Multiple Instance Learning for Peptide-MHC Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for predicting peptide-MHC presentation, particularly for MHC class II molecules, face challenges due to limited training data and inefficiencies in using available data, leading to suboptimal performance in predicting peptide binding and presentation.
Innovation Solution
A computer-implemented method using multiple instance learning (MIL) with a loss function that calibrates model confidence through a probability calibration function, weighting positively labeled instances by their predicted probabilities, and employing a neural network model like BERT for accurate peptide-MHC presentation prediction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional supervised learning methods are used for peptide-MHC presentation prediction, then the model can be trained with available data, but the prediction performance is suboptimal due to limited training data and inefficient data utilization
Solution Approach 1:
The patent inverts the traditional learning approach by treating the set of MHC molecules as instances and the peptide presentation outcome as the bag label. Instead of predicting whether a single MHC-peptide pair binds, the model learns to predict whether any MHC molecule in a set can present a peptide, thereby inverting the supervision hierarchy to better utilize limited positive examples.
Solution Approach 2:
The patent applies partial supervision by providing labels only at the bag level (peptide presentation outcome) rather than requiring labels for every MHC molecule instance. This partial labeling approach allows the model to learn effective representations without needing complete annotation of all instance-level interactions, making efficient use of limited experimental data.
2Quantity of substance
If mass spectrometry experiments are used to generate training data, then peptide presentation data can be obtained, but the specific MHC molecule responsible for presenting each peptide cannot be precisely identified
Solution Approach 1:
The patent introduces an intermediary modeling approach where instead of directly identifying which specific MHC molecule presents each peptide (losing information), the model uses the set of all MHC molecules as an intermediary representation. The MIL framework aggregates signals from all instances in the bag, allowing the model to learn from the collective presentation capability without requiring precise instance-level attribution.
Solution Approach 2:
The patent transitions from a one-to-one mapping dimension (specific MHC-peptide pairs) to a one-to-many dimension (peptide presented by any MHC in the set). This dimensional shift allows the model to capture presentation patterns across multiple alleles simultaneously, transforming the loss of specific identification into a advantage for learning generalizable presentation rules.
3Reliability
If conventional learning methods are used, then the model training process is simple, but the model confidence calibration is poor leading to unreliable predictions
Solution Approach 1:
The patent incorporates feedback mechanisms through the probability calibration function that adjusts instance weights based on predicted probabilities. The loss function receives feedback from the model's own predictions and dynamically adjusts the weighting of training instances, creating a self-correcting training process that improves confidence calibration while managing complexity through iterative refinement.
Data Source
Figure 1
Figure 2
Figure 3~4
AI summary
Embodiments of the disclosed invention provide a computer-implemented method for predicting binding and presentation of peptides by MHC molecules. The method comprises collecting or generating training data, wherein the training data includes a set of MHC molecules present in a biological sample as well as a set of observed peptide sequences that are presented by at least one of the MHC molecules present in the biological sample, wherein it is not known to which specific of the MHC molecules a peptide sequence is bound, and wherein the training data are organized in bags with each bag having a set of training instances, wherein labels are known for the bags, but unknown for the training instances; using a loss function to train an MIL classifier fθ at an instance-level; and predicting the label of new instances by applying the MIL classifier fθ directly and/or predicting the label of new bags by applying the MIL classifier fθ to each instance of a respective bag and aggregating the results among all instances of the respective bag.