AI-designed neoantigen vaccines just hit a Phase 3 milestone —

Moderna and Merck's mRNA-4157 data dropped last week: 110% stock pop, but the real signal is that personalized neoantigen vaccines work at scale. The pipeline — tumor sequencing → mutation calling → HLA binding prediction → epitope ranking → mRNA construct — is where ML has been eating the workflow for years. Most of the heavy lifting (NetMHCpan, MHCflurry, pVACseq) is already open. What's missing is a clean, reproducible end-to-end stack you can run on your own GPU cluster without pharma-grade infra.

I've been stitching together a minimal viable version for a side project. Rough outline if anyone wants to replicate:

1. Somatic variant calling

# tumor/normal pair → filtered VCF
gatk Mutect2 \
  -R hg38.fa \
  -I tumor.bam \
  -I normal.bam \
  -tumor TUMOR \
  -normal NORMAL \
  -O somatic.vcf.gz

2. HLA typing from RNA-seq (or WES)

optitype --rna -i tumor_rna.fastq.gz -o hla_typing.tsv

3. Neoantigen prediction + ranking

# pVACseq wrapper — handles NetMHCpan binding, expression filter, clonal fraction
pvacseq run \
  somatic.vcf.gz \
  hla_typing.tsv \
  sample_id \
  output_dir \
  -e 8,9,10,11 \
  --allele-specific-binding-threshold 0.5 \
  --top-score-metric lowest

4. Construct design
Feed top 20-30 epitopes into your mRNA backbone (5' UTR, CDS optimization, poly-A tail). I'm using a custom codon optimizer + UTR selector trained on ribosome profiling data — happy to share the notebook if there's interest.

5. In silico immunogenicity sanity check
Run the final construct through MHCflurry 2.0 presentation score + TCRdist similarity to known immunogenic epitopes. Filters out ~40% of candidates that look good on binding alone.

The bottleneck isn't prediction accuracy anymore — NetMHCpan-4.1 hits AUC 0.92 on presented ligands. It's manufacturing turnaround. Moderna quotes 6-8 weeks from biopsy to vial. Academic cores run 12-16. If you're building in this space, the lever is automating the QC/release pipeline (residual DNA, dsRNA, integrity) not the ML.

Anyone running a similar stack? Curious what HLA typing method you trust — OptiType, HLA-LA, or something newer.