Neoantigen Immunogenicity Prediction With Joint MHC Presentation Modeling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for predicting immunogenic neoantigens in cancer patients lack predictive accuracy, leading to ineffective personalized cancer vaccines due to low positive predictive value and time-consuming, laborious processes.
Innovation Solution
A novel method using a neural network model to jointly predict MHC class I or MHC class II binding affinity and the likelihood of tumor-specific neoantigen presentation on a cell-surface, incorporating peptide sequences and flanking regions, with training on both positive and negative data sets to enhance predictive accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If current in silico methods are used to predict MHC-binding affinity, then the prediction process is simplified, but the predictive accuracy and positive predictive value remain low
Solution Approach 1:
The patent combines multiple prediction tasks (MHC-binding affinity prediction and immunogenicity prediction) into a single integrated neural network model. This merging of functions allows the system to simultaneously optimize for both binding affinity and immunogenicity, thereby improving overall predictive accuracy while maintaining computational efficiency through a unified architecture rather than separate sequential models.
Solution Approach 2:
The neural network model is designed with multi-functionality to handle both MHC-binding affinity prediction and immunogenicity prediction within a single framework. The model accepts peptide-MHC pairs as input and outputs both binding affinity scores and immunogenicity predictions, making it a universal predictor that addresses multiple aspects of neoantigen characterization simultaneously, thus improving measurement precision without proportionally increasing complexity.
2Reliability
If existing predictors (MHCflurry-1.4, MHCflurry-2.0) are used, then the prediction process is established and reliable, but the positive predictive value for identifying immunogenic neoantigens is insufficient
Solution Approach 1:
The patent incorporates feedback mechanisms by training the neural network on experimentally validated datasets that include both binding affinity measurements and immunogenicity outcomes. The model learns from positive examples (immunogenic neoantigens) and negative examples (non-immunogenic peptides), continuously improving its predictive reliability and positive predictive value through this feedback-driven training process that adjusts weights to better distinguish truly immunogenic candidates.
Solution Approach 2:
The invention changes the predictive parameters by moving beyond simple MHC-binding affinity scores to include immunogenicity-specific parameters. The neural network outputs both binding affinity predictions and separate immunogenicity predictions, using transformed input features and specialized output layers that capture immunogenicity-related patterns. This parameter expansion allows the model to achieve higher positive predictive value by considering multiple dimensions of neoantigen quality simultaneously.
3Measurement precision
If comprehensive training data sets are used to improve model accuracy, then the predictive value increases, but the data processing time and computational resources increase
Solution Approach 1:
The patent applies preliminary action by pre-processing and curating training datasets before model training, organizing peptide-MHC binding data and immunogenicity data into structured formats with standardized features. The training data is pre-filtered and annotated with relevant labels, allowing the neural network to learn efficiently from high-quality prepared data. This preliminary preparation reduces the computational burden during actual prediction by ensuring the model receives optimized input data, thereby balancing predictive value improvement with reasonable processing time.
Data Source
AI summary
Disclosed herein are methods for predicting tumor-specific neoantigen MHC class I or MHC class II immunogenicity by jointly predicting MHC class I or MHC class II binding affinity and predicting the likelihood a tumor-specific neoantigen will be presented by a MHC class I or class II protein on a cell-surface.