13/ In contrast, TCRbridge could not reliably distinguish whether non-validating (mislabeled) TCRs were reactive to YLQ or GLC.
This demonstrates how TCRvdb helps reveal that models like AlphaFold3 may perform better than previously assumed—if tested on validated data.

12/ We also show that TCRbridge can distinguish whether validating TCRs are reactive to YLQ or GLC.

11/ We introduce TCRbridge, a model that combines AlphaFold3’s per-residue confidence metrics at TCR–peptide interfaces.
TCRbridge revealed strong ability to distinguish validating from non-validating GLC-annotated TCRs, despite AlphaFold3 not being trained for this task.

10/ Can more complex models like AlphaFold3 distinguish validating from non-validating GLC-annotated TCRs? Interestingly, AlphaFold3 assigned lower confidence to non-validating TCR structures.

9/ tcrdist3 groups TCRs by sequence similarity. tcrdist3 distinguishes truly YLQ-reactive TCRs from non-validating ones, but its performance for GLC-reactive TCRs is modest, likely due to validating and non-validating GLC- annotated TCRs being similar in sequence space.

8/ We introduce TCRvdb, a functionally validated TCR-pMHC database to support TCR specificity model development, along with a standardized platform for rapid expansion. Using TCRvdb, we evaluated the performance of tcrdist3 and AlphaFold-based models.

7/ Why do some TCRs validate and others don’t? One key predictor: studies with more donors who had recent or ongoing COVID-19 showed a higher fraction of truly YLQ-reactive TCRs.

6/ Strikingly, this functional analysis demonstrates that claimed TCR reactivity is only confirmed for 50% of evaluated entries.

5/ To generate a high-confidence dataset for TCR specificity modeling, we validated TCRs annotated as reactive to YLQ (SARS-CoV-2) and GLC (EBV) epitopes using pooled functional genetic screening.

4/ Nanopore QC analysis of 5 TCR libraries showed 96.9% successful assembly, with an average of 74.2% sequence-perfect unique TCRs. Libraries were highly uniform—only 5.6-fold difference between the 5th and 95th percentiles.

3/ To create a Nanopore TCR QC pipeline with single base error detection capacity, we designed a UMI-based sequence error correction method to create highly accurate TCR UMI consensus sequences.

2/ Comprehensive QC of TCR libraries requires full-length TCR sequencing to resolve correct Vα–CDR3α–Jα and Vβ–CDR3–Jβ pairings. While Nanopore can sequence full-length TCRs, its raw base-call accuracy isn’t sufficient to detect mispairings, mutations, or indels.

1/ To evaluate a large fraction of TCR-pMHC entries in the database without bias from T cell phenotype or TCR abundance, we developed and utilized a high-throughput synthetic platform for TCR assembly.