Self-Training Base Caller for Embedded Sequencing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deploying deep Convolution Neural Networks (CNNs) on portable and embedded systems is challenging due to large data volume, intensive computation, varying algorithm structures, and frequent memory accesses, which affects the efficiency of Graphics Processing Unit (GPU) and other general-purpose platforms, necessitating specialized acceleration hardware like Field Programmable Gate Arrays (FPGAs) for efficient convolution operations.
Innovation Solution
A neural network-based base caller system is developed using FPGAs, which includes a configurable processor configured with a multi-cycle neural network architecture that segregates processing of data for different sequencing cycles, utilizing spatial and temporal convolution layers to handle misalignment in image data and provide real-time base calling, and iteratively trains itself using labeled training data to improve accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep CNNs are deployed on portable and embedded systems, then base calling accuracy is improved, but device complexity and computational requirements increase
Solution Approach 1:
The patent segments the base calling process into multiple cycles, where each cycle processes a portion of the sequencing data. The neural network is trained to handle multiple cycles iteratively, breaking down the complex task into manageable segments that can be processed sequentially on resource-constrained devices.
Solution Approach 2:
The patent implements dynamic base calling where the system adapts its processing based on the number of cycles completed. The neural network dynamically adjusts its predictions by incorporating information from multiple cycles, allowing the system to optimize between accuracy and computational resource usage in real-time.
2Measurement precision
If multiple sensing cycles are used for base calling, then base calling accuracy is improved, but processing time and data volume increase
Solution Approach 1:
The patent enables continuous base calling across multiple sensing cycles without requiring external training data generation between cycles. The neural network maintains continuous learning and prediction capabilities, processing data from each cycle immediately and iteratively improving base calling accuracy throughout the sequencing run.
Solution Approach 2:
The system performs self-training using its own generated predictions and comparisons against known oligo sequences. This self-service capability eliminates the need for external training infrastructure, allowing the system to improve its own accuracy continuously during operation without additional time overhead for separate training phases.
3Measurement precision
If external training data generation systems are used, then model accuracy is improved, but system complexity and external dependencies increase
Solution Approach 1:
The neural network performs self-training by comparing its base calling predictions against known oligo sequences that are already part of the sequencing data. This self-service approach allows the model to improve its accuracy without requiring external training data generation systems, reducing system complexity and eliminating external dependencies.
Solution Approach 2:
The system uses the same sequencing run data for both base calling and training purposes. The known oligo sequences serve dual functions as both the target to be sequenced and as training labels, making the system universally applicable without requiring separate training infrastructure.
Data Source
AI summary
A method of progressively training a base caller is disclosed. The method includes iteratively initially training a base caller with analyte comprising a single-oligo base sequence, and generating labelled training data using the initially trained base caller. At operations (i), the base caller is further trained with analyte comprising multi-oligo base sequences, and labelled training data is generated using the further trained base caller. Operations (i) are iteratively repeated to further train the base caller. In an example, during at least one iteration, a complexity of neural network configuration loaded within the base caller is increased. In an example, labelled training data generated during an iteration is used to train the base caller during an immediate subsequent iteration.