Model has no local sequence modeling beyond ESM embeddings

[jeffreyHoelzel]

The FFNN flattens residues and predicts each residue independently after the ESM embedding. It does not model neighboring residue interactions during classification.

Evidence:

• src/pepseqpred/core/models/ffnn.py
  • PepSeqFFNN.forward flattens (B, L, D) into (B * L, D).
  • Output is reshaped back to (B, L).

Why this can hurt:

• Residue-level epitope signals are often regional.
• The ESM embedding contains context, but the classifier cannot enforce local smoothness or peptide-level consistency.
• Weak peptide-expanded labels may need a model/objective that understands windows rather than independent residues.

Planning direction:

• First fix diagnostics, objective, and weighting before changing architecture.
• Then consider light local context heads, such as 1D convolution, CRF-like smoothing, or pooling over peptide windows.