Neural network variant calling with symmetric convolutional layers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current variant calling methods for genomes, especially tumoral DNA, face challenges due to the rapid mutation rate of tumor cells, leading to genetic heterogeneity and inaccuracies in identifying genomic variants.
Innovation Solution
A machine-learning based neural network is developed for variant calling, utilizing symmetric functions to process read data aligned to a reference genome, incorporating convolutional layers, reduction layers, and fully connected layers for accurate classification, and including features like insertion and deletion size descriptors, haplotype support, and Bayesian variant estimation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional variant calling methods are used, then the process is simple and fast, but accuracy deteriorates due to genetic heterogeneity in tumoral DNA
Solution Approach 1:
The neural network processes reads in segments through convolutional layers that apply filters to local regions of the input data. Each convolutional layer segments the read sequence into overlapping windows, allowing the network to process heterogeneous tumoral DNA by analyzing local patterns independently while maintaining global context through the pooling operations.
Solution Approach 2:
The patent replaces traditional mechanical/algorithmic variant calling methods with a neural network system that uses convolutional neural networks (CNNs) and recurrent neural networks (RNNs). This substitution enables the system to handle genetic heterogeneity in tumoral DNA through learned representations and transformations, achieving superior accuracy compared to conventional algorithms.
2Measurement precision
If machine learning methods are applied to improve accuracy, then variant calling precision improves, but computational complexity and training requirements increase
Solution Approach 1:
The neural network architecture is designed to be universally applicable to different types of variant calling scenarios (germline and somatic variants) through a single unified model that processes multiple data types. The convolutional and recurrent layers can handle various read depths and quality scores, reducing the need for separate training datasets for different variant types and lowering overall training data requirements.
Solution Approach 2:
The system dynamically adjusts processing parameters based on input characteristics, such as read depth and quality scores. The neural network adapts its feature extraction and classification behavior according to the specific characteristics of each read, allowing efficient processing across varying data conditions without requiring extensive retraining or multiple specialized models.
3Reliability
If redundant sequencing is used to counterbalance errors, then reliability improves, but data processing complexity increases
Solution Approach 1:
The neural network merges information from multiple reads at each genomic position into a unified variant calling decision. By pooling data from redundant sequencing reads through convolutional and pooling operations, the network integrates error-corrected signals while filtering out noise, achieving reliable variant detection without proportionally increasing processing complexity.
Solution Approach 2:
The neural network performs self-calibration and adaptive processing based on the quality scores and characteristics of incoming reads. The model automatically adjusts its processing behavior based on the redundancy and quality of the sequencing data, eliminating the need for manual tuning or complex post-processing steps while maintaining high reliability in error correction.
Data Source
AI summary
A computer-implemented method for machine-learning a neural network for variant calling with respect to a reference genome. The neural network takes as input one or more sets of data pieces each specifying a respective read aligned relative to a genomic position of the reference genome. The neural network outputs information with respect to presence of a variant at the genomic position. The neural network includes, for each set of data pieces, a respective function configured to take as input and process the set of data pieces. The respective function is symmetric. The machine-learning method improves variant calling with respect to a reference genome.