TL;DR
Researchers have created the Observed T cell receptor Space (OTS) database, a comprehensive resource of paired alpha-beta T cell receptor (TCR) sequences. This database enables new analyses of TCR pairing biases, public TCRs, and chain coherence, and has been used to train a novel paired-chain TCR language model for better understanding TCR function.
What Problem Does This Paper Address?
T cells are crucial for adaptive immunity, recognizing fragments of antigens presented by MHC molecules via their T cell receptors (TCRs). TCRs are typically composed of two chains, alpha (α) and beta (β), which work together to bind antigens. Historically, sequencing technologies often only captured one chain (usually β, due to its higher diversity), leading to an incomplete understanding of TCR recognition. While advances in single-cell sequencing now allow for the capture of paired αβ TCR sequences, there's been a lack of a centralized, curated database for this data. This gap hinders comprehensive analysis of TCR pairing, coherence, and the development of sophisticated computational tools like language models that leverage this paired information.
Methods
The core of this work is the creation and analysis of the Observed T cell receptor Space (OTS) database. The authors developed a processing pipeline, adapted from their previous antibody database (OAS), to curate publicly available paired TCR sequencing data. This pipeline involves:
- Data Acquisition: Sourcing raw 10x Genomics V(D)J sequencing data from studies involving humans.
- Processing: Using CellRanger to process fastq files, translating nucleotide sequences, and aligning them using IgBLAST.
- Filtering & Numbering: Applying IMGT numbering with ANARCI to ensure sequences are full-length and free from ambiguous amino acids.
- Pairing: Reconstructing paired αβ TCR sequences using 10x barcodes.
- Metadata Curation: Manually curating metadata from SRA tables and study manuscripts.
Beyond database creation, the authors employed several analytical methods:
- V-allele Pairing Analysis: Comparing observed V-gene pairings against expected frequencies derived from unpaired chain data to identify biases.
- Public TCR Identification: Mining the database for TCR clonotypes found in multiple independent studies.
- Chain Coherence Analysis: Quantifying the degree to which α and β chains are consistently paired within specific TCR populations (e.g., antigen-sorted vs. baseline).
- Paired-Chain TCR Language Model: Training a language model (TCRLang-Paired) on the curated paired TCR sequences to generate embeddings and perform residue-infilling tasks.
Datasets
The primary dataset is the Observed T cell receptor Space (OTS) database, which contains approximately 5.35 million redundant (1.63 million non-redundant) paired TCR sequences from 50 studies and at least 75 individuals. This dataset is predominantly composed of human TCR sequences.
For training the paired-chain TCR language model, the authors utilized:
- Unpaired TCR sequences: Approximately 9 million curated non-redundant sequences from iReceptor.
- Paired TCR sequences: Around 1.35 million non-redundant paired sequences from OTS.
The language model was evaluated on unseen MAIT TCR sequences from a study by Garner et al.
Key Insights and Ideas
- The Need for Paired Data: The study strongly emphasizes that understanding TCR function requires analyzing both α and β chains together, as their pairing and coherence reveal biological insights not apparent from single-chain analysis.
- OTS as a Central Resource: The creation of OTS addresses a critical gap, providing a standardized, accessible repository for paired TCR data, which is essential for advancing the field.
- Unexpected Pairing Biases: The analysis revealed specific TRAV:TRBV gene pairings that deviate significantly from expected frequencies, some of which are linked to known semi-invariant TCR populations (like MAITs and NKT cells) and responses to pathogens (like COVID-19), while others remain unexplained.
- Public TCRs as Signatures: Identifying public TCRs (found across multiple studies) highlights stereotypic immune responses, particularly to common viral epitopes, and also uncovers "orphan" TCRs with unknown specificities that warrant further investigation.
- Balanced Chain Coherence in TCRs: Unlike antibody light chains, both TCR α and β chains exhibit a more balanced coherence signal, suggesting that both chains contribute more equally to antigen specificity and orientation.
- Language Models Learn Paired Information: The development of TCRLang-Paired demonstrates that language models can effectively learn from paired TCR sequences, capturing inter-chain dependencies that improve representation and prediction tasks.
Key Results
- OTS Database Size: Contains over 5.3 million redundant (1.63 million non-redundant) paired TCR sequences from 50 studies.
- V-allele Pairing Deviations: Over 97% of V-allele pairings were within 0.020% of expected frequencies, but several significant deviations were identified, including those linked to MAIT cells, iNKT cells, and COVID-19 responses.
- Public TCRs: Identified 2276 public TCR sequences, with 45 found in five or more independent investigations.
- Coherence Analysis: Antigen-sorted TCRs showed significantly higher coherence (e.g., ~45.59% for α-chains) compared to baseline populations (~9.55%), indicating a stronger pairing preference when TCRs are selected for function.
- TCRLang-Paired Performance: The paired language model could distinguish TCRs based on V/J gene usage and group functionally distinct subsets (like MAIT cells) in embedding space. While residue-infilling performance was not significantly improved by paired context, likelihood analysis showed that the model learned inter-chain dependencies.
Limitations and Future Work
- Baseline Coherence Approximation: The "baseline" coherence dataset is an imperfect approximation of a true naïve TCR repertoire due to the scarcity of studies specifically sorting for naïve T cells.
- Data Biases: The database is still growing, and certain populations (e.g., CD8+ T cells) are more represented than others, potentially skewing some analyses.
- Uncharacterized TCRs: A significant portion of public TCRs identified in OTS remain unannotated in existing databases like VDJdb, highlighting a need for further functional characterization.
- Future Work: The authors plan to continuously update OTS with new studies, expanding its depth and diversity. They also suggest that future language models could be rationally designed based on the specific biological features they aim to capture, and that incorporating data from other sequencing methodologies will be important as they become more prevalent. The potential for using OTS data to direct further improvements in public paired sequencing data coverage is also noted.
Summary updated