# NCBI Protein Database Conserved Domains: Step-by-Step Guide By the end of this guide you will be able to take any [protein sequence](/blog/guides/protein-sequence), find the conserved domains inside it using the [NCBI protein database](/blog/guides/ncbi-protein-database) conserved domains resources, read the results with confidence, and scale the same analysis to hundreds or thousands of sequences. You will run a single-sequence CD-Search in the browser, run a Batch CD-Search on a set of proteins, and reproduce the same search on the command line with RPS-BLAST plus rpsbproc. Here is what you need on hand: - A web browser for the CD-Search and Batch CD-Search web tools - One or more protein sequences in FASTA format, or [NCBI protein](/blog/guides/ncbi-protein) accessions such as NP_000537.3 - For the command-line workflow, the NCBI BLAST+ executables and the rpsbproc package installed on a Linux or macOS machine, plus a few gigabytes of disk space for the domain profile database - Optional, a spreadsheet program for sorting batch results ## What a Conserved Domain Is and Why It Matters A protein domain is a segment of a protein that folds and functions as a recognizable unit. A conserved domain is a domain whose sequence has been maintained over evolution because it performs a specific biochemical job, such as binding DNA, hydrolyzing ATP, or catalyzing a phosphoryl transfer. When you find a conserved domain in a new protein, you inherit a large body of knowledge about what that region probably does. The NCBI Conserved Domain Database (CDD) is the resource that formalizes this. CDD annotates protein sequences with the locations of conserved domain footprints, together with functional sites and motifs inferred from those footprints [1]. A domain model in CDD is a [multiple sequence alignment](/blog/guides/multiple-sequence-alignment) of representative domain members, distilled into a position-specific scoring matrix that a search tool can compare against your query. The current generation of CDD contains more than 57,000 domain models, of which almost 15,000 were curated in house by NCBI staff, with the rest imported from external providers such as Pfam, SMART, COG, KOG, and TIGRFAMs [1][11]. Two properties of CDD matter for your analysis: 1. Curated models are organized into hierarchies of related families and subfamilies, so a hit can tell you not just that a protein is a kinase, but which functional subfamily of kinases it belongs to [2]. 2. The SPARCLE system (Subfamily Protein Architecture Labeling Engine) classifies whole proteins by their complete domain architecture, meaning the ordered combination of domains across the sequence, which is often more informative than any single domain hit [2]. The domain models themselves are searched with RPS-BLAST (reverse position-specific BLAST), which compares a query sequence against a database of precomputed position-specific scoring matrices. The E-value you see in your results is the expectation value, the number of hits with a score at least this good that you would expect to find by chance in a database of this size. Lower is better, and anything near 1e-3 or below is usually a genuine domain relationship. CDD serves two kinds of data. For proteins tracked by the NCBI Entrez [protein database](/blog/guides/protein-database), there is an archive of precomputed domain annotations. For any sequence you supply yourself, including novel or predicted proteins, you can run a live search with CD-Search on the web, Batch CD-Search for many sequences, or standalone RPS-BLAST plus the rpsbproc annotation package locally [1]. ## How to Use NCBI to Get Your Query Sequence If your protein is already in the NCBI protein database, fetch it in FASTA format with E-utilities. The worked example throughout this guide uses human cellular tumor antigen p53, accession NP_000537.3, a 393 amino acid protein. ```bash curl "https://eutils.ncbi.nlm.nih.gov/entrez/eutils/efetch.fcgi?db=protein&id=NP_000537.3&rettype=fasta&retmode=text" > p53.fa head -2 p53.fa ``` Expected output: ``` >NP_000537.3 tumor protein p53 [Homo sapiens] MEEPQSDPSIEPPLSQETFSDLWKLLPENNVLSPLPSQAMDDLMLSPDDIEQWFTEDPGP... ``` If your sequence comes from a genome annotation, a transcriptome assembly, or your own translation of an open reading frame, save it as a standard FASTA file with a single line starting with `>` and the [amino acid sequence](/blog/guides/amino-acid-sequence) below it. Translate any nucleotide query to protein first, since CDD models describe protein domains. Searching a [nucleotide sequence](/knowledge/molecular-biology/nucleotide-sequence) against protein domain profiles is one of the most frequent beginner errors, covered in the mistakes section below. ## Running a Single-Sequence CD-Search Step by Step ![Six-step flow for submitting one protein to CD-Search, from opening the page to opening the alignment view of a hit.](/article-images/illustrations/ncbi-protein-database-conserved-domains-single-search-steps.webp) This is the fastest path and the first of the three official CDD protocols [1]. 