VDJDB: A curated database of T-cell receptor sequences of known antigen specificity

The primary goal of VDJdb is to facilitate access to existing information on T-cell receptor antigen specificities, i.e. the ability to recognize certain epitopes in certain MHC contexts.

Our mission is to both aggregate the scarce TCR specificity information available so far and to create a curated repository to store such data.

In addition to routine database updates providing the most up-to-date information, we make our best to ensure data consistency and fight irregularities in TCR specificity reporting with a complex database validation scheme:

• We take into account all available information on experimental setup used to identify antigen-specific TCR sequences and assign a single confidence score to highligh most reliable records at the database generation stage.
• Each database record is also automatically checked against a database of V/J segment germline sequences to ensure standardized and consistent reporting of V-J junctions and CDR3 sequences that define T-cell clones.

This repository hosts the submissions to database and scripts to check, fix and build the database itself.

To build database directly from submissions, go to src directory and run groovy -cp . BuildDatabase.groovy script (requires Groovy).

To query the database for your immune repertoire sample(s) use the vdjmatch software.

A web-based GUI for the database can be found in VDJdb-web repository.

Citing

Please cite the database using the most recent paper Mikhail Goncharov, Dmitry Bagaev, Dmitrii Shcherbinin, Ivan Zvyagin, Dmitry Bolotin, Paul G. Thomas, Anastasia A. Minervina, Mikhail V. Pogorelyy, Kristin Ladell, James E. McLaren, David A. Price, Thi H. O. Nguyen, Louise C. Rowntree, E. Bridie Clemens, Katherine Kedzierska, Garry Dolton, Cristina Rafael Rius, Andrew Sewell, Jerome Samir, Fabio Luciani, Ksenia V. Zornikova, Alexandra A. Khmelevskaya, Saveliy A. Sheetikov, Grigory A. Efimov, Dmitry Chudakov & Mikhail Shugay. VDJdb in the pandemic era: a compendium of T cell receptors specific for SARS-CoV-2. Nature Methods 2022. doi:10.1038/s41592-022-01578-0.

Submission guide

To submit previously published sequence follow the steps below:

• Create an issue(s) labeled as paper and named by the paper pubmed id, PMID:XXXXXXX. Note that if paper is a meta-study, you can mark it as meta-paper and link issues for its references in a reply to this issue. Also note that in case submitting unpublished sequences, choose any appropriate issue name with details on submitter (name, organization, etc) in issue comments.

• Create new branch and add chunk(s) for corresponding papers named as PMID_XXXXXXX. Don't forget to close/reference corresponding issues in the commit message.

• Create a pull request for the branch and check if it passes the CI build. If there are any issues, modify them by fixing/removing entries as necessary.

The structure of submission chunk is provided below, but first a couple of notes:

STYLE Try avoiding spaces (e.g. TRBV7,TRBV5, not TRBV7, TRBV5) and leave fields that have no information as blank (don't use any placeholder). Stick to listed field values at all cost! In case a critical part of your submission doesn't fit in current specification: 1) Create an issue in the issues section (and tag it as maintainance), 2) provide us with an example (e.g. open a pull request). Do not insert critical information into the comment field.

FORMAT Please ensure that Variable/Joining and MHC names in your submission come from IMGT nomenclature (this does not apply to donor MHC typing fields).

The BuildDatabase routine will be executed during CI tests upon each submission and prior to every database release implements table format checks, CDR3 sequence checks and fixes (if possible), and VDJdb confidence score assignment (see below).

To view the list of papers that were not yet processed follow here.

An XLS template is available here.

CAUTION make sure that nothing is messed up (x/X frequencies are transformed to dates, bad encoding, etc) when importing from XLS template. The format of all fields is pre-set to text to prevent this case.

Database specification

Each database submission in chunks/ folder should have the following header and columns:

Complex information columns (required)

These columns convey full information about TCR:peptide:MHC complex and are mandatory for any submission.

Notes:

In case given record represents a clonotype with either TCR alpha or beta sequence unknown, missing CDR3/V/(D)/J fields should be left blank.

V/(D)/J fields can be left blank, however this will abrogate CDR3 fixing/verification procedure for a given record.

Any record should have at least one of CDR3 alpha/beta fields that are not blank.

