TL;DR
This paper introduces Seq2Neo, a comprehensive, open-source pipeline for predicting cancer neoantigens and their immunogenicity. By integrating various data sources and employing a novel deep learning model, Seq2Neo aims to improve the identification of effective targets for cancer immunotherapy, particularly for patients who don't respond to existing treatments.
What Problem Does This Paper Address?
Cancer immunotherapies, like PD-1/PD-L1 blockers, have revolutionized cancer treatment, but a significant portion of patients do not respond. Combining these therapies with neoantigen vaccines shows promise, especially for non-responders. Neoantigens, derived from tumor-specific mutations, are ideal targets because they are unique to cancer cells. However, not all mutations lead to immunogenic neoantigens. Current tools for predicting neoantigen immunogenicity are often limited, focusing on specific mutation types or lacking comprehensive feature prediction. This gap hinders the development of effective neoantigen-based therapies. Seq2Neo aims to bridge this gap by providing a unified, robust solution for predicting neoantigens from raw sequencing data and assessing their immunogenicity.
Methods
Seq2Neo is a command-line-based pipeline designed for ease of use. It automates several key steps:
- Data Preprocessing: Handles raw sequencing data (FASTQ, SAM, BAM) for quality control, alignment, and variant calling.
- HLA Typing: Utilizes HLA-HD for determining patient-specific HLA types, crucial for antigen presentation.
- Mutation Calling: Employs Mutect2 for point mutations and INDELs, and STAR-Fusion for gene fusions.
- Neoantigen Candidate Generation: Annotates somatic variants to identify potential mutant peptides.
- Feature Prediction: Predicts several key features for candidate neoantigens:
- HLA Binding Affinity: Uses NetMHCpan BA, selected for its high accuracy and precision.
- TAP Transport Efficiency: Leverages NetCTLpan.
- Gene Expression: Quantified using TPMCalculator.
- Immunogenicity Score: This is the core novelty, predicted by a custom Convolutional Neural Network (CNN) model named Seq2Neo-CNN.
The Seq2Neo-CNN model is trained on a curated dataset from the IEDB and validated on an independent dataset. It incorporates peptide sequences, HLA types, TAP transport efficiency, and binding affinity to predict immunogenicity.
Datasets
- Training and Validation Data:
- IEDB (Immune Epitope Database): Used for training the Seq2Neo-CNN model. The dataset was filtered to include linear epitopes, T cell assays, MHC class I restriction, human origin, and specific assay types (ELISPOT, 51 Chromium, ICS, Multimer/Tetramer, ELISA). This resulted in 8975 peptides (3633 positive, 5342 negative).
- TESLA (Tumor Neoantigen Selection Alliance): An independent dataset of 599 experimentally tested tumor-specific neoantigens used for model validation.
- Benchmark Data:
- IEDB: Used to evaluate and select the best HLA-I binding affinity prediction algorithms.
- Cancer Patient Data: WES and RNA-seq data from five patients with experimentally validated neoantigens were used to test the Seq2Neo pipeline's performance in a real-world scenario.
Key Insights and Ideas
- Beyond Binding Affinity: The paper strongly emphasizes that HLA-peptide binding affinity alone is insufficient for predicting immunogenicity. Other factors like TAP transport efficiency play a significant role and are incorporated into Seq2Neo.
- Deep Learning for Immunogenicity: The development of a CNN-based model (Seq2Neo-CNN) specifically for predicting neoepitope immunogenicity is a key contribution, showing improved performance over traditional machine learning methods.
- Comprehensive Pipeline: Seq2Neo's strength lies in its "one-stop" nature, integrating raw data processing, mutation calling, HLA typing, and multiple feature predictions into a single, user-friendly pipeline.
- Handling Diverse Mutations: The pipeline's ability to handle various neoantigen types (point mutations, INDELs, gene fusions) makes it broadly applicable.
Key Results
- Seq2Neo-CNN Performance: The Seq2Neo-CNN model achieved the highest performance in predicting immunogenicity compared to other machine learning algorithms (ExtraTree, Random Forest, Logistic Regression, SVM, XGBoost) on both internal and independent datasets.
- Benchmarking HLA Predictors: NetMHCpan BA was identified as the best-performing algorithm for HLA-I binding affinity prediction, achieving an accuracy of 0.75 and precision of 0.96.
- Feature Importance: HLA-binding affinity and TAP transport efficiency were found to be significantly associated with immunogenicity, while proteasomal C-terminal cleavage was not. These two features were incorporated into the Seq2Neo-CNN model.
- Validation on Patient Data: Seq2Neo successfully identified 10 out of 16 validated neoantigenic sites in experimental cancer samples, with a lower rank percentage compared to other tools like pVACseq, TSNAD 2.0, and NeoPredPipe, indicating improved performance in identifying real immunogenic neoantigens.
Limitations and Future Work
- Data Imbalance: The paper notes that the training dataset had a significant imbalance between negative and positive reactive instances, which was addressed by weighting the classes during model training.
- HLA Allele Representation: While a pseudo-sequence approach was used for HLA allele representation, the paper doesn't explicitly discuss limitations of this method or potential improvements.
- Future Work: The authors anticipate Seq2Neo will be a valuable tool for neoantigen prediction and cancer immunotherapy. Further development could involve expanding the range of input data types, incorporating additional features, or refining the deep learning architecture for even greater accuracy. The open-source nature of Seq2Neo encourages community contributions and further advancements.
Summary updated