1. Open the CDD home page at https://www.ncbi.nlm.nih.gov/Structure/cdd/cdd.shtml and click the link for CD-Search. 2. Paste your FASTA sequence into the query box, or upload the FASTA file. For the example, paste the contents of `p53.fa`. 3. Choose the database. The default is the full CDD collection. Selecting the NCBI-curated subset restricts hits to models curated in house, which tend to carry richer functional site annotations and hierarchical classification [1][11]. For a first look at a well-studied protein, the default full collection is fine. 4. Leave the E-value threshold at the default unless you have a reason to change it. You can tighten it later if the results feel noisy. 5. Click "Submit" or "Query sequence". The job runs in seconds for a single protein. 6. Read the graphical summary first. Each detected domain appears as a colored footprint drawn on a bar representing your protein, so you immediately see domain boundaries, overlaps, and coverage. 7. Open the tabular hit list below the graphic. Each row shows the domain name, the source database (curated, Pfam, SMART, COG, and so on), the E-value, the bitscore, the aligned interval on your query, and the alignment status. 8. Click any hit to open the multiple sequence alignment of your query against the domain model. This alignment view also marks functional sites conserved across the model, for example catalytic residues or metal-binding positions [1]. For the p53 example, you should see a large, highly significant hit covering roughly the central DNA-binding portion of the protein, from a p53-family domain model, with an E-value far below 1e-40, followed by smaller models toward the C-terminus corresponding to the tetramerization region. The alignment view annotates residues that contact DNA and coordinate a zinc ion, which is exactly the kind of inferred functional site CDD is designed to report [1]. ## Running a Batch CD-Search for Many Sequences When you have dozens to thousands of proteins, the Batch CD-Search tool saves you from submitting sequences one at a time [1]. 1. Prepare a single text file. Either a multi-FASTA file with one protein per record, or a plain list of NCBI accessions, one per line. Both are accepted input types. 2. Open the Batch CD-Search page linked from the CDD home page. 3. Upload your file or paste its contents, then set the search database and E-value options as you would for a single search. 4. Submit the job. You receive a request identifier, and the results become available as a downloadable tabular report rather than an interactive page. 5. Download the tabular results. Each line corresponds to one domain hit, with the query identifier, domain accession and name, source database, E-value, bitscore, and query coordinates. Depending on the output option you select, you also get conserved site annotations and the classification of each hit as a specific or superfamily-level match. 6. Sort and filter in a spreadsheet. A practical first pass is to filter on E-value, then collapse to one best hit per query, then check how many queries were left with no hits at all. Two practical notes for batch runs. First, keep a record of the request identifier together with your input file, because results are served through a session and you want to be able to reproduce the exact run. Second, if your accession list contains many redundant near-identical proteins, consider deduplicating first, since highly similar sequences inflate runtime without adding information. ## Reproducing the Search on the Command Line with RPS-BLAST and rpsbproc The third official protocol uses standalone tools, which is the right choice when you want the analysis inside a reproducible pipeline or need to run it on a server [1]. Install BLAST+ through your package manager, for example `conda install -c bioconda blast`, and download rpsbproc from its NCBI distribution page. Then download the prebuilt CDD profile database volumes from the NCBI FTP server and extract them all into one directory: ```bash mkdir -p cdd_db && cd cdd_db curl -s https://ftp.ncbi.nlm.nih.gov/blast/db/ | \ grep -o 'Cdd\.