TL;DR
This paper introduces PoreOver, a novel computational method that leverages paired sequencing reads from Oxford Nanopore's 1D2 protocol to significantly improve basecalling accuracy. By aligning the probability profiles of two complementary DNA strands, PoreOver effectively halves the median sequencing error rate, making nanopore sequencing more reliable for applications requiring high precision.
What Problem Does This Paper Address?
Nanopore sequencing offers direct readout of DNA molecules but has historically been hampered by a higher error rate compared to other methods. This limitation restricts its use in scenarios demanding high accuracy, such as detecting rare genetic variants. While Oxford Nanopore's 1D2 protocol sequences both strands of a DNA duplex, most existing neural network basecallers are designed for single-strand analysis. This paper addresses the challenge of effectively utilizing the redundant information from these paired reads to boost basecalling accuracy.
Methods
The core of this work is the development of a "pair consensus decoding" algorithm. Instead of relying on a single read's basecalled sequence, PoreOver takes the probability profiles (output by neural network basecallers) from two complementary reads. It then employs a constrained dynamic programming heuristic, specifically a banded 2D beam search, to align these probability profiles. This alignment process effectively finds the consensus sequence that maximizes the combined probability from both strands. To accelerate this computationally intensive process, an "alignment envelope" is generated by first performing a preliminary Viterbi decoding on each read individually and then aligning these initial sequences. This envelope guides the beam search to focus on the most likely regions of alignment. The authors also developed a basic recurrent neural network basecaller, PoreOverNet, for demonstration purposes.
Datasets
The primary dataset used for evaluation consisted of 5,000 R9.4 E. coli 1D2 read pairs (10,000 reads total) provided by Oxford Nanopore Technologies. The paper also mentions testing their algorithm on reads processed by other basecallers like DeepNano-blitz and Bonito, implying these were also used for evaluation.
Key Insights and Ideas
- Leveraging Redundancy is Key: The paper demonstrates that simply having two reads of the same DNA molecule isn't enough; intelligently combining their probabilistic outputs is crucial for accuracy gains.
- Probability Profiles are Richer than Sequences: Instead of just using the final basecalled sequence, PoreOver works with the underlying probability distributions, which contain more nuanced information about basecalling confidence.
- Alignment of Probabilities, Not Just Sequences: The novel aspect is aligning the probability profiles of the two strands, rather than just aligning the already decoded sequences. This allows for a more robust consensus.
- Computational Efficiency Matters: The development of the "alignment envelope" is a critical insight for making this computationally demanding approach practical for real-world nanopore data.
Key Results
- When applied to the Bonito basecaller, PoreOver reduced the median sequencing error rate by more than half.
- For their own PoreOverNet basecaller, pair decoding improved median accuracy from 87.6% for single reads to 93.2% for 1D2 read pairs.
- Consensus decoding with PoreOver lifted Bonito's median accuracy from 94.7% to 98.1%, surpassing the consensus accuracy of ONT's Guppy basecaller.
- The PoreOver software is open-source and modular, designed to be compatible with various neural network basecallers.
Limitations and Future Work
- Scalability to More Than Two Reads: The current method is designed for pairs of reads. Generalizing to multiple reads (for tasks like genome polishing) presents a significant computational challenge due to the "curse of dimensionality" (O(T^N) complexity for N reads). The authors suggest that future work could involve generalizing alignment envelopes to multiple sequences or employing stochastic search methods.
- Potential for Further Speedups: While the banded alignment and envelope approach improve speed, the authors note that exploiting recent advances in efficient pairwise alignment algorithms could yield further improvements in both accuracy and speed.
- Integration into Existing Tools: The paper highlights that the Bonito basecaller has since incorporated an implementation of their pair decoding algorithm, indicating successful integration and adoption.
Summary updated