Method information columns (optional)

Optional columns (i.e. it is not required to fill them, but they should be present in table header) that ensure correct confidence ranking of a given entry. Used to calculate a single confidence score based on various factors, e.g. fraction of a given TCRab sequence among tetramer+ clones sequenced and verification experiments performed.

Notes:

In case method.identification is left blank, the record is automatically assigned with a lowest confidence score possible.

For special cases such as CD8-null tetramers that utilize HLA with mutated residues that abrogate CD8 binding, specify cd8null-tetramer in method.identification field rather than using mhc.a field.

During database build phase, the information from columns mentioned above is collapsed to a JSON string and stored in a single method column, e.g.:

{
   "identification":"tetramer-sort",
   "frequency":"5/13",
   "sequencing":"sanger",
   "verification":"antigen-loaded-targets"
}

Meta-information columns (optional)

Note:

While these columns are optional, subject identifier, replica identifier, etc are used when scanning submission for duplicates. Normally duplicate records (with identical complex information columns) are not allowed, but they will not be considered as duplicates in case they have distinct id fields mentioned above.

During database build phase, the information from columns mentioned above is collapsed to a JSON string and stored in a single meta column, e.g.:

{
   "cell.subset":"CD8+",
   "subject.cohort":"HSV-2+",
   "subject.id":12,
   "clone.id":46,
   "tissue":"PBMC"
}

Condition association columns (for extended database, TBA)

Condition metadata:

Association metadata:

Ambiguous antigens (for extended database, TBA)

Peptide pools, long peptides for T-cell culture expansion, non-peptide ligands

Non TRAB columns (for extended database, TBA)

Information for non alpha-beta T-cells, CAR-T, etc

Database processing

CDR3 sequence fixing

At this stage, a series of checks is performed for CDR3 sequence and reported V/J segments:

• In case of canonical (starting with conserved C and ending with F/W) CDR3 sequences: checks if 5' and 3' germline parts match corresponding V/J segment sequences.
• In case of truncated CDR3 sequences: adds conserved C/F/W residues. Can add more missing residues in case a relatively large contiguous V/J germline match is present.
• In case excessive germline part is reported (e.g. FGXG instead of simply F at CDR3 3' part), excessive residues are removed.
• Can correct mismatches in V/J germline regions in case a reliable non-contiguous V/J match is found.

The main reason behind that is that current immune repertoire sequencing (RepSeq) data processing software reports canonical clonotype sequences, high number antigen-specific TCR sequences present in literature are reported inconsistently. The latter greatly complicates annotation of RepSeq data using known antigen-specific TCR sequences.

In case of good V/J germline matching and errors in CDR3 sequence, the final CDR3 sequence in the database is replaced by its fixed version. The following report of CDR3 fixer is placed under cdr3fix.alpha and cdr3fix.beta columns, e.g.

{
    "fixNeeded":true,
    "good":false,
    "cdr3":"CASSQDVGTGGVFALYF",
    "cdr3_old":"CASSQDVGTGGVFALY",
    "jFixType":"FixAdd",
    "jId":"TRBJ1-6*01",
    "jCanonical":true,
    "jStart":14,
    "vFixType":"FailedBadSegment",
    "vId":null,
    "vCanonical":true,
    "vEnd":-1
    }

and

{
    "fixNeeded":true,
    "good":true,
    "cdr3":"CASSLSRGGNQPQYF",
    "cdr3_old":"CASSLSRGGNQPQY",
    "jFixType":"FixAdd",
    "jId":"TRBJ1-5*01",
    "jCanonical":true,
    "jStart":9,
    "vFixType":"NoFixNeeded",
    "vId":"TRBV14*01",
    "vCanonical":true,
    "vEnd":4
}

Field descriptions:

Note:

Possible V and J fix types: NoFixNeeded, FixAdd, FixReplace, FixTrim, FailedReplace (too many mismatches), FailedBadSegment (bad segment entry), FailedNoAlignment (no alignment at all)

VDJdb scoring

At the final stage of database processing, TCR:peptide:MHC complexes are assigned with confidence scores. Scores are computed according to reported method entries.

