Build Neoantigen Prediction Tool for Brain Therapeutics

JOB: Computational Biologist / ML Engineer - Brain-Specific Neoantigen Prediction Tool

PROJECT OVERVIEW:
I am building a brain-specific neoantigen prediction tool for mRNA/LNP/PNA therapeutics. The goal is to predict which mutated peptides (neoantigens) are most likely to be presented by HLA in brain metastasis and trigger an immune response. This is a 7-stage pipeline with a defined step list.

WHAT YOU WILL BUILD:

Stage 1 - Data Sourcing
- Search and download brain metastasis MS immunopeptidomics data from CPTAC, SysteMHC, GEO, PRIDE
- Download TCGA primary tumour WES/RNA-seq (20 samples)
- Download BrainMetShare brain metastasis WES/RNA-seq (20 samples)
- Download GTEx normal brain expression, AFND HLA frequencies, IEDB self-antigens
- Record sample metadata (source, cancer type, ancestry, treatment history)

Stage 2 - Data Preparation and Labelling
- Label positive samples (MS-confirmed peptides)
- Generate negative samples (matched length + AA composition)
- Remove overlap between positive and negative sets
- Split 70/15/15 train/validation/test with batch-aware splitting
- Handle missing values, normalise features

Stage 3 - Feature Engineering (23 Features)
- NetMHCpan %Rank and IC50
- MHCflurry %Rank and IC50
- Ensemble binding score
- Proteasomal cleavage (NetChop), TAP transport score
- Cancer Cell Fraction (PyClone-VI), VAF, gene expression, mutant allele expression
- Normal brain expression (GTEx), CNS GO flag, microglia signature
- Human proteome homology (BLAST), self-antigen flag (IEDB)
- HLA allele frequency (AFND), HLA category
- Peptide length, hydrophobicity, net charge, stability
- Assemble 23-feature matrix

Stage 4 - Model Development
- Install Python 3.10+ (scikit-learn, xgboost, torch, pandas, numpy)
- Train XGBoost (n_estimators=300, max_depth=6, learning_rate=0.05) with 5-fold CV
- Train Random Forest (n_estimators=500, max_features='sqrt') with 5-fold CV
- Train Neural Network (23 -> 64 -> 32 -> 1, dropout 0.3, Adam, BCE loss)
- Apply early stopping on validation loss
- Train logistic regression meta-learner on base learner predictions
- Define Composite Immunogenicity Score (CIS)

Stage 5 - Validation and Benchmarking
- Evaluate on held-out test set: AUC-ROC, AUC-PR, Precision@20, Recall@20, F1
- Compare to NetMHCpan, MHCflurry, pVACtools
- Generate feature importance plots (XGBoost, Random Forest)
- Generate SHAP values, identify top 10 features

Stage 6 - Equity Pre-Screen (CPV/pPVI)
- Implement CPV = 1 - product(1 - f_i)
- Implement pPVI = sum(w_i * f_i) / sum(w_i)
- Calculate for 5 populations (EUR, AFR, EAS, SAS, AMR)
- Thresholds: PASS (CPV >= 60%), CAUTION (30-60%), FAIL (< 30%)
- Generate Equity Scorecard for test samples
- Implement mitigation: add promiscuous peptides binding multiple common HLA alleles
- Recalculate CPV after mitigation

Stage 7 - Reporting and Handover
- Document pipeline, methods, and results
- Provide reproducible code and environment file
- Handover to my team

REQUIREMENTS:
- Strong Python skills (pandas, numpy, scikit-learn, xgboost, pytorch)
- Experience with immunoinformatics tools (NetMHCpan, MHCflurry, NetChop, pVACtools)
- Experience with WES/RNA-seq data processing
- Understanding of HLA typing and neoantigen prediction
- Experience with ensemble ML models and SHAP
- Familiarity with population genetics (AFND, CPV/pPVI concepts)

NICE TO HAVE:
- Background in cancer immunology or immuno-oncology
- Experience with brain metastasis or CNS immunology
- Publications in neoantigen prediction or immunopeptidomics

WHAT I PROVIDE:
- Full 7-stage step list with detailed tasks
- Access to public datasets (all downloads are free)
- Clear communication and feedback
- Preprocessed data

TIMELINE: 8-12 weeks

BUDGET: 450 to 800usd.Open to discussion and negotiation

TO APPLY: Send me a message with:
1. Examples of similar ML/bioinformatics pipelines you have built
2. Your experience with immunoinformatics tools
3. Your estimated timeline and budget