Status: Draft, EDA only.

Objective

Classify immune receptor repertoires using V/J gene segment usage and CDR3 sequences.

A repertoire is basically a bag of receptor sequences. Each one has a V gene, a J gene, and the CDR3 region in between, which is where most of the variety lives. The question is whether that bag carries enough signal to say which class the person belongs to.

Approach

1. Batched sequence embeddings: embed CDR3 sequences in batches so a whole repertoire can be summarised without loading everything at once.
2. N-gram kernel (PyTorch): predict V/J gene usage from context windows over the sequence. If the context around a CDR3 predicts its V/J genes well, that structure might be what separates the classes.

Where it's at

• ✅ EDA on V/J gene usage and CDR3 sequences
• ✅ Batched embedding pass
• ❌ N-gram data processor. Self note: it's still a stub, so the kernel has nothing real to train on yet.
• ❌ Repertoire-level classifier