HPL-APMS (pHLA)

Hierarchical Progressive Learning for Zero-Shot Peptide-HLA Binding Prediction and Automated Antigenic Peptide Design

Description

Predicting peptide binding to human leukocyte antigen (HLA) alleles is crucial for immune response initiation, essential for therapies like vaccines. While current AI-based computing methods have advanced, they're limited by training data covering less than 1% of known alleles, causing issues with generalization to unseen alleles.

To address this, we propose the Hierarchical Progressive Learning (HPL) framework. With the help of protein pre-trained language models, HPL learns sequence patterns from universal proteins to specific peptide-HLA complexes, improving prediction for unseen alleles by 60.8% (by the way, 1414.0% for non-classical alleles!) compared to TransPHLA model.

Additionally, we develop the Automated Peptide Mutation Search (APMS) program. APMS automatically modifies weak or non-binding peptides' amino acid residues for any target HLA class I allele using the HPL prediction model. It successfully generates high-affinity binder candidates for the target allele in over 38.1% of test cases while adhering to mutation restrictions.

Here is our paper: to be update. This Repo provides a detailed project instruction of our study.

Get Started

Installation

Check env.yml or set up a Python environment with the listed packages:

• python==3.8
• numpy pandas scikit-learn tqdm
• torch==1.7.0
• jupyter notebook seaborn matplotlib
• tape-proteins==0.5
  • TAPE model, a protein pre-trained language model used in our study, details see the source repo
• transformers==4.22.1
  • [Optional] You can install it if you want to use other pre-trained protein language models, like ProtBERT

Data download

We've uploaded all the data to Google Drive. You can download and store it wherever you prefer. Just remember to update the datapath or data_path variable in the code with the correct path!

View the source and purpose of each downloaded document by clicking here.

raw_data folder:

document	property/purpose	source
iedb_neg/	exported IEDB HLA immunopeptidome datasets	IEDB
hla_prot.fasta	HLA alleles and the corresponding amino acid sequences
Pos_E0101.fasta Pos_E0103.fasta Pos_G0101.fasta Pos_G0103.fasta Pos_G0104.fasta	experimentally validated binding peptides of five non-classical HLA alleles, i.e., HLA-E*01:01, HLA-E*01:03, HLA-G*01:01, HLA-G*01:03, HLA-G*01:04	HLAncPred Web server
new_hla_ABC_list.xlsx	list of HLA alleles mentioned in paper [a large peptidome...]
mhc_ligand_table_export_1677247855.csv	binding peptide-HLA pairs published by paper [a large peptidome...], exported from IEDB	IEDB

main_task folder:

document	property/purpose	source
train_data_fold4.csv val_data_fold4.csv	training dataset, common classical HLA alleles, consistent with TransPHLA	TransPHLA repo
independent.csv	testing dataset, common classical HLA alleles, consistent with TransPHLA	TransPHLA repo
HLA_sequence_dict_ABCEG.csv (old versions: hla_seq_dict.csv, HLA_sequence_dict_new.csv, HLA_sequence_dict_new.csv)	HLA name and corresponding full/clip/short(pseudo) sequence: common classical, zero-shot classical and zero-shot non-classical HLA alleles
IEDB_negative_segments.npy	negative peptides extracted from all possible peptide segments from the exported IEDB HLA immunopeptidome dataset	./Data_preprocess/build_candidate_pools.ipynb
allele2candidate_pools.npy	possible candidate peptide segments for each common classical HLA allele	./Data_preprocess/build_candidate_pools.ipynb
allele2positive_segs.npy	all possible peptide segments of positive peptides for each common classical HLA allele	./Data_preprocess/build_candidate_pools.ipynb
zeroshot_set.csv	zero-shot non-classical dataset	./Data_preprocess/prepare_EG_peptides.ipynb
zeroshot_allele2candidate_pools.npy	possible candidate peptide segments for each zero-shot non-classical HLA allele	./Data_preprocess/prepare_EG_peptides.ipynb
zeroshot_allele2positive_segs.npy	all possible peptide segments of positive peptides for each zero-shot non-classical HLA allele	./Data_preprocess/prepare_EG_peptides.ipynb
zeroshot_abc_set.csv	zero-shot classical dataset	./Data_preprocess/prepare_new_ABC_data.ipynb
zs_new_abc_allele2candidate_pools.npy	possible candidate peptide segments for each zero-shot classical HLA allele	./Data_preprocess/prepare_new_ABC_data.ipynb
zs_new_abc_allele2positive_segs.npy	all possible peptide segments of positive peptides for each zero-shot classical HLA allele	./Data_preprocess/prepare_new_ABC_data.ipynb
Supertype_HLA.xls	supertype category of HLA alleles	paper link

Hints:

• We denote the HLA alleles observed in model training as common classical HLA alleles since they are all commonly used in previous studies and classical (i.e., HLA-A/B/C).
• We call classical HLA alleles not seen in training as zero-shot classical HLA alleles, which are all from a recent study [a large peptidome...]
• Because non-classical HLA alleles (i.e., HLA-E/F/G) are not included in the model training, we refer to them as zero-shot non-classical HLA alleles.
• Build Negative peptide pool: see Data_preprocess/build_candidate_pools.ipynb

Usage

HPL

1. HPL-pan

• training：enter HPL/jobs/ directory and run finetune1.sh in the commandline

  • our official HPL-pan model can be downloaded at Google Drive
  • we train using four RTX3090 GPUs with a total batch size of 256.

• evaluating: enter HPL/jobs/ directory and run finetune1_eval.sh in the commandline

• inference: see HPL/inference_demo.ipynb

2. HPL-Cluster

HPL-Cluster is designed for target HLA allele. Firstly, you should check ./Data_preprocess/allele_cluster.ipynb notebook to obtain a certain amount of clustered HLA alleles.

• training：enter HPL/jobs/ directory and run finetune2.sh in the commandline
  • HPL-Cluster is based on HPL-pan, so you need to train HPL-pan model first or directly use trained HPL-pan that we provided. Just remember to update the load_path and model_name variable in fine_tune_tape2.py script with the correct path!
• evaluating: enter HPL/jobs/ directory and run finetune2_eval.sh in the commandline

3. HPL-Allele

HPL-Allele consists of a group of HPL-Cluster models for a specific target HLA allele. No training is required.

• evaluating: enter HPL/jobs/ directory and run ensemble_eval.sh in the commandline
  • remember to provide model names of HPL-Cluster models in Evaluation_HPL/evaluation_ft_ensemble.py script

APMS program

Refer to the APMS/mutation_release.py script for the complete algorithm implementation. Additionally, examples are available in the APMS/run_mutation_release.ipynb notebook.

Citation

If you found our code/work useful in your own research, please consider citing the following:

to be update

Contact Us

Free feel to create an issue under this repo or contact to be update [email] if you have any questions!

Acknowledgements

• tape-proteins for pre-trained model
• TransPHLA for baseline model and common classical dataset
• IEDB for data collection
• paper [A large peptidome dataset improves HLA class I epitope prediction across most of the human population] for zero-shot classical dataset
• HLAncPred Web server / paper [HLAncPred: a method for predicting promiscuous non-classical HLA binding sites] for zero-shot non-classical dataset
• paper [Classification of Human Leukocyte Antigen (HLA) Supertypes] for the idea of supertype categorization