To construct a comprehensive database of unconventional TCRs, we searched the published literature for unconventional TCR data on PubMed following the search terms: (CD8αα intraepithelial lymphocytes[Title/Abstract]) OR (mucosal-associated invariant T cells[Title/Abstract]) OR (MR1-restricted T cells[Title/Abstract]) OR (type I NKT T cells[Title/Abstract]) OR (MAIT[Title/Abstract]) OR (γδ T[Title/Abstract]) OR (iNKT[Title/Abstract]) OR (CD1b[Title/Abstract]) AND (T cell receptor). All results were then checked for inclusion of TCR sequences. The corresponding data listed in the literature tables or supplementary files were manually downloaded and archived. For those that do not provide processed TCR data, we downloaded the raw FASTQ files from Sequence Read Archive (SRA) (35) and extracted the TCR sequences using MiXCR (36). Besides, TCR sequences of γδ T cells (dataset ID: 20-36) were collected from the immuneAccess website by setting the locus to TRAD OR TRG in the data browser page (https://clients.adaptivebiotech.com/immuneaccess/browse). To ensure the reliability and quality of the data, all the TCR sequences were subject to quality control by a uniform workflow, and only sequences with high confidence were retained. The sequences which lack V/J gene annotation or contain any illegal amino acid abbreviated characters were discarded. For the sequences that contain multiple V/J gene annotations, the first ones were kept as the most confident annotated gene. Duplicated sequences of identical V gene, J gene, and CDR3 amino acid sequence in each dataset were removed. By re-organizing the data format, the corresponding T cell type information and PubMed ID were retained in addition to the V gene, J gene, and TCR CDR3 amino acid sequences.

UcTCRdb allows users to analyze selected data in the database or uploaded data files. V and J gene segment usage and gene-gene pairing landscapes are analyzed and shown using vertical stacks connected by curved paths whose thickness is proportional to the number of TCR clones with the respective gene pairing. The length distribution of the TCR CDR3 amino acid sequences was shown by the normalized frequency of each TCR chain. The normalized mutual information values (NMI) between each pair of CDR3α/β residues were calculated for different CDR3 lengths to represent the association of amino acid usage at different positions. We use Shannon entropy to measure the uncertainty of amino acid usage at different CDR3 positions. The entropy H(i) was calculated according to formula (1), where P(xi) represents the probability amino acid x appearing at position i, and |Xi| is the set of all different amino acids in position i. Then the mutual information (MI) of amino acid distribution for position i and j was calculated according to formula (2), and P(xi, yj) is the joint probability of x in position i and y in position j. The NMI was then calculated by using formula (3).

The NMI takes values in the range of 0-1, with a larger NMI indicating a higher amino acid correlation. We plotted the NMI matrix with a heatmap. The CDR3 sequence logos including amino acids as well as V and J genes were depicted by calculating the occurrence weights of different amino acids at each position. The larger the amino acid char in the sequence logo represents the higher frequency of amino acid occurrence. The amino acids were colored by their chemical properties in the sequence log plot. The scatter plot of amino acid usage was generated by calculating the proportion of each amino acid. To compare the differences between cell subsets, UcTCRdb has included 8 different datasets as references in the analysis module. To visualize the global similarity between TCR sequences, we calculated the Hamming distance between each pair of TCRs. We plotted a network graph where each node represents a TCR sequence and each edge represents a distance of less than 2 between two sequences. The visualization allows users to easily identify and explore groups of highly similar sequences.

UcTCRdb was implemented using Vue (version 3.2.13), an approachable, performant, and versatile framework for building web user interfaces and Flask (version 1.1.2), a lightweight Web Server Gateway Interface (WSGI) web application framework. All statistics and figures were calculated using Python (version 3.6.9) scripts and plotted using Echarts (version 5.3.2). The website was proxied on Nginx, an HTTP and reverse proxy server. And the server was running on the AliCloud server.

UcTCRdb consists of five main modules: Home, Browser, Analysis, Search, and Download ( Figure 1 ). Users can access the introduction, significant functions, and summarization of the database from the home page. The browser module allows users to interactively look through the characteristics of TCRs in different unconventional T cell subpopulations from different species. To facilitate the visualization and comparison of the TCR characteristics between different datasets, UcTCRdb provides a codeless analysis module. Users can easily analyze the repertoire data without any coding experience by selecting the datasets in UcTCRdb or uploading their data files, which will provide a great convenience for users with no programming background to quickly analyze and grasp the features of TCRs. As a comprehensive database, one of the core applications of UcTCRdb is to allow users to search for identical or similar TCR sequences by specifying species and cell types. In addition, the download page offers a user-defined manner that allows users to directly select and download the data.

