What is limiting the accuracy of genome sequencing today?

Question

What are the current limiting factors of genome sequencing accuracy? By accuracy I mean a closeness relation between the sequenced genome and the finally assembled (I am not sure whether there is a proper name for this metric). I hope this way of measuring accuracy is useful since it also captures errors introduced during read-alignment (if short-read sequencing technology is used) and assembly.

As I understand it, there are two sources of errors limiting accuracy: errors in determining the correct bases and errors made in data analysis (read alignment, assembly, etc.). Which of these two sources is responsible for most errors in long- and short-read sequencing techniques? Are a significant amount of errors stemming from the data analysis?

As @David guessed correctly, I am a student (engineering) and wondering whether accuracy may be significantly improved by better algorithms.

As I currently understand it, short-read sequencing techniques are accurate but the repetitive regions are hard/impossible to align, while long-read sequencing techniques are more error-prone, and long-read & accurate sequencing (HiFi) is very expensive. Hence, my overly simplified perspective suggests that algorithmic improvements may continue