VDJdb scoring is performed by evaluating TCR sequence, identification and verification confidence based on the following criteria:

1. Ensuring TCR sequence is correctly identified according to method.sequencing and method.singlecell (1-3 points)
   • sanger - several cells sequenced (2+ cells sequenced according to method.frequency) - 2 points, otherwise 1
   • amplicon-seq - frequency is higher than 0.01 - 2 points, otherwise 0
   • single-cell - 3 points if performed
2. Initial identification of TCR:pMHC is correct according to method.identification (0-1 point)
   • sort-based - frequency is higher than 0.1 according to method.frequency)
   • culture-based - frequency is higher than 0.5
   • limiting dilution/culture prior to sequencing - the method.frequency becomes somewhat ambigous, check if it is higher than 0.5
3. Verification T-cell specificity (0-3 points)
   • direct method - 3 points, e.g. has PDB id (meta.structure.id is not empty) or some other method that directly evaluates TCR:pMHC binding
   • target stimulation-based - 2 points
   • staining-based - 1 points
   • If verification is performed, then the TCR sequence is assumed to be known, so score from 1. is set to 3

The final score is then calculated as minimal between score from part 1. and sum of scores from part 2. and part 3..

Maximal score is then selected among different records (independent submissions, replicas, etc) pointing to the same unique complex entry (i.e. set of unique complex fields).

Database build contents

The final database assembly can be found in the database/ folder upon execution of BuildDatabase.groovy script:

• vdjdb_full.txt - combined chunks with TCRalpha/beta records, antigen information, etc. All method and meta information are collapsed into two columns with corresponding names. VDJdb scores and CDR3 fixing information for TCR alpha and beta are given in separate columns. This is the raw version of VDJdb.
• vdjdb.txt - a collapsed version of database used for annotation of single-chain TCR sequencing data by VDJdb-standalone software. Each line corresponds to either TCR alpha or TCR beta record as specified by the gene column. TCR records coming from the same alpha-beta pair have the same index in complex.id column. In case complex.id is equal to 0 a record doesn't have either TCRalpha or TCRbeta chain information. This table is used by VDJdb-standalone and VDJdb-server.
• vdjdb.meta.txt - metadata for vdjdb.txt table, used by VDJdb-standalone and VDJdb-server.
• vdjdb.slim.txt - a slim database used for annotation of single-chain TCR sequencing data by VDJdb-standalone software. This is a collapsed version of vdjdb.txt containing unique records for each CDR3:antigen pair and comma-separated lists of values for other columns (*.segm,mhc.*, complex.id and reference.id). This table can be easily parsed with R and Python/Pandas, it is intended for end users exploring VDJdb.
• vdjdb.slim.meta.txt - metadata for vdjdb.slim.txt table.
• motif_pwms.txt and cluster_members.txt - position-weight matrices for antigen-specific TCR motifs and representative sets of TCR sequences that constitute them. These tables are computed separately using code from vdjdb-motifs repository.

Note that some statistics can be generated by running R markdown templates in summary/ folder.

Building database release and generating summary figures

First make sure that you clone both vdjdb-db repo and vdjdb-motifs repo to the same folder, say ~/vcs.

Then navigate to vdjdb-db and run bash release.sh. You can then find the output in ~/vcs/vdjdb-db/database, ~/vcs/vdjdb-db/summary and ~/vcs/vdjdb-motifs folders. Note that you have to check .Rmd files that will be executed and manually install missing R packages, as well as get VDJtools binary and place it in the path specified in ~/vcs/vdjdb-motifs/compute_vdjdb_motifs.Rmd.

Database build process with Docker

The repository contains Dockerfile to simplify the database building process. Dockerfile instantiates the correct environment needed to build the database. If you have Docker Desktop installed and running on your machine use the following command to build local Docker image:

docker build -t vdjdbdb .

NOTE You may need sudo to run docker.

In order to build the database using the newly created local Docker image create some folder (e.g. /tmp/output) and use it as a external volume when running Docker image. Docker image always puts the result in /root/output folder within docker container.

NOTE: Host path, e.g. /tmp/output, should be absolute.

NOTE: Database building process requires at least 64GB of RAM.

mkdir -p /tmp/output
docker run -v /tmp/output:/root/output vdjdbdb

Pre-built images can be found at DockerHub, N.B. replace vdjdb with mikessh/vdjdb:legacy if running this image.