Comparison of Oxford Nanopore basecalling tools
Ryan R. Wick, Louise M. Judd and Kathryn E. Holt
Department of Biochemistry and Molecular Biology, Bio21 Molecular Science and Biotechnology Institute, University of Melbourne, Australia
This repository uses a bacterial genome to assess the read accuracy and consensus sequence accuracy for Oxford Nanopore Technologies (ONT) basecallers. Albacore v2.1.10, Guppy v0.5.1 and Scrappie raw v1.3.0 (all developed by ONT) were the best performers for read accuracy, and Chiron v0.3 produced the best assemblies. Consensus sequence accuracies reached approximately 99.75%, revealing that even the best basecallers still have systematic error. Nanopolish, used with its methylation-aware option, was able to raise consensus accuracy to about 99.9%. Most post-Nanopolish assemblies have similar accuracy, making basecaller choice relatively unimportant if Nanopolish is used.
Table of contents
- Basecallers tested
This repository contains a comparison of available basecallers for Oxford Nanopore Technologies (ONT) sequencing reads. It's a bit like a paper, but I decided to put it here on GitHub (instead of somewhere more papery like bioRxiv) so I can come back and update it as new versions of basecallers are released.
Basecallers are programs which translate the raw electrical signal from an ONT sequencer to a DNA sequence. Basecalling is interesting because it's a hard machine learning problem (modern basecallers all seem to tackle it with neural networks) and because it's a huge part of what makes ONT sequencing good or bad. Getting a piece of DNA through a pore and measuring the current is only half the battle; the other half is in the computer. It's a very active field, with both ONT and independent researchers developing methods.
For each basecaller, I assessed the accuracy of the reads and of the resulting assembly. Read accuracy is interesting for obvious reasons – more accurate reads are nice! Assembly accuracy is interesting because shows whether the read errors can 'average out' with high sequencing depth. This provides a window into the nature of the basecalling errors. For example, consider a hypothetical set of reads with a mediocre accuracy of 80% but a truly random error profile. Despite their error rate, these reads could result in a perfect assembly. Now consider a set of reads with an excellent 98% accuracy but they all make the same mistakes (i.e. error is systematic, not random) – their assembly will have an accuracy of 98%, no better than the raw reads and worse than our first hypothetical assembly. Which read set is better? That probably depends on how you're using them, but in my line of work, I'd prefer the first.
I hope these results help to answer the question: Should I go back to old reads and re-basecall them with a newer basecaller? Doing so could take a lot of CPU time, so you probably don't want to do it unless it would bring a significant improvement.
I used a 1D R9.4 dataset of native DNA Klebsiella pneumoniae reads for this analysis, so my results may be biased toward that kind of data. I'm not sure how consistent these results are with other data types, e.g. eukaryote genomes, R9.5 flow cells, 1D2 kits, PCRed DNA, etc.
For each basecaller I have only used the training model(s) included with the program. Custom training of the neural net is out of scope for this analysis. Similarly, whenever low-level tuning parameters were available, I used the defaults.
When a basecaller had many versions, I skipped superseded patch versions. E.g. I tested Albacore v1.1.2 but not v1.1.0 and v1.1.1.
Nanonet is ONT's first generation neural network basecaller. I used the latest commit (at the time of writing) on the master branch. This seems to be functionally identical to v2.0.0, so that's what I've called it. Nanonet no longer appears to be under active development, so this may be the final version.
nanonetcall --chemistry r9.4 --write_events --min_len 1 --max_len 1000000 --jobs 40 raw_fast5_dir > /dev/null
I set the
--max_len options so Nanonet wouldn't skip any reads. While Nanonet outputs its basecalls to stdout in fasta format, I've discarded that and instead used the
--write_events option so it stores fastq basecalls in the fast5 files, which I can extract later. Unlike Albacore, which makes a copy of fast5 files, Nanonet modifies the original ones in the
Albacore is ONT's official command-line basecaller. I tested versions 0.8.4, 0.9.1, 1.1.2, 1.2.6, 2.0.2 and 2.1.10. I previously also tested v1.0.4, but its results were extremely similar to v1.1.2, so I've removed it from the analyses. All tested versions were released in 2017, which shows the rapid pace of basecaller development. The transducer basecaller (helps with homopolymers) was added in v1.0. Basecalling from raw signal (without segmenting the signal into events) first appears in v2.0. Albacore can be downloaded from the Nanopore community, but you'll need an account to log in.
