RNA-seq Falls Short for TCR Repertoire Recovery, Review Finds

A new review published in Genomics & Informatics delivers a sober verdict on one of immunogenomics’ most tempting shortcuts: mining standard RNA-seq datasets to recover T cell receptor sequences. Drawing together quantitative evidence from head-to-head benchmarking studies that sequenced both RNA-seq and dedicated TCR-seq libraries from the same biological samples, the authors conclude that conventional short-read RNA-seq typically recovers only 3 to 5 percent of the unique clonotypes identified by TCR-seq, and that no amount of deeper sequencing or smarter software can close that gap entirely.

The appeal of the shortcut is easy to understand. Public repositories such as the Gene Expression Omnibus and the Sequence Read Archive hold more than a million RNA-seq samples, and if T cell receptor information could be computationally extracted from them, researchers could retrospectively map the immune landscape of thousands of diseases, treatments and patient cohorts without the cost, time and fresh-sample requirements of dedicated TCR sequencing. The review argues that this expectation rests on a critical and often false assumption: that RNA-seq data contain sufficient, high-quality TCR transcripts to be reliably assembled.

The technical reasons are rooted in how each technology works. Dedicated TCR-seq enriches TCR transcripts by 1,000- to 10,000-fold using PCR primers or hybridization probes targeting TCR constant regions, and its amplicon design spans the complete V-CDR3-J region, preserving the contiguous linkage between the V segment, the hypervariable CDR3 loop and the J segment. That linkage matters because the CDR3 region, generated by somatic V(D)J recombination and diversified by random nucleotide deletion and template-independent addition of P- and N-nucleotides, acts as a unique barcode for each T cell clone. The theoretical sequence space is estimated at roughly 10^18 unique TCRβ chains.

Standard RNA-seq destroys exactly this linkage before sequencing begins. Library preparation fragments RNA into pieces of a few hundred base pairs, so the V region, the CDR3 sequence and the J region of the roughly 500-base-pair TCR transcript often end up on separate fragments. Even paired-end 150-base-pair reads cover only about 300 base pairs, insufficient to traverse the full junction. In addition, TCR transcripts are vanishingly rare: the receptor is one of more than twenty thousand genes expressed in a cell, and even in a T cell the TCR transcript constitutes far less than one tenth of one percent of the mRNA library. A known three-prime bias in standard protocols, which preferentially captures sequence near the poly-A tail, further under-represents the CDR3 region, which sits in the middle of the transcript.

The benchmarking numbers quantify the consequences. Beyond the headline 3 to 5 percent clonotype recovery reported by Dabran and colleagues, Barennes and colleagues found that RNA-seq-based approaches consistently underestimated clonotype diversity, with recovery strongly dependent on tissue type and TCR transcript abundance. In T cell-rich samples such as peripheral blood mononuclear cells, recovery improved substantially, whereas in solid tumors and tissues with low T cell infiltration it dropped below 1 percent. In matched PBMC samples, RNA-seq captured approximately 40 to 60 percent of the top 100 high-frequency clonotypes but fewer than 10 percent of low-frequency clones, and clonotype frequency estimates showed poor concordance with TCR-seq, with Spearman correlation coefficients of roughly 0.3 to 0.5, primarily due to three-prime bias and variable capture efficiency.

The biases extend across repertoire metrics. In a study of tumor-infiltrating lymphocytes, Li and colleagues showed that RNA-seq reliably detected clonotypes making up more than 0.1 percent of the repertoire but missed clones below 0.01 percent, a critical limitation given that most tumor-reactive T cell clones are rare. Diversity measures such as Shannon entropy and clonal evenness were systematically underestimated, with RNA-seq-based values 30 to 50 percent lower than TCR-seq-derived estimates. When expanded clonotypes were defined as those with at least two copies, RNA-seq achieved a sensitivity of about 50 percent and specificity of about 70 percent in PBMC samples, but sensitivity fell below 20 percent in solid tumors. Pairing information fared worst: in a direct comparison using 10x Genomics single-cell RNA-seq and single-cell TCR-seq on the same cells, paired alpha-beta chain recovery from scRNA-seq alone was below 20 percent, compared with more than 90 percent for scTCR-seq.

