Convolutional Neural Network for Single Molecule Variant Calling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Accurate and efficient variant calling in single molecule sequencing is challenging due to high sequencing error rates, especially for indel errors, and existing approaches fail to handle these errors effectively, leading to suboptimal performance in identifying genomic variants underlying phenotypic differences and diseases.
Innovation Solution
A convolutional deep neural network is employed for variant calling, transforming genomic dataset properties into multi-dimensional tensors and passing them through a trained network to predict categorical output variables, including alternate bases, zygosity, variant types, and indel lengths, using a multi-task architecture with pooled convolutional and fully-connected layers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If single molecule sequencing is used, then sequencing throughput and read length are improved, but sequencing error rate increases
Solution Approach 1:
A convolutional neural network serves as an intermediary system between the sequencing data and variant identification. The neural network processes sequencing reads with high error rates and outputs accurate variant calls, effectively mediating the relationship between low-cost high-throughput sequencing and accurate variant detection. The network learns patterns from training data to distinguish true variants from sequencing errors.
Solution Approach 2:
The patent transforms sequencing data into multi-dimensional tensors with specific dimensional parameters (read depth, base quality scores, positional information, flanking sequence context). This parameter transformation enables the neural network to process and interpret sequencing data effectively, converting raw sequencing outputs into a representation format suitable for accurate variant calling despite high error rates.
2Measurement precision
If deep learning models are used for variant calling, then accuracy is improved, but computational complexity increases
Solution Approach 1:
The computational task is segmented into distinct functional components: data preprocessing (converting sequencing reads to tensors), feature extraction (convolutional layers identifying local patterns), and classification (fully-connected layers determining variant types). This segmentation allows each component to be optimized independently and facilitates parallel processing, reducing overall computational complexity while maintaining accuracy.
Solution Approach 2:
The patent introduces multi-dimensional tensor representation to capture diverse aspects of sequencing data simultaneously (read depth, quality scores, position, sequence context). By transforming one-dimensional sequence data into multi-dimensional tensors, the neural network can process information more efficiently and accurately, improving variant calling performance without proportionally increasing computational burden.
3Productivity
If traditional variant calling methods are used, then computational efficiency is maintained, but accuracy for indel errors deteriorates
Solution Approach 1:
The patent replaces traditional mechanical/algorithmic variant calling methods with a neural network-based system. Traditional methods use hand-crafted algorithms that struggle with indel errors, while the neural network learns complex patterns from training data, substituting simple algorithms with a sophisticated learning-based system that achieves superior accuracy for difficult indel cases while maintaining reasonable computational efficiency.
Data Source
AI summary
Systems and methods for variant calling in single molecule sequencing from a genomic dataset using a convolutional deep neural network. The method includes: transforming properties of each of the variants into a multi-dimensional tensor; passing the multi-dimensional tensors through a trained convolutional deep neural network to predict categorical output variables, the convolutional deep neural network minimizing a cost function iterated over each variant, the convolutional deep neural network trained using a training genomic dataset including previously identified variants, the convolutional neural network including: a plurality of pooled convolutional layers and at least two fully-connected layers connected sequentially after the last of the pooled convolutional layers, the at least two fully-connected layers comprising a second fully-connected layer connected sequentially after a first fully-connected layer; and outputting the predicted categorical output variables. In some cases, the categorical output variables include an alternate base, zygosity, variant type, and length of an indel.