan, and try fold-recognition methods such as HHpred that work at a more distant level [9]. Then re-check CDD in a later release, since curators actively add models for newly characterized families emerging from structure determination efforts [4]. ### Can I get pre-computed CDD annotation for proteins without running a search? Yes, for proteins tracked in Entrez, pre-computed domain annotation can be retrieved as "Conserved Domain" links on protein records, for single sequences or in bulk [1] [3]. Live search is still the right choice for sequences not yet in the archive or when you need the newest models. ## Sources 1. [CDD: NCBI's conserved domain database](https://pubmed.ncbi.nlm.nih.gov/25414356/) 2. [CDD: specific functional annotation with the Conserved Domain Database](https://pubmed.ncbi.nlm.nih.gov/18984618/) 3. [CDD: a conserved domain database for interactive domain family analysis](https://pubmed.ncbi.nlm.nih.gov/17135202/) 4. [The conserved domain database in 2023.](https://pubmed.ncbi.nlm.nih.gov/36477806/) 5. [CDD: conserved domains and protein three-dimensional structure](https://pubmed.ncbi.nlm.nih.gov/23197659/) 6. [CDD/SPARCLE: functional classification of proteins via subfamily domain architectures](https://pubmed.ncbi.nlm.nih.gov/27899674/) 7. [Whole genome sequence of Salmonella Rissen SCSW714, a porcine strain harbouring a novel multidrug-resistant Tn7-like pco- and sil-containing transposon.](https://pubmed.ncbi.nlm.nih.gov/35500837/) 8. [CDD: a Conserved Domain Database for the functional annotation of proteins](https://pubmed.ncbi.nlm.nih.gov/21109532/) 9. [In silico analysis of a novel hypothetical protein (YP_498675.1) from Staphylococcus aureus unravels the protein of tryptophan synthase beta superfamily (Try-synth-beta_ II)](https://pubmed.ncbi.nlm.nih.gov/37995054/) 10. [Protein Domain Annotations of the SARS-CoV-2 Proteomics as a Blue-Print for Mapping the Features for Drug and Vaccine Designs](https://doi.org/10.5614/JMS.2020.25.2.1) 11. [Seq2DomML: Protein domain boundary prediction from bacterial protein sequences using a machine learning framework](https://doi.org/10.1186/s12859-026-06666-y) 12. [A Novel Multi-label Human Protein Subcellular Localization Model Based on Gene Ontology and Functional Domain](https://doi.org/10.1145/3592686.3592754) 13. [cddApp: a Cytoscape app for accessing the NCBI conserved domain database](https://pubmed.ncbi.nlm.nih.gov/25212755/)