[0-9]*\.tar\.gz' | sort -u | \ while read f; do curl -sO "https://ftp.ncbi.nlm.nih.gov/blast/db/$f" tar xzf "$f" && rm "$f" done ls | head ``` Expected output is a directory listing of many extracted profile database files rather than a single archive, for example: ``` Cdd.00.phr Cdd.00.pin Cdd.00.psq Cdd.01.phr ... ``` Also fetch the auxiliary annotation data package linked from the rpsbproc distribution page, since rpsbproc needs it to convert raw RPS-BLAST hits into curated annotations. With the database in place, run the search: ```bash rpsblast -query p53.fa \ -db cdd_db/Cdd \ -evalue 1e-5 \ -num_threads 8 \ -outfmt "6 qseqid sseqid evalue bitscore qstart qend" \ -out p53_rps.tsv head -3 p53_rps.tsv ``` Expected output shape, with values illustrative of the format: ``` NP_000537.3 cd00001 2.1e-58 212.4 94 286 NP_000537.3 pfam00870 3.7e-55 201.8 95 286 NP_000537.3 smart... 6.0e-40 160.2 96 283 ``` Raw RPS-BLAST output gives you coordinates and scores but no curated functional annotation. That is what rpsbproc adds. It takes the same query file, runs against the same database, and post-processes hits into CDD's curated hierarchy with functional sites: ```bash rpsbproc -i p53.fa -o p53_annotated.tsv -m 4 ``` The `-m` option controls verbosity of the output, and mode 4 reports the fullest annotation, including query data, the specific hit and its superfamily, domain intervals, and site information. The exact column layout is documented with the rpsbproc package, and the general structure is one row per hit with the query identifier, the specific domain model, its parent superfamily, the E-value, the aligned interval, and any annotated functional sites. The following table summarizes the key knobs across all three routes to CDD data. | Option or parameter | Where it appears | What it does | Practical guidance | |--|--|--|--| | Search database selection | CD-Search, Batch CD-Search | Full CDD collection versus NCBI-curated subset | Use curated subset when you want site annotations and subfamily detail, full set for maximum sensitivity [1] | | E-value threshold | All three routes | Maximum expectation value reported per hit | Keep the default for single proteins, tighten for large noisy batches | | Query format | CD-Search, Batch CD-Search | FASTA sequence versus accession list | FASTA for your own sequences, accession lists for proteins already in Entrez | | Output format | Batch CD-Search | Interactive report versus tabular download | Use tabular output for any set you plan to filter or count | | `-outfmt 6` fields | RPS-BLAST | Columns of the tabular output | Include qseqid, sseqid, evalue, bitscore, qstart, qend as a minimum | | `-m` mode | rpsbproc | Amount of annotation written per hit | Mode 4 for full annotation including sites [1] | | Request identifier | Batch CD-Search | Retrieval handle for a submitted job | Record it so you can retrieve the same run later | ## Reading Your Results Correctly ![Comparison table of specific versus superfamily domain hits by depth, annotation, and what each supports in a report.](/article-images/illustrations/ncbi-protein-database-conserved-domains-specific-vs-superfamily.webp) Three distinctions in the output decide whether your interpretation holds up. **Specific hits versus superfamily hits.** A specific hit is the deepest, best-matching model, often a curated subfamily model that carries functional site annotations. A superfamily hit tells you which broad, evolutionarily related group the region belongs to. A protein may have a strong superfamily hit and a weaker specific hit, which usually means it belongs to the family but not to any characterized subfamily. SPARCLE formalizes this by labeling proteins by their complete subfamily architecture [2]. **Significance.** Judge each hit by its E-value in combination with alignment coverage. A hit with an E-value of 1e-60 covering 180 amino acids is a different animal from a hit with an E-value of 0.05 covering 40 residues near the end of a protein. The standard practice is to treat the default threshold as a screen and confirm borderline hits by opening the alignment and checking that the conserved catalytic or binding residues are actually present in your query [1]. **Multiple overlapping hits.