Do Pseudosequences Matter in Neoantigen Prediction?

This repository contains the code, notebooks, training scripts, and figure-generation workflows used to evaluate multiple MHC-I allele representations for neoantigen prediction. This project examines how different MHC-I allele representations influence neoantigen prediction. We compare several pseudosequence definitions (BigMHC, NetMHCpan-4.1, and a random baseline), evaluate a range of pseudosequence lengths, test ESM-based embeddings, and introduce a graph-derived embedding built from HLA nomenclature, P-groups, and supertypes.

---

Repository Structure

nb/

Preprocessing

• Annotation-Based Embedding.ipynb — generates the annotation-based embeddings
• Pseudoseqs.ipynb — generates BigMHC EL, NetMHCpan-4.1, and random pseudosequences
• Gather_Testing_Results.ipynb — collects raw BigMHC outputs into a single file for downstream analysis

Visualization
Notebooks used to generate the plots for the figures in the paper:

• Panel_1.ipynb
• Panel_2.ipynb
• Panel_3.ipynb
• Plot_Length_Values.ipynb

src/

Training and evaluation code

• run_bigmhc.sh — script to retrain BigMHC as done in the study.
  Usage: sbatch run_bigmhc.sh pseudoseqs.csv mhclen_value
  Example: sbatch run_bigmhc.sh Original_BigMHC_Pseudosequences.csv 414

• mhcenc.py — replace the original mhcenc.py in the BigMHC source code when training with float-valued embeddings.

• esm.py — code to generate the ESM embeddings.

• compute_auprc.py — computes AUPRC to identify the best training epoch.

---

Main Experiments

1. Standard Pseudosequences

• BigMHC (30-res positions)
• NetMHCpan (34-res positions)
• Random pseudosequence baseline

2. Embedding-Based Representations

• Frozen ESM-2 embeddings (mean-pooled)
• Node2Vec embeddings using:
  • locus
  • two-field allele
  • supertypes
  • P-groups

3. Length Sweep

Systematic evaluation of pseudosequence windows from 5 to 100 residues.

---