Neoantigen Prediction - Usage Guide

Overview

Identify tumor neoantigens from somatic variants with pVACtools, with the methodological discipline the field actually requires: binding prediction is the easy part, and the analysis must center the downstream attrition (clonality, HLA loss-of-heterozygosity, expression, proximal-variant phasing, agretopicity/foreignness quality, and validation tiers), because a binding-only pipeline has single-digit-percent positive predictive value.

Prerequisites

# pVACtools (current 7.x; dependencies are heavy - use a dedicated env)
conda create -n pvactools python=3.11
conda activate pvactools
pip install pvactools vatools
pvacseq install_vep_plugin $VEP_PLUGINS   # installs Wildtype + Frameshift
# Ensembl VEP, a local IEDB install, and an HLA typer (OptiType/arcasHLA/HLA-HD) are also needed.

Quick Start

Tell your AI agent what you want to do: - "Find neoantigens from my VEP-annotated somatic VCF for this patient's HLA" - "Rank these neoantigens by tumor-specific quality, not just IC50" - "Drop candidates on HLA alleles the tumor lost (LOHHLA)" - "Incorporate proximal germline variants by phasing before prediction"

Example Prompts

End-to-End Calling

"Run VEP with the Wildtype and Frameshift plugins, annotate expression, then run pVACseq with my class I alleles"

"Call fusion-derived neoantigens from my Arriba output with pVACfuse"

Prioritization and Quality

"Compute agretopicity (DAI) correctly as the WT/MT binding ratio and flag anchor-position mutations"

"Rank candidates by cancer cell fraction and foreignness, and mark subclonal ones"

Tumor-Specific Pitfalls

"Which candidates are presented by an allele lost to HLA LOH?"

"These are frameshift neoantigens - how should I weight them given the MS-validation gap?"

What the Agent Will Do

Confirm the VCF is VEP-annotated with Wildtype + Frameshift plugins and a protein FASTA
Annotate expression and read counts (VAtools) so the filters are not silent pass-throughs
Phase proximal somatic/germline variants and supply the phased VCF
Run pVACseq with the patient HLA and sane VAF/coverage/expression filters
Run LOHHLA and drop candidates on lost alleles; estimate purity/CCF for clonality
Rank by quality (agretopicity, foreignness, expression, clonality) and re-tier in pVACview
Frame the output as a tier-1 hypothesis list for MS and functional validation

Tips

VEP plugins - Wildtype + Frameshift (NOT Downstream, which was dropped in pVACtools 2.0)
Binding is the easy part - spend effort on clonality, LOH, expression, and quality features
HLA LOH - run LOHHLA and remove candidates on deleted alleles; it fails silently if skipped
Clonality - use cancer cell fraction (purity + copy number), not raw VAF
Phasing - supply --phased-proximal-variants-vcf or the predicted peptides are ones the tumor never makes
Predicted != presented != immunogenic - the in-silico list is the input to validation, not the answer
Frameshifts/fusions - high-value, high-foreignness, but under-validated and rarely seen by MS

immunoinformatics/mhc-binding-prediction - the binding step (the solved, low-leverage part)
immunoinformatics/mhc-class-ii-prediction - class II neoantigens for CD4 help
immunoinformatics/immunogenicity-scoring - quality ranking of the candidate list
clinical-databases/hla-typing - the genotype substrate
variant-calling/variant-calling - upstream somatic calls
workflows/neoantigen-pipeline - end-to-end orchestration