Source-linked AI summary
DeepNano: Deep Recurrent Neural Networks for Base Calling in MinION Nanopore Reads
Vladimír Boža, Broňa Brejová, Tomáš Vinař
TL;DR
MinION offers long reads and portable, real-time sequencing but suffers from a high sequencing error rate. The paper develops DeepNano, an open-source deep recurrent-neural-network base caller, and reports improved accuracy over Metrichor, especially for 1D reads.
Problem
MinION produces long reads but has a high sequencing error rate, creating a need for improved base calling.
Method
DeepNano uses deep recurrent neural networks to convert MinION event features into probability distributions over called bases.
Results
DeepNano is significantly more accurate than Metrichor for 1D reads on both strands and data sets, with slightly higher accuracy for 2D reads.
Takeaways & Limitations
DeepNano provides a more accurate and computationally efficient alternative to the HMM-based methods used in Metrichor.
Takeaways & Limitations
DeepNano still relies on Metrichor-generated event data and, for 2D base calling, an alignment between template and complement sequences from the original base caller.
Abstract
from arXiv · showhide
Motivation: The MinION device by Oxford Nanopore is the first portable sequencing device. MinION is able to produce very long reads (reads over 100~kBp were reported), however it suffers from high sequencing error rate. In this paper, we show that the error rate can be reduced by improving the base calling process. Results: We present the first open-source DNA base caller for the MinION sequencing platform by Oxford Nanopore. By employing carefully crafted recurrent neural networks, our tool improves the base calling accuracy compared to the default base caller supplied by the manufacturer. This advance may further enhance applicability of MinION for genome sequencing and various clinical applications. Availability: DeepNano can be downloaded at http://compbio.fmph.uniba.sk/deepnano/. Contact: boza@fmph.uniba.sk
1 Introduction
MinION enables portable, real-time sequencing and very long reads but has a high error rate. DeepNano introduces an open-source recurrent-neural-network base caller designed to improve accuracy beyond the manufacturer’s default approach.
- MinION is a 90-gram, low-cost, portable sequencer that can analyze data in real time for applications including infectious-disease monitoring.
- DeepNano is presented as the first open-source base caller for the MinION platform, addressing base-calling accuracy rather than consensus generation from multiple reads.
- Nanopore current measurements are segmented into events summarized by current mean, variance, and duration before translation into DNA sequence.
- Reads can reach 100,000 bp, while combining template and complement strands reduces default-base-caller error from roughly 30% for 1D reads to 13–15% for 2D reads.
- The conventional process can be represented by an HMM whose states encode overlapping k-tuples and whose emissions model expected current with variance.
- DeepNano uses recurrent neural networks, whose large hidden state space can potentially capture long-distance dependencies unlike fixed-k-mer HMMs.
2 Basecalling using deep recurrent neural networks
DeepNano represents MinION event sequences with deep recurrent neural networks. The model processes event-level features through recurrent hidden states to produce sequence-labeling outputs, using three or four layers.
- The network inputs the mean, standard deviation, and length of each event, while its output vectors provide the corresponding sequence predictions.
- For each event, a recurrent network updates its hidden state from the current input and previous hidden state, then derives an output vector.
- Using several hidden layers can improve prediction accuracy, and DeepNano employs networks with three or four layers.
- Different layers use separate functions f1, f2, and f3, each with its own parameter set.
2.2 Bidirectional recurrent neural networks
The base caller uses bidirectional recurrent neural networks so predictions can incorporate sequence information from both directions. Gated units address training difficulties associated with ordinary recurrent networks.
- Bidirectional processing: Bidirectional networks let each prediction use data occurring both before and after the current input.They scan in both directions and concatenate hidden outputs before passing them to the next layer.
- Recurrent layers: A typical recurrent layer applies a linear transformation followed by a hyperbolic tangent nonlinearity.The matrices W and U and bias vector b are model parameters, with separate parameters for each layer and direction.
- Training difficulty: Ordinary recurrent networks suffer from vanishing gradients, making training harder as backpropagation moves toward earlier layers.Gradients in layers farther from the output become much smaller than those closer to it.
- Gated units: Gated recurrent units use update and reset gates to improve gradient flow through the network.Both gates produce values in the interval (0, 1); the reset gate can reduce the impact of the previous state.
2.4 Output layer
The output layer accommodates mismatches between sequencing events and called bases by predicting up to two bases per event, including no-base outputs. Final calls select the most probable character for each output.
- Output representation: Each event produces two probability distributions over Σ = {A, C, G, T, −}, allowing up to two bases or no base per event.The dash denotes no base, and the latter case is limited to two bases per event.
- Output probabilities: The network uses softmax probabilities from the last hidden layer to score each possible base.
- Final calls: Final basecalling chooses the most probable base for each output, or no base when the dash is most probable.