Based on the data processing criteria, a total of 659,804 single-chain TCR sequences and 10,096 paired-chain TCR sequences from 34 corresponding studies in humans, mice, and cattle were collected and integrated into UcTCRdb ( Supplemental Table 1 ). Among them, over 85% of the data were derived from human samples ( Supplementary Figure 1A ) and are mostly γδ TCRs ( Supplementary Figures 1B, C ). This was expectable since we retrieved most datasets from studies on the human γδ T cells. Besides, MR1-restricted MAIT cells, CD1d-restricted iNKT cells, and CD1b-restricted germline-encoded mycolyllipid-reactive (GEM) T cells, which have invariant TCRα gene usage and limited TCRβ diversity, account for the majority of unconventional TCRαβ datasets ( Supplementary Figure 1D ). Moreover, two datasets of TCRαβ⁺CD8αα⁺ IELs were also included in the database.

The browser page focuses on the features of TCRs in the database, which allows users to look through the analysis results of the data from different views. To optimize the demonstration of the sequences data, unconventional T cells were categorized into two groups, unconventional αβ T cells, and γδ T cells. Here, we take human MAIT TCRs as an example to introduce the browser module ( Figure 2A ). Consistent with the previous studies (26, 38), we observed invariant TCRα gene usage and limited TCRβ diversity in the integrated MAIT TCR data. As expected, the dominant conserved TCRα gene is TRAV1-2 and TRAJ33 ( Figure 2B ) and the length distribution of TCRα CDR3 is concentrated at 12 amino acids ( Figure 2C ). In contrast, MAIT TCRβ shows more diversified gene usage and CDR3 length ( Figures 2B, C ). As shown in Figure 2B , the dominant TCRβ V genes are TRBV20-1, TRBV6-1,2,4, and TRBV4-1,2 suggesting that the usage bias of these genes might be related to antigen recognition by MAIT cells.

To further display the conservation of MAIT TCRβ sequences, we select the CDR3 sequences with a length of 15 amino acids, which represents the highest length frequency, for downstream analysis. By calculating the amino acid fraction and comparing it to the controlled conventional TCRβ CDR3s, we found differences in the distribution of amino acid usage between them, particularly the higher use frequency of glycine (G) and Glutamic acid (E) in MAIT TCRβ CDR3 sequences ( Figure 2D ). The heatmap of CDR3 NMI scores falls into three blocks, corresponding to the V gene motif region, the CDR3 highly variable region, and the J gene motif region ( Figures 2E, F ). In addition, multiple clusters were found in the similarity network graph of MAIT TCRβ sequences, which also indicated the conserved patterns of MAIT TCRβ ( Figure 2G ). The above examples showed that users can easily access TCR sequence features of different unconventional T cells across different species on the UcTCRdb browser page.

Further data mining of the unconventional TCR repertoire is critical for improving our understanding of the interaction between UcTCRs and their corresponding antigens. To simplify the analysis workflow and facilitate more users, UcTCRdb integrated several TCR analysis tools ( Figure 3 ). For an overall analysis of the UcTCRs, the length distribution of CDR3 and Sankey graph display of VJ gene usage were provided to access the intrinsic characters of the data. In addition, the NMI heatmap and sequence logo char of the CDR3 sequences were provided to depict the correlation and probability of amino acids in each position. The NMI takes values from 0-1, with a larger number indicating a higher amino acid correlation.

To identify the differences between UcTCRs and conventional TCRs, we compared the amino acid usage between unconventional TCRs and background TCRs [(100,000 sequences randomly sampled from the conventional TCR CDR3 sequences (37)]. The result can be visualized by a scatter diagram with the horizontal and vertical coordinates indicating the amino acid proportion in UcTCRs and references. To obtain the global similarity of TCR s, we calculated the Hamming distance between any pair of CDR3 sequences and constructed a distance matrix. The results were visualized via a network graph, where nodes represent TCR sequences, and edges were drawn if the distance between two nodes is less than 2. Nodes were further colored by different TRBV genes. Users can click on the node to trigger all adjacent nodes and get the corresponding sequence information. The above tools were all integrated into the analysis interface of UcTCRdb, allowing for personalized analysis on both selected datasets and uploaded data files ( Figure 3 ).

UcTCRdb allows users to search for TCR sequences in the database that is identical or highly similar to the query sequence. Users can set different query conditions, such as specific cell types, species, or whether contain user-defined mismatches ( Figure 4A ). Currently, UcTCRdb provides two options for calculating sequence similarity, i.e. Hamming distance and Levenshtein (edit) distance. The search result can be further filtered by a specific V or J gene. Due to different sequencing methods, the unconventional TCR sequence data that we obtained contain both single-chain and paired chains. We use an asynchronous loading technique to fetch the corresponding pairing chain data so that the results can be reformatted before being displayed. Users can click the display button to view the result and can download the search result by directly clicking the ‘Export To CSV’ button.

On the download page, UcTCRdb enables users to download datasets that are specified by species and cell types ( Figure 4B ). For the convenience of the users who are interested in processing the unconventional TCR sequence data, UcTCRdb also allows users to download all selected datasets by clicking the ‘Download’ button.