Raw protein sequence length is appended as an unnormalized feature

[jeffreyHoelzel]

The embedding pipeline appends the original sequence length as an additional scalar feature to every residue embedding similar to BepiPred-3.0.

Evidence:

• src/pepseqpred/core/embeddings/esm2.py
  • append_seq_len appends float(seq_len) as a column.
  • Both embedding generation and prediction embedding paths use this behavior.

Why this can hurt:

• Raw length can be orders of magnitude larger than normalized embedding features.
• It can become a source/pathogen shortcut if sequence length correlates with dataset or label source.
• Multi-pathogen training is particularly susceptible to spurious source-specific features.
• It worked with BP3 but may actually be hurting us here.

Planning direction:

• Run an ablation without the length feature.
• If length is retained, normalize it, bucket it, or pass it through a controlled transform.