Peptide-MHC Immunogenicity Prediction Using Two-Model Calibration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for predicting immunogenic responses based on major histocompatibility complex (MHC) data are limited by the availability and specificity of training data, leading to suboptimal performance in predicting peptide-MHC interactions and immunogenic responses.
Innovation Solution
A two-model approach is employed, where a machine learning (ML) model is trained with MHC-restricted data to predict peptide-MHC binding and cell-surface presentation, and a calibration model is configured with MHC-unrestricted data to transform these predictions into accurate immunogenic response probabilities, using an expectation-maximization algorithm to iteratively refine the model parameters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If MHC-restricted training data is used to train the ML model, then the model can learn specific peptide-MHC binding patterns, but the limited availability and specificity of this data leads to suboptimal prediction performance
Solution Approach 1:
The patent introduces MHC-unrestricted data as an intermediary resource to bridge the gap between limited MHC-restricted training data and the need for accurate predictions. This intermediary data type allows the model to learn general immunogenic response patterns that can be combined with specific peptide-MHC binding information, effectively expanding the available training signal without requiring more scarce MHC-restricted data.
Solution Approach 2:
The patent segments the prediction task into two distinct components handled by separate models: (1) an ML model that handles peptide-MHC binding prediction using MHC-restricted data, and (2) a calibration model that handles immunogenic response calibration using MHC-unrestricted data. This segmentation allows each model to be optimized for its specific data type and function, resolving the contradiction by distributing the learning burden across specialized components.
2Measurement precision
If a single ML model is used for both peptide-MHC binding prediction and immunogenic response prediction, then the system is simpler, but the prediction accuracy is suboptimal due to the different data requirements and characteristics
Solution Approach 1:
The patent divides the prediction system into two specialized models: an ML model for peptide-MHC binding affinity prediction and a calibration model for immunogenic response calibration. Each model is trained on appropriately matched data types (MHC-restricted for the ML model, MHC-unrestricted for the calibration model), allowing both to achieve optimal performance for their specific functions while maintaining a relatively simple overall architecture through clear separation of concerns.
Solution Approach 2:
The calibration model serves as an intermediary layer that transforms the raw binding affinity predictions from the ML model into calibrated immunogenic response probabilities. This intermediary component bridges the gap between binding affinity and actual immunogenicity by incorporating population-level immunogenic response patterns learned from MHC-unrestricted data, thereby improving overall prediction accuracy without requiring the primary ML model to directly handle both tasks.
Data Source
AI summary
Immunogenic response prediction techniques are described. In an example, a system receives first training data that identifies a peptide, a first set of multiple major histocompatibility complex (MHC) molecules, and an immunogenic response associated with the peptide and the first set. The system updates a parameter of a first model based at least in part on the first training data. The first model is configured to determine a probability of causing the immunogenic response by a pair formed by the peptide and a MHC molecule from the first set. The system also receives second data that identifies a second set of MHC molecules, and generates, by using the first model and a second model, an immunologic response prediction of pairing the peptide with a MHC molecule from the second set. The second model is configured to generate at least one of peptide-MHC binding predictions or peptide-MHC cell surface presentation predictions.