Computational tools cannot fully compensate. Established assemblers such as MiXCR and TRUST4, both optimized for TCR-enriched samples, improve recovery over naive alignment: in one tool benchmark, TRUST4 recovered roughly twofold more clonotypes than MiXCR from RNA-seq data, though both remained far below TCR-seq recovery. Machine learning-based error correction shows promise in reducing false-positive assemblies, but no current method can reconstruct clonotypes for which no reads span the V-CDR3-J junction. The review frames the problem bluntly: when the input consists of short, fragmented, biased reads with broken linkage, assemblers either discard reads or, worse, merge fragments from different transcripts into incorrect assemblies.

The authors also scrutinize the more speculative hope that artificial intelligence could infer TCR sequences from gene expression profiles alone. They identify a fundamental information asymmetry: the theoretical repertoire space of roughly 10^18 beta chains vastly exceeds the dimensionality of a transcriptome of about 20,000 genes with continuous expression values, and there is currently no evidence that gene expression profiles contain sufficient information to reconstruct TCR sequences de novo. The review therefore distinguishes two categories of AI application: error correction and assembly improvement, which is supported by existing evidence, and imputation of missing TCR sequences from transcriptomes, which remains highly speculative and lacks empirical validation. Generative models that design novel TCR sequences with desired specificity are described as a separate, promising field that is fundamentally different from inferring native receptors from transcriptomic data.

The review does identify genuine openings. Its scope is explicitly limited to conventional short-read platforms, which dominate public datasets; long-read technologies such as PacBio Iso-Seq and Oxford Nanopore can generate reads exceeding 10 kilobases, sufficient to span the entire V-CDR3-J region in a single read, and have demonstrated full-length TCR alpha and beta reconstruction and paired chain analysis when combined with single-cell partitioning. However, long-read approaches carry higher per-base error rates, particularly for nanopore sequencing, lower throughput and higher cost, and cannot be applied retroactively to existing short-read archives. On the experimental side, probe-based hybridization capture can enrich TCR transcripts 100- to 1,000-fold from complex libraries, including from formalin-fixed paraffin-embedded tissue; FUME-TCRseq couples 5′ RACE with a single highly efficient multiplex PCR reaction to recover full-length, paired alpha-beta TCRs from low-input samples such as fine-needle aspirates and small biopsies; and rhTCRseq uses barcoded soluble peptide-HLA multimers to physically pull down T cells by their antigen specificity and sequence the matched receptor, directly linking TCR sequence to cognate antigen.

The overall conclusion is that RNA-seq cannot replace TCR-seq, which remains the gold standard whenever correct clonotype identification, preserved V-CDR3-J linkage or alpha-beta pairing information is required, whether for neoantigen-reactive T cell discovery or TCR-engineered therapy design. Yet the verdict is not purely negative. In T cell-rich tissues, and when high sequencing depth is combined with optimized computational pipelines, RNA-seq can capture the majority of high-frequency clonotypes, even if frequency estimation accuracy and alpha-beta pairing recovery remain poor. The authors argue that AI-enhanced tools focused on error correction, combined with targeted library improvements such as enrichment and long-read strategies, could expand the utility of existing datasets and degraded specimens, augmenting rather than supplanting the gold standard in a multi-omic approach to T cell repertoire analysis.

Subject of Research: Recovering T cell receptor CDR3 sequences from RNA-seq data and the limits of computational and experimental strategies

Article Title: Recovering CDR3 sequences from RNA-seq data: a critical assessment of computational and experimental strategies

Article References: AmeliMojarad, M., & AmeliMojarad, M. (2026). Recovering CDR3 sequences from RNA-seq data: a critical assessment of computational and experimental strategies. Genomics & Informatics, 24(1), Article 22. https://doi.org/10.1186/s44342-026-00084-5

Image Credits: AI Generated

DOI: 10.1186/s44342-026-00084-5

Keywords: TCR sequencing, CDR3, RNA-seq, immun repertoire, MiXCR, TRUST4, single-cell RNA-seq, long-read sequencing, AI error correction, probe-based enrichment, FUME-TCRseq, rhTCRseq