🧬 Designing Better Cancer Vaccines — UCL-CCC Hackathon 2025 🏆

This repository contains our project developed for the UCL Cancer Collaborative Centre Hackathon 2025, where our team designed an end-to-end computational pipeline to improve personalised cancer vaccine development. Our aim is to identify high-value neoantigen targets that a patient's immune system has not already failed to recognise.

🏆 Result: Winner — UCL Cancer Collaborative Centre Hackathon 2025

---

🚀 Project Overview

Cancer vaccines provide a powerful therapeutic avenue, but high non-response rates remain a key challenge. The main bottleneck is the identification and prioritisation of effective neoantigen targets from an enormous search space (≈10¹⁷ peptides). Our solution integrates tumour genomics, patient HLA typing, and TCR repertoire data to generate a ranked list of optimal vaccine candidates.

Key Innovation

We exclude peptides already recognised by exhausted or ineffective T-cell responses, increasing the likelihood of inducing a strong and durable vaccine response.

---

🧠 High-Level Pipeline

1. Neoantigen Identification
   Extract mutated peptides (9-mers) present in cancer cells but not in normal tissue.

2. MHC Binding Prediction
   Predict HLA-specific binding affinities to determine surface-presentable peptides.

3. TCR Binding Prediction
   Identify and remove peptides already targeted unsuccessfully by the patient's TCRs.

4. Ranking & Output
   Score the remaining peptides using immunogenicity, presentation likelihood, clonality, conservation, and safety metrics.

---

🧬 Data Requirements

Training Data

• Cancer genomes with validated neoantigens
• HLA–peptide binding datasets (IEDB)
• TCR–peptide binding datasets

Patient-Specific Inputs

• Tumour genome sequencing
• Patient HLA genotype
• TCR repertoire sequencing

---

🧩 Model Architecture

Our architecture incorporates:

• Neoantigen identification module
• Transformer-based MHC binding prediction
• Structural TCR–peptide interaction modelling (AlphaFold-based)

---

📊 Ranking Metrics

• MHC binding affinity
• TCR engagement strength
• Surface presentation probability
• Peptide abundance
• Conservation across cancer subclones
• Phylogenetic clonality
• Cross-reactivity and safety assessment

---

📁 Repository Structure

.
├── Human_AF3_inputs/          # AlphaFold 3 JSON input files (6 TCR-pMHC jobs)
├── hack-1.ipynb               # Main pipeline notebook
├── human_tcr_dataset.xlsx     # Raw IEDB export
├── updated_reduced_data.xlsx  # Filtered IEDB data
├── new_reduced_data_TCRonly.csv            # Cleaned TCR-only dataset
├── Processed_Human_TCR_MHC_Dataset.csv    # Enriched dataset with MHC sequences
├── requirements.txt
└── README.md

---

🗄️ Datasets

Data Sources

Dataset	Source	Format	Public?
TCR–peptide–MHC binding data	IEDB	Excel export	✅ Yes
MHC heavy chain sequences (HLA allotypes)	UniProt REST API	FASTA	✅ Yes
Beta-2-microglobulin sequence (P61769)	UniProt REST API	FASTA	✅ Yes
AlphaFold 3 input files	Generated by pipeline	JSON	—

Dataset Details

human_tcr_dataset.xlsx / updated_reduced_data.xlsx

• Source: IEDB query export
• Columns: peptide, hla, tcr_alpha, tcr_beta
• ~70 raw rows filtered to 6 complete entries (both TCR chains required)

new_reduced_data_TCRonly.csv

• Cleaned/filtered version of the IEDB export
• 6 rows with complete TCR alpha + beta sequences

Processed_Human_TCR_MHC_Dataset.csv

• Enriched dataset combining IEDB data with UniProt-fetched sequences
• Columns: peptide, hla, tcr_alpha, tcr_beta, mhc_heavy_chain, beta_2_microglobulin
• 6 complete TCR-pMHC entries ready for structure prediction

Human_AF3_inputs/

• 6 AlphaFold Server JSON files (af3_job_0.json – af3_job_5.json)
• Each file encodes one TCR-pMHC complex with 5 protein chains: TCR-α, TCR-β, MHC heavy chain, β2-microglobulin, peptide (9-mer)

Peptides & HLA Allotypes in Current Dataset

Peptide	HLA Allotype
IMDQVPFSV	HLA-A*02:01
TRLALIAPK	HLA-B*27:05
LRVMMLAPF	HLA-B*27:05

HLA → UniProt ID Mapping (used for MHC sequence retrieval)

HLA Allotype	UniProt ID
HLA-A*02:01	P01892
HLA-A*02:05	P30512
HLA-B*27:05	P03989
HLA-B*27:09	P30480
HLA-B*08:01	P01889
HLA-E*01:03	P30511

---

🔮 Scaling This Further

To extend the pipeline beyond the current 6-entry proof of concept:

• More TCR-pMHC data: Relax IEDB query filters (e.g. allow single-chain TCRs, broader HLA coverage)
• Neoantigen prediction: Tools like pVACseq, NetMHCpan, or MHCflurry applied to somatic mutation data
• Tumour mutation data: TCGA (GDC portal) or ICGC — somatic mutation calls in MAF/VCF format
• HLA population frequencies: Allele Frequency Net Database to prioritise broadly immunogenic alleles
• Structural validation: PDB templates of TCR-pMHC complexes to benchmark AF3 outputs
• Immunogenicity scoring: NetTCR, ERGO, or ImRex for TCR-antigen binding prediction

---

🧑‍🔬 Team TC-AWARE

• Matthew Cowley
• Zhen Wei Yap
• Mohammad Alawwami
• Zarif Shafiei
• Linh Hoang
• Julia Sala-Bayo
• Graham Bonomo-Jackson
• Gleb Gmyzov
• Nick Keatley