2.5 Training
Training maximizes the log likelihood of correct output bases when event-level labels are known, but uses an EM-like alignment procedure when exact event-to-base pairs are unavailable. Optimization combines SGD with Nesterov momentum and, for 2D basecalling, later L-BFGS fine-tuning.
- Supervised training: When correct bases are known for each event, training maximizes the summed log likelihood of the correct output sequence.The target outputs are the correct bases associated with the input events.
- Optimization: SGD with Nesterov momentum is used to increase convergence rate during optimization.For 2D basecalling, optimization switches to L-BFGS after several iterations.
- Optimization: For 2D basecalling, SGD is used initially and L-BFGS during final fine-tuning.The authors report that SGD better avoids poor local optima early, whereas L-BFGS is faster during final fine-tuning.
- Alignment-based training: When exact event-to-base pairs are unknown, training alternates optimization with approximate realignment in an EM-like procedure.Initial alignments use differences between expected and observed event means plus penalties for undetected and split events.
2.6 1D basecalling
For 1D basecalling, DeepNano uses separate recurrent-neural-network models for template and complement strands because the strands have different properties. The networks process event data from Fast5 files after signal scaling and shifting.
- Strand-specific models: Separate neural-network models are used for template and complement strands because the strands have different properties.Both models contain three hidden layers with 100 hidden units.
- Input preprocessing: The networks receive event data from Metrichor-produced Fast5 files as input.Preprocessing uses scaling and shifting parameters stored in the same Fast5 files.
2.7 2D basecalling
The paper combines template and complement event information in a neural network to improve 2D basecalling. This design avoids the unsatisfactory accuracy of separately predicting each strand and then aligning their probability sequences.
- 2.7 2D basecalling: Separate strand predictions aligned by dynamic programming produced the maximum-likelihood DNA sequence but yielded the same or slightly worse accuracy than Metrichor.The authors attribute this to independent base probabilities that lack dependencies between adjacent basecalls.
- 2.7 2D basecalling: DeepNano instead inputs corresponding events from both strands and combines them into a single prediction.Aligned event pairs become one input vector, while strand-specific events are filled with special values.
2.8 Implementation details
DeepNano was implemented in Theano for Python, and the authors omitted regularization because their dataset showed almost no overfitting.
- 2.8 Implementation details: Theano for Python provided symbolic differentiation and other features used to train the neural networks.
- 2.8 Implementation details: The authors did not use regularization because they observed almost no overfitting with their dataset size.
3 Experimental results
Experiments on E. coli and K. pneumoniae testing data compared DeepNano with Metrichor using reference-aligned accuracy. DeepNano was significantly better for 1D reads, slightly more accurate for 2D reads, showed less GC-content bias on K. pneumoniae, and processed reads faster under the reported setup.
- 3. Experimental results: Table 2 accuracy was computed as alignment matches divided by alignment length after BWA-MEM reference alignment.
- 3. Experimental results: DeepNano was evaluated on existing E. coli and K. pneumoniae reads, using mapped reads with full 2D base calls and separate training and testing regions for E. coli.The E. coli training reads mapped to the first 2.5 Mbp, while testing used the remainder and K. pneumoniae reads.
- 3. Experimental results: DeepNano was significantly better than Metrichor for 1D reads on both strands and both datasets, while its 2D accuracy was slightly higher.
- 3. Experimental results: On K. pneumoniae, DeepNano underestimated 57.5% GC content by 1%, compared with 2% for Metrichor.
- 3. Experimental results: 0.0003 seconds per event was DeepNano's 1D processing time on one CPU thread, versus approximately 0.01 seconds for Metrichor.For the same example read, DeepNano took 1.5 seconds for template, 1.3 seconds for complement, and 11.3 seconds for 2D data.
- 3. Experimental results: DeepNano required extensive training computation: three weeks on one CPU for 1D networks and five weeks across GPU and 24-CPU training for the 2D network.The same parameters worked across different genomes produced with the same MinION chemistry.
4 Conclusion and further research
The paper presents DeepNano as a more accurate and computationally efficient alternative to Metrichor's HMM-based basecalling. Further work targets accuracy improvements, independence from Metrichor preprocessing, and richer event representations.
- 4. Conclusion and further research: DeepNano is presented as a new basecaller offering a more accurate and computationally efficient alternative to Metrichor's HMM-based methods.
- 4. Conclusion and further research: Increasing network size could improve accuracy, but would require more training data and make training and basecalling slower.
- 4. Conclusion and further research: Ensembling networks could improve accuracy but would also slow basecalling, while dark knowledge could retain fast inference with a small network at the cost of time-consuming training.
- 4. Conclusion and further research: DeepNano still depends on Metrichor-generated scaling parameters and template-complement alignments, so bypassing Metrichor requires broader testing and alternative alignment methods.
- 4. Conclusion and further research: Alternative event segmentation or additional raw-signal features might improve accuracy, but tested options produced mixed results.