Skip to main content
Use tcr-workbench to run TCR-Workbench analyses of processed human or mouse alpha-beta TCR data on your own computer. The Skill picks the workflow that matches the quantity you ask for, runs the python3 tcr.py launcher, and explains what the outputs do and do not show.
This Skill calls no Biohub tools. Your AI client runs the TCR-Workbench launcher in a local checkout of TCR-Workbench, so it needs a shell on your own computer. Model workflows also need a completed setup, which the tcr-workbench-setup Skill covers.

​
Before you begin

​
Run the Skill

Score peptides for an HLA allele

​
What the Skill guides

  1. It checks the local setup with python3 tcr.py doctor and hands off to tcr-workbench-setup when the environment is missing or lacks what the analysis needs.
  2. It chooses the workflow from the requested quantity, such as pmhc-score for peptide scores in an MHC context, tcr-score for a paired receptor, repertoire-score for panel screening, or screen for curated reference evidence without model weights.
  3. It ingests Cell Ranger V(D)J, AIRR, or paired receptor tables, using the supplied V/J genes and anchored junctions without repairing sequences or inventing cell pairing.
  4. It keeps species, MHC class II alpha/beta pairs, model size, device, and precision explicit, and does not switch them silently.
  5. It writes each analysis to a new output folder and inspects the result tables, unresolved reasons, QC, and manifests, not only the HTML report.

​
Expected result

The result includes:
  • An output folder with an offline report.html, the complete result tables, and manifests that record the model, device, precision, and file hashes.
  • Scored and unresolved counts, with the reason each unresolved row could not be scored.
  • For score reports, the top-ranked peptides per receptor, MHC, and peptide length, compared with actually scored random peptides.
DecoderTCR scores are uncalibrated model outputs, not binding probabilities. Unresolved means insufficient support, not non-binding, and random reference peptides are not known nonbinders.

​
Skill files