# Albacore v0.8.4 and v0.9.1: read_fast5_basecaller.py -c FLO-MIN106_LSK108_linear.cfg -i raw_fast5_dir -t 40 -s output_dir # Albacore v1.1.2 and v1.2.6: read_fast5_basecaller.py -f FLO-MIN106 -k SQK-LSK108 -i raw_fast5_dir -t 40 -s output_dir -o fast5 # Albacore v2.0.2 and v2.1.10: read_fast5_basecaller.py -f FLO-MIN106 -k SQK-LSK108 -i raw_fast5_dir -t 40 -s output_dir -o fast5 --disable_filtering
Albacore v1.1 and later can basecall directly to fastq file with
-o fastq. This saves disk space and is usually more convenient (especially since Nanopolish v0.8), but for this experiment I used
-o fast5 to keep my analysis options open.
Guppy is ONT's new basecaller that can use GPUs to basecall much faster than Albacore. Both the GridION X5 and PromethION contain GPUs and use Guppy to basecall while sequencing. Guppy can also use CPUs and scales well to many-CPU systems, so it may run faster than Albacore even without GPUs. ONT employees have told me that its neural network structure is the same as Albacore's, and I got the vibe that it's ultimately destined to replace Albacore as the main production-ready ONT basecaller.
Guppy is not yet publicly available, so you'll need to sign the ONT developer license agreement or contact ONT to give it a try.
# Guppy v0.3.0: guppy --input_path raw_fast5_dir --config dna_r9.4_450bps.cfg --save_path output_dir # Guppy v0.5.1: guppy_basecaller --input_path raw_fast5_dir --config dna_r9.4_450bps.cfg --save_path output_dir
Scrappie is ONT's research basecaller. Successful developments here seem to eventually work their way into Albacore. I tested versions 1.0.0 and 1.3.0, and I skipped v1.1.1 and v1.2.0 because they are functionally equivalent to v1.3.0 for the models they have in common.
Scrappie can be run as
scrappie events (basecalls from event segmentation) or as
scrappie raw (basecalls directly from the raw signal). For Scrappie v1.0.0, running as
scrappie events relies on pre-existing event data in the fast5s, so I used the fast5s produced by Albacore 1.2.6 (the last version to do event segmentation). In Scrappie v1.3.0, there are four different raw basecalling models to choose from (raw_r94, rgr_r94, rgrgr_r94 and rnnrf_r94) and I tried each. As a side note, the rgr_r94 and rgrgr_r94 models are referred to as 'pirate' models, for reasons explained here.
# Scrappie v1.0.0: scrappie events --albacore --threads 40 albacore_v1.2.6_fast5 > scrappie_v1.0.0_events.fasta scrappie raw --threads 40 raw_fast5_dir > scrappie_v1.0.0_raw.fasta # Scrappie v1.3.0: scrappie events --threads 40 raw_fast5_dir > scrappie_v1.3.0_events.fasta scrappie raw --model raw_r94 --threads 40 raw_fast5_dir > scrappie_v1.3.0_raw_raw_r94.fasta scrappie raw --model rgr_r94 --threads 40 raw_fast5_dir > scrappie_v1.3.0_raw_rgr_r94.fasta scrappie raw --model rgrgr_r94 --threads 40 raw_fast5_dir > scrappie_v1.3.0_raw_rgrgr_r94.fasta scrappie raw --model rnnrf_r94 --threads 40 raw_fast5_dir > scrappie_v1.3.0_raw_rnnrf_r94.fasta
Unlike Albacore, Scrappie does not have fastq output, either directly or by writing it into the fast5 files – it only produces fasta reads.