** CDD deliberately includes models from many source databases, so one region of your protein can match a curated model, a Pfam model, and a SMART model that all describe the same domain family. This is redundancy by design, not an error. Collapse these to the most specific model when you report results, and note the source database. A compact way to report per-protein results is a table like this, using the p53 example: | Query | Best specific hit | Source | E-value | Query coordinates | Interpretation | |--|--|--|--|--|--| | NP_000537.3 | p53-family DNA-binding domain | Curated | far below 1e-40 | central region of a 393 aa protein | DNA-binding domain with annotated DNA-contact and zinc-binding sites | | NP_000537.3 | p53 tetramerization region | Curated | very low | C-terminal region | oligomerization domain | ## How Published Studies Use CDD in Real Workflows Seeing how the tool is used end to end helps you place it in your own pipeline. **Annotating an unknown protein.** A study of a hypothetical protein from Staphylococcus aureus used NCBI Conserved Domain Search together with InterProScan and Pfam to assign the protein to the YqeG family of HAD IIIA-type phosphatases within the haloacid dehalogenase superfamily, then proceeded to secondary structure prediction, homology modeling, and docking [3]. The domain search came first because it anchored every downstream step to a functional family. **Predicting domains before docking and dynamics.** A study of placental proteins against osteoarthritis targets used the NCBI conserved domain database to predict functional protein domains, then combined that with motif discovery using the MEME suite and the MyHits database before protein-protein docking and molecular dynamics [4]. Domain locations told the authors which residues to constrain at the binding interface. **Vaccine and epitope work.** For the Brucella melitensis L7/L12 protein, a 124 amino acid ribosomal protein, the authors ran CD-Search to confirm ribosomal protein-related conserved domains before mapping phosphorylation sites and predicting B-cell and T-cell epitopes [5]. A similar pattern appears in epitope prediction work on HHLA2 and MAGEB5 protein variants, where NCBI Conserved Domain analysis established which regions were conserved before immune epitope prediction [6]. **Prioritizing mutations.** A screen of non-synonymous SNPs in the Mycobacterium tuberculosis PPE68 protein focused structural modeling and simulation on mutations falling in conserved regions of the N-terminal domain, including substitutions at positions 26, 44, 92, 163, and 167, because mutations inside conserved domains are the ones most likely to change function [7]. **Building phylogenetic markers.** For cotton leaf curl Multan virus, the conserved CRESS domain of roughly 300 bp within the replication-associated protein served as an alternative molecular marker, and comparative phylogenies built from 278 retrieved domain sequences clustered consistently with full-length Rep trees [8]. Conserved domains make excellent short phylogenetic markers precisely because they are the parts of a protein that resist change. **Family-wide surveys.** A data-mining survey of 646 PELP1 proteins across NCBI, UniProt, and EukProt showed that every member contained one RIX1 domain and one or two NUC domains, with LXXLL and PXXP motifs, and used that conserved architecture to trace the protein's distribution to the Opisthokonta clade [9]. In the Alba domain family, iterative PSI-BLAST searches against the NCBI non-redundant protein database recovered 15,161 domain-containing proteins, which were then clustered into 13 subgroups by sequence similarity network analysis [10]. Gene family papers on plant proteins follow the same recipe, for example the pomegranate sucrose synthase family analysis, where conserved motifs and domains supported the subdivision of five genes into subgroups [11]. **Spotting divergent members.** In a survey of the AlkB family in poultry, conserved domain comparison revealed that ALKBH4 lacked the typical AlkB structural motifs, a finding that changed how the authors interpreted that gene's function [12]. This is one of the most valuable uses of the tool, since a family member missing expected motifs often behaves differently from its relatives. **Counting conserved residues across large sets.