Nanopore Basecalling With Dual LSTMs for Streaming Sequencing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current nanopore sequencing technologies face challenges in accurately and efficiently segmenting raw data into individual events representing base pairs and determining the corresponding bases, requiring significant computational resources and time, which limits their applicability and throughput.
Innovation Solution
A system utilizing a pair of unidirectional recurrent neural networks, specifically long short-term memory (LSTM) networks, processes raw nanopore sequencing data directly for basecalling, enabling streaming basecalling with improved accuracy and throughput by deferring segmentation until after sequence content analysis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional basecalling methods (HMM or RNN) are used to segment and process raw nanopore data, then basecalling can be performed, but computational resources and processing time are excessive
Solution Approach 1:
The patent applies preliminary action by pre-processing the raw nanopore signal to identify and remove artifacts (such as adapter sequences, hairpins, and other non-target signals) before the basecalling process begins. This preliminary cleaning of the data reduces the computational burden during basecalling while maintaining accuracy, as the neural network only needs to process clean, valid signals rather than filtering through noisy raw data.
Solution Approach 2:
The patent extracts and separates the basecalling function from the signal processing function. By using a dedicated neural network model that operates on pre-processed signals, the system extracts the core basecalling task from the complex pipeline of signal segmentation and event processing. This extraction allows for more efficient computation, as the neural network can focus solely on translating cleaned signals into base calls without the overhead of simultaneous segmentation and artifact removal.
2Productivity
If raw nanopore data is processed directly with neural networks, then processing speed improves, but data quality and signal accuracy deteriorate
Solution Approach 1:
The patent segments the basecalling process into distinct stages: first, signal processing and artifact removal; second, neural network basecalling on the cleaned signal. This segmentation allows each stage to optimize for its specific function—signal processing for accuracy and the neural network for speed—without compromising the other. The segmented approach ensures that high-speed basecalling is performed on high-quality, artifact-free data.
Solution Approach 2:
The patent introduces an intermediary step—signal cleaning and normalization—that sits between the raw nanopore signal and the neural network basecalling process. This intermediary acts as a mediator that transforms the raw signal into a format suitable for both high-speed processing and high-accuracy analysis. The intermediary processing ensures that the neural network receives clean, standardized input, enabling both speed and precision.
Data Source
AI summary
Disclosed herein are systems and methods for nanopore sequencing basecalling. In one embodiment, the method can include: receiving raw nanopore sequencing data comprising a plurality of continuous data acquisition (DAC) values corresponding to a biomolecule; normalizing the raw nanopore sequencing data to generate normalized nanopore sequencing data comprising a plurality of normalized DAC values; generating, using a first neural network (NN) and a normalized DAC value of the plurality of normalized DAC values, a vector of transformed probability values, segmenting the plurality of normalized DAC values into a plurality of discrete events; generating, using a second neural network and the event vector, an element determination of the biomolecule.