DeepNano is developed by Vladimír Boža and colleagues at Comenius University and is described in this paper. I could not find version numbers associated with DeepNano, so I used the current commit (at the time of this writing): e8a621e.
export OMP_NUM_THREADS=1 python basecall.py --chemistry r9.4 --event-detect --max-events 1000000 --directory raw_fast5_dir --output deepnano.fasta
--max-events parameter is important – without it DeepNano will use a default of 50000 which results in a maximum read length of about 30 kbp (longer reads are truncated). It's also worth noting that the basecalling script I ran was not the
basecall.py in the DeepNano base directory, but rather the
basecall.py in the
r9 directory, which can handle R9.4 reads.
Chiron is a third-party basecaller developed by Haotian Teng and others in Lachlan Coin's group at the University of Queensland and described in this paper. Versions 0.1 and 0.1.4 had some bugs which caused issues with my data, so I tested versions 0.2 and 0.3.
chiron call -i raw_fast5_dir -o output_dir --batch_size 1000
While testing Chiron, I noticed a curious effect. The
--batch_size parameter was described as controlling performance: a larger value runs faster but increases RAM requirements. While I found this to be true, I also saw another effect: larger
--batch_size values improved the read accuracy. The default is 100 and I saw modest read accuracy improvements up to about 500, after which accuracy plateaued. I used
--batch_size 1000 to improve both performance and accuracy.
Version 0.3 added a beam search option (
--beam) which improves accuracy but slows down the basecalling. The default value is 50, and setting that option to 0 turns off the beam search. In a brief test of different values, I found that read accuracy improved up to a beam search setting of about 25 to 50. Higher values failed to improve accuracy but resulted in worse performance, so I stuck with the default of 50.
basecRAWller is a third-party basecaller developed by Marcus Stoiber and James Brown at the Lawrence Berkeley National Laboratory. I previously included basecRAWller in these tests, but it performed poorly. I spoke to Marcus as the Nanopore Community Meeting, and he confirmed my suspicion: the model included with basecRAWller was trained only on human DNA, not bacterial. I therefore removed it from this comparison, but you can still view the results in a past version of this repository.
Unfortunately, I cannot compare with the old cloud-based Metrichor basecalling, as it's no longer available. I also cannot test the integrated basecalling in MinKNOW (ONT's sequencing software). I believe MinKNOW's integrated basecalling shares much in common with Albacore, but I don't know which Albacore versions correspond to which MinKNOW versions.
If you'd like to try this analysis for yourself, here's what you need to do. I tried to make this process reproducible, but some aspects may be specific to my setup, so you may need to modify parts to make it work for you. In particular, the
nanopolish_slurm_wrapper.py script assumes you're using a SLURM-managed cluster, so other users will probably need to change that one.
You'll obviously need a set of ONT reads. Put them in a directory named
01_raw_fast5. I used the same ones from our recent paper: Completing bacterial genome assemblies with multiplex MinION
sequencing (and its corresponding repo). Check out that paper if you're interested in the wet lab side of things.
You'll also need Illumina reads for the sample (named
illumina_2.fastq.gz) and a good reference sequence (named
reference.fasta), e.g. a completed hybrid assembly. For the reference-based assembly step later, it's important that circular replicons in
circular=true in their fasta header.
My reads came from a barcoded run, so I first had to collect only the fast5 files for my sample. I did this by analysing the fastq file of our confidently-binned reads (see the paper for more info). This process should have excluded most of the very low-quality reads, because such reads would not have been confidently binned. I also discarded any fast5 files less than 100 kilobytes in size to remove shorter reads.