** An analysis of 57,311 SARS-CoV-2 spike proteins from the NCBI Virus database found the receptor-binding domain and three surface N-glycosylation sites highly conserved at 99.67 percent [13]. Large-scale retrieval plus domain-focused comparison is a standard way to quantify conservation before designing therapeutics. ## Tips That Make the Results Better - **Run the curated subset when annotation quality matters.** NCBI-curated models carry the hierarchical classification and site annotations that imported models often lack [1]. - **Always open the alignment for any hit you plan to report.** The graphical footprint tells you where the domain is. The alignment tells you whether your protein actually retains the conserved residues that justify the functional label [1]. - **Use SPARCLE when the whole architecture matters.** For enzymes and signaling proteins, the ordered combination of domains frequently separates proteins with different functions inside one superfamily [2]. - **Pair CDD with an independent resource.** Published annotation workflows routinely run CDD alongside InterProScan and Pfam and report hits from both [3]. Agreement between independent systems is cheap, strong evidence. - **Deduplicate before batch searches.** Near-identical sequences from the same species or strain slow the run and clutter the output table. - **Use domains to focus downstream work.** Published pipelines map mutations [7], design docking constraints [4], and choose epitope scaffolds [3][4] inside or near conserved domains rather than at random across the protein. - **Mine the conserved site annotations.** Sites inferred from the domain model, such as catalytic residues, metal-binding positions, and glycosylation-relevant regions, are reported directly in the alignment view [1][9]. - **For family surveys, combine retrieval with iterative searching.** The Alba domain study's approach, PSI-BLAST against the non-redundant protein database followed by clustering, recovered a large and well-partitioned dataset [10], and the same logic applies to any domain family you want to enumerate. ## Common Mistakes and Limitations - **Searching a nucleotide sequence against protein domain profiles.** CDD models describe amino acid sequences. Translate your coding sequence in the correct frame first, using a tool such as the ORF finder on the NCBI site or `transeq` from EMBOSS, and search the protein. A nucleotide query returns nothing useful or misleadingly weak hits. - **Trusting hits below your significance threshold.** A hit with an E-value near 1 or above has a substantial chance of being a coincidence. Filtering on the threshold is not pedantry. Borderline hits in long, compositionally biased proteins are a common source of false annotation in manuscripts. - **Treating a superfamily hit as a specific function.** "Member of a large enzyme superfamily" is not the same as "catalyzes reaction X." Only the specific, deepest hit with retained functional sites supports a precise functional claim [2]. - **Assuming the best hit is the only hit.** Multi-domain proteins produce many hits. Read the full footprint map before summarizing a protein as a single domain type. - **Confusing redundancy with conflict.** Overlapping models from Pfam, SMART, and curated CDD that describe the same region are expected. Report one representative model per region and cite its source database. - **Letting local databases go stale.** A standalone RPS-BLAST run is only as current as the profile database you downloaded. CDD releases update on their own schedule [1], and an old local database can disagree with the live web service. If you see discrepancies between your local run and the web tool, check your database version first. - **Batch job hygiene.** Forgetting the request identifier, submitting an accession list with typos, or mixing FASTA and accession formats in one file are the three most common reasons a batch run fails or returns fewer results than expected. - **Overstating what a footprint proves.** A domain hit is a computational inference from sequence similarity. It supports hypotheses about function. It does not replace expression data, biochemical assays, or structural confirmation. Studies that treat domain findings as the start of a pipeline, rather than the conclusion, model the correct standard of use. - **Ignoring the N- and C-terminal tails.