If you'd like to try this analysis using the same data I used, here are the relevant links:
- Reference hybrid assembly (
- Illumina reads (
- Raw fast5 files
I've indicated the versions I used, but the exact versions may or may not be important (I haven't checked). However, it is necessary to use a recent version of Nanopolish. Since v0.8, Nanopolish can be run without event-data-containing fast5 files, which lets it work with any basecaller! However, for non-Albacore basecallers I did have to alter read names – more on that later.
The commands I used are described above in the Basecallers tested section. Regardless of which basecaller was used, the reads need to be put in a
02_basecalled_reads directory in either
Read ID to fast5 file
It is also necessary to make a
read_id_to_fast5 file which contains two columns: read ID in the first and fast5 filename in the second. For example:
0000974e-e5b3-4fc2-8fa5-af721637e66c_Basecall_1D_template 5210_N125509_20170425_FN2002039725_MN19691_sequencing_run_klebs_033_restart_87298_ch173_read25236_strand.fast5 00019174-2937-4e85-b104-0e524d8a7ba7_Basecall_1D_template 5210_N125509_20170424_FN2002039725_MN19691_sequencing_run_klebs_033_75349_ch85_read2360_strand.fast5 000196f6-6041-49a5-9724-77e9d117edbe_Basecall_1D_template 5210_N125509_20170425_FN2002039725_MN19691_sequencing_run_klebs_033_restart_87298_ch200_read1975_strand.fast5
This will be used by the
fix_read_names.py script to ensure that all basecalled reads are named consistently and compatible with Nanopolish.
analysis.sh script automates most of the remaining steps. It will:
- Change the read names to a consistent, Nanopolish-friendly format (
- Align the reads to a reference and make a tsv file of read accuracies (
read_length_identity.py). This only uses the aligned parts of the read to calculate the read's identity. The definition of 'identity' is the same as how BLAST defines it: the number of matches in the alignment divided by the total bases in the alignment (including gaps). If less than 50% of a read aligned, it is deemed unaligned and given an identity of 0%. This script also determines the read length to reference length ratio for each read, to see if insertions or deletions are more likely.
- Prepare reads for assembly (
- Do a reference-based assembly (
rebaler). Rebaler conducts multiple rounds of Racon, so the final assembly accuracy is defined by the Racon consensus (which in my experience is a bit higher accuracy than a Canu assembly).
- Assess the assembly accuracy. By chopping the assembly into 10 kbp pieces (
chop_up_assembly.py) and treating them like reads (
read_length_identity.py), we can get a distribution of assembly identity instead of just a single value.
- Run Nanopolish (
- Assess the Nanopolished assembly (
Put all your resulting tsv files in a
results directory and run
plot_results.R to generate figures. Edit the
basecaller_colours vectors at the top of that script to control which results are included in the plots.
I did not quantify speed performance in this analysis, mainly because I ran different basecallers on different hardware, which makes a fair comparison hard. There are, however, a couple points worth making.
Chiron was the slowest basecaller tested. When run on CPUs, it is so slow that it could only be used for very small datasets. It is much faster on GPUs (I ran it on a GTX 1070), but it still took Chiron v0.3 over two weeks to basecall my read set of 1.2 Gbp (~1 kb/sec). Guppy, on the other hand, is by far the fastest. On the same hardware (GTX 1070), it basecalled the read set in less than 30 minutes (~700 kb/sec). Running on CPUs, Scrappie had a mediocre performance of about 7 kb/sec for most of its available models, but the rnnrf_r94 model was much slower at about 1 kb/sec.
You might expect that each basecaller would produce approximately the same total yield. E.g. a read that makes a 10 kbp sequence in one basecaller would be about 10 kbp in each basecaller. That's mostly true, but Nanonet is a notable exception. For most reads, it produced a much shorter sequence than other basecallers, sometimes drastically so. For example, all versions of Albacore basecalled one read (d2e65643) to a 34+ kbp sequence. Nanonet produced 518 bp for the same read. I don't have an explanation for this odd behaviour.
Other oddities you might notice are Albacore v0.9.1, which produced more sequence than the others, and Scrappie events, which produced less. These are explained in the Relative read length section.
This addresses the most obvious question: how accurate are the basecalled reads? The plot above shows read identity distributions, with the medians (weighted by read length) marked as a horizontal line. Unaligned reads were given an identity of 0% and fall to the bottom of the distribution. Reads with an actual identity below 65% often fail to align and end up at 0%.
Nanonet performed poorly, with a low median and a significant proportion of unaligned reads. Its curiously high peak of about 99% results from the short output sequences discussed above. While a few Nanonet 'reads' did indeed align to the reference with up to 99% identity, these were actually just small fragments (hundreds of bp) of larger reads.
Albacore v2.1.10, Guppy v0.5.1 and Scrappie v1.3.0 performed the best overall. All three of these are developed by ONT and share much of their design, so similar performance makes sense. In particular, Albacore and Guppy produced nearly identical results, a trend that will continue in more analyses below. Scrappie's rnnrf_r94 model did the best overall, but only by a small margin, and it was very slow to run.
Relative read length
This plot shows the distribution of read length to reference length for each alignment. It shows whether the basecaller is more prone to insertions or deletions. 100% (same length) means that insertions and deletions are equally likely. <100% means that deletions are more common than insertions. >100% means that insertions are more common than deletions. Albacore v0.9.1 stands out with many overly-long reads, while Scrappie events tends to make short reads. This explains the total yield differences we saw earlier.
I found it curious that many basecallers had a distinctly bimodal distribution (particularly pronounced for Chiron v0.3). This effect seems to be related to the timing of the MinION run. It was started at about 4 pm and MinKNOW crashed at 10:30 pm, halting the run. Nobody was in the lab to notice, and the next day was a public holiday. Thankfully Louise came in that afternoon, saw the crashed run and restarted it at about 3 pm. That meant the flow cell sat for about 16.5 hours not being used. When I plot read length against signal length and colour by the restart, the effect is obvious. It's still not entirely clear why the restart has resulted in shorter basecalled reads, but the effect is present in all basecallers. A possible clue is that the raw signal values are lower after the restart: with a median value of about 450 before and 370 after.
This analysis is my personal favourite: how accurate are the consensus sequences? I don't particularly care if individual reads have low identity if they can produce an accurate assembly.
In previous versions of my analysis, Albacore consistently led the pack with consensus accuracy, but the newly released Chiron v0.3 produces the best consensus sequences by a huge margin. Its assemblies have approximately half the errors of the second-best assembly (Albacore v2.1.10). I'm not sure if this improvement is solely due to Chiron v0.3's new beam search functionality, but whatever the cause, kudos to the Chiron developers!
It's also interesting to look at the assembly relative length, like we did for reads:
This shows which basecallers are more prone to consensus sequence insertions (e.g. Albacore v1, Scrappie raw v1.0.0 and DeepNano) and which are more prone to deletions (most of the rest).
Read vs assembly identity
Here I've plotted the median read identity and median assembly identity for all basecallers – zoomed out on the left, zoomed in on the right. The shaded zone is where assembly identity is worse than read identity. That should be impossible (unless you've got a very bad assembler).
This shows how much of each basecaller's error is random vs systematic. If a basecaller had the same read and assembly identities (i.e. on the edge of the shaded zone), that would imply that all its error is systematic and every read is making the same mistakes. Thankfully, assembly identities are nowhere near that low. Conversely, if a basecaller had an assembly identity of 100%, that would imply a there was little systematic error so all problems could be fixed in the consensus.
You might expect that a basecaller's read and assembly identities would be tightly correlated: low-identity reads produce low-identity assemblies and vice versa. That is mostly the case, with Albacore v0.9.1 and Chiron v0.3 being the strongest outliers – they produce better assemblies than you might expect due to relatively low systematic error.