** Domains cover the conserved core of a protein. Long low-complexity or disordered tails outside domain footprints can still carry motifs and modification sites, so a domain search is not a complete functional annotation of a protein. - **Expecting the tool to catch everything.** Rapidly evolving or small domains can fall below detection. Cross-checking with an additional domain resource reduces this risk. ## Frequently Asked Questions ### What does the NCBI protein database conserved domains tool actually do? It locates conserved protein domain footprints within a query sequence and reports the functional sites and motifs inferred from those footprints, using a database of more than 57,000 domain models [1]. ### Do I need to install anything to run a conserved domain search? No. A single sequence or a batch of sequences can be analyzed entirely in the browser with CD-Search and Batch CD-Search [1]. Local installation is only needed for pipeline work. ### What E-value cutoff should I use for domain hits? Start with the tool's default threshold and inspect borderline hits by eye in the alignment view. Tighten the cutoff for large noisy batches rather than loosening it for small queries. ### What is the difference between a specific hit and a superfamily hit? A specific hit is the deepest matching model, often a curated subfamily with functional site annotations, while a superfamily hit places the region in the broader evolutionary group [2]. ### Can I search many proteins at once? Yes. Batch CD-Search accepts a multi-FASTA file or a list of accessions and returns tabular results for every query in one run [1]. ### How current are the domain annotations? The live web service reflects the current CDD release, and the local command-line route reflects whatever database version you last downloaded, so keep local copies updated [1]. ### Why did my protein return no domain hits? The protein may genuinely lack recognizable domains, fall below the detection threshold, contain sequencing errors or frameshifts, or have been submitted as nucleotide sequence instead of protein. Check each of these before concluding there is nothing there. ### How should I cite or report domain findings in a paper? Report the domain model name, its source database, the E-value, and the query coordinates, and state that annotation was performed with the CDD tools, following the reporting style used in published annotation workflows [1][7]. ## Sources 1. [NCBI's Conserved Domain Database and Tools for Protein Domain Analysis.](https://pubmed.ncbi.nlm.nih.gov/31851420/) 2. [CDD/SPARCLE: functional classification of proteins via subfamily domain architectures.](https://pubmed.ncbi.nlm.nih.gov/27899674/) 3. [Bioinformatics Approaches for Functional and Structural Annotation and Molecular Docking Study of a Hypothetical Protein From Staphylococcus aureus.](https://pubmed.ncbi.nlm.nih.gov/42252783/) 4. [Exploring Placental Protein-Target Protein Interactions: In Silico and In Vitro Approaches for Osteoarthritis Therapy.](https://pubmed.ncbi.nlm.nih.gov/40231497/) 5. [Bioinformatics analysis of the antigenic epitopes of L7/L12 protein in the B- and T-cells active against Brucella melitensis.](https://pubmed.ncbi.nlm.nih.gov/39687919/) 6. [Targeted immune epitope prediction to HHLA2 and MAGEB5 protein variants as therapeutic approach to related viral diseases.](https://pubmed.ncbi.nlm.nih.gov/34320928/) 7. [An in silico approach for screening non-synonymous SNPs in Mycobacterium tuberculosis PPE68 protein and impact on structure.](https://pubmed.ncbi.nlm.nih.gov/41350799/) 8. [CRESS domain-based molecular characterization of Begomovirus gossypimultanense.](https://pubmed.ncbi.nlm.nih.gov/42461357/) 9. [Proline-, Glutamic Acid-, Leucine-Rich Protein 1 (PELP1): Diversity, Structural Conservation, and Evolutionary Origins Across the Species.](https://pubmed.ncbi.nlm.nih.gov/41465416/) 10. [Evolution of sequence, structural and functional diversity of the ubiquitous DNA/RNA-binding Alba domain.](https://pubmed.ncbi.nlm.nih.gov/39638848/) 11. [Identification and expression analysis of the sucrose synthase gene family in pomegranate (Punica granatum L.).](https://pubmed.ncbi.nlm.nih.gov/35047243/) 12. [Structural Analysis of the AlkB Family in Poultry.](https://pubmed.ncbi.nlm.nih.gov/40646842/) 13. [Targeting conserved N-glycosylation blocks SARS-CoV-2 variant infection in vitro.](https://pubmed.ncbi.nlm.nih.gov/34839261/)