Nanopolish assembly identity
I ran Nanopolish v0.9.0 with two different configurations, with and without the
--methylation-aware dcm,dam option (added in v0.8.4).
First, here are the results without that option. The plot shows the assembly identity distributions after Nanopolish, with pre-Nanopolish distributions lightly drawn underneath:
In most every case, Nanopolish improved the assembly accuracy, and most post-Nanopolish assemblies are quite similar to each other and near 99.7% accurate. Chiron v0.3 is the only case where Nanopolish failed to make a significant improvement, as its assembly already reached 99.7%.
In a few cases, Nanopolish resulted in a ~100 bp insertion, which caused the distribution to extend down to 98.5%. I'm not sure if this is an issue with Nanopolish (the insertion was erroneous) or with the reference sequence (the insertion was correct). Three of the assemblies did not reach the 99.7% accuracy of the others. Nanonet's truncated reads may have caused problems for Nanopolish. Scrappie raw v1.0.0 and DeepNano probably suffered due to the low accuracy of their pre-Nanopolish assemblies.
The upside seems to be that if you're planning to use Nanopolish, then your basecaller choice may not be very important. Any basecaller, as long as it isn't awful, should be fine. The downside is that Nanopolish makes little to no improvement for an already good assembly.
Now here are the results with the methylation-aware option:
The improvement is huge! Most assemblies now reach about 99.9% accuracy. Most of the remaining errors are deletions homopolymers:
assembly: AACCGCTACCACTCATCTTCCCCCGCCTCGC-GGGGA-TTTTTTGCTTG reference: AACCGCTACCACTCATCTTCCCCCGCCTCGCGGGGGATTTTTTTGCTTG ^ ^
The Nanopolish results show that, at least for this dataset, methylation is a major factor in consensus accuracy. For example, when a pre-Nanopolish assembly is aligned to the reference, many of the errors correspond to the Dcm motif (
assembly: CCCGG CCCGGG CCGGG CC-GG CC-GG CCA-G CACGG CCCGGG reference: CCTGG C-CAGG CCAGG CCAGG CCTGG CCAGG CCTGG C-CTGG ^ ^ ^ ^ ^ ^ ^ ^^ ^ ^
While Nanopolish can correct many of these errors, it would be better if the basecallers themselves could properly recognise modified bases. While calling modified bases as modified bases would be fascinating and useful, a first step would be to call modified bases as their canonical base. E.g. 5-mC called as a C, 4-mC as a C, 6-mA as an A, etc. This could be a tricky problem, as different organisms have different enzymes which modify bases at different sequence motifs. Perhaps basecallers need very broad training sets which include as many such motifs as possible. Or perhaps each basecaller needs multiple trained models, each on different organisms, and an automatic method for choosing the appropriate one.
Medaka is trying to solve a similar problem to Nanopolish: improving the consensus sequence accuracy using the alignment of multiple reads. It differs from Nanopolish in two significant ways. First, Medaka uses neural networks where Nanopolish uses HMMs. Second, it uses basecalled reads, not the raw signal (though this is likely to change in the future). Here I test Medaka v0.2.0:
While Medaka could improve most assemblies, it was overall less effective than Nanopolish. It particularly seemed to struggle with older basecallers that use event segmentation (as opposed to modern basecallers which call directly from the signal). Medaka crashed when working with a few read sets, which is why there are missing plots in the figure.
Combining different basecallers
This section previously looked at how well a combination of Albacore and Chiron reads assemble. The idea was that perhaps two different basecallers can somewhat 'cancel out' each other's systematic error, leading to a better assembly. This was the case with Albacore and Chiron v0.2, but Chiron v0.3 reads assemble so well that combining them with Albacore reads gives no improvement (it actually makes the assembly a bit worse).
I don't think this is relevant anymore, so I've removed it. You can see my earlier results in a past version of this repository if you're still interested.
Combining different polishers
I tried assembly polishing with both Medaka and Nanopolish (methylation-aware) to see if a joint approach could yield better accuracies. I tried both Medaka followed by Nanopolish and vice versa, but neither combination could improve upon Nanopolish alone:
All supervised learning depends on a good training set, and basecalling is no exception. A nice example comes from the rgrgr_r94 model in Scrappie v1.1.0 and v1.1.1. The primary difference between these two versions is that in v1.1.0, only human DNA was used to train the basecaller, whereas v1.1.1 was trained with a mixed set of genomes (described here by Scrappie author Tim Massingham). I didn't include v1.1.0 in the above plots because it's a superseded version – it's here only to show the difference a training set makes. The difference in read identity is huge, but assembly identity had a subtler improvement:
There are three obvious basecaller recommendations to make: Albacore v2.1.10, Guppy v0.5.1 and Chiron v0.3.
The current version of Albacore (v2.1.10) is probably the best basecaller choice for most users. It has very good read accuracy, produces decent assemblies, runs reasonably quickly, is easy to use and has many useful features such as barcode demultiplexing. If you have a GPU, Guppy v0.5.1 can produce essentially the same basecalls in much less time, but it is not yet publicly available and lacks barcode demultiplexing. I suspect the day is coming when Guppy replaces Albacore and it may then become my definitive recommendation.
My last recommendation, Chiron v0.3, is more complicated. Its pre-Nanopolish assembly accuracy is outstanding, and it also had the best post-Nanopolish (methylation-aware) assembly, though only by a small margin. It may therefore be the best choice when assembly accuracy is paramount. However, Chiron is much slower than Albacore and only a viable option if you have a powerful GPU to accelerate the process. Even with powerful GPUs, basecalling an entire MinION run could take a very long time. I would therefore only recommend Chiron to users with a small volume of reads.
Scrappie raw v1.3.0 (rgr_r94, rgrgr_r94 and rnnrf_r94 models) also did quite well and had the highest read accuracy. However, Scrappie is a research product, labelled as a 'technology demonstrator' and lacks nice features present in Albacore, such as fastq output and barcode demultiplexing. I therefore think Albacore is a better choice for most users. Nanonet and DeepNano should probably be avoided, but I'm happy to revisit them if they are updated.
Anyone interested in maximising assembly accuracy should be using Nanopolish. It improved all assemblies and took most up to about 99.9% (with the methylation-aware option). If you only care about assembly identity, Nanopolish makes your basecaller choice relatively unimportant.
While Medaka does not improve assemblies as well as Nanopolish, it requires only a fasta/fastq file, not the raw fast5 files. It may therefore be the best choice for assembly polishing when raw reads are not available. However, the 'Future directions' section of Medaka's documentation indicates that signal-level processing may be in its future. Furthermore, Medaka uses neural networks, unlike Nanopolish's HMMs. The authors suggest that just as neural networks have outperformed HMMs in basecallers, they will also prove superior in consensus algorithms. Watch this space!
My future work is easy: trying new versions and new basecallers as they are released and adding them to this analysis. Check back occasionally for new data! The much harder task lies with the basecaller authors: reducing systematic error.
In my opinion, low consensus accuracy is ONT sequencing's biggest issue and the one area where PacBio still has a distinct advantage. It makes it hard to recommend an ONT-only approach for many types of genomics where accuracy matters (read more in our paper on this topic). One approach to improving consensus accuracy is with polishing tools such as Nanopolish and Medaka. Even better would be for the problem to be solved in the basecallers themselves: reads with a truly random error profile could yield exact consensus sequences. Either way, if and when assemblies approach 100% accuracy, ONT will be a true alternative to Illumina.
Did I miss anything important? Can you shed any light on oddities that I couldn't explain? Please let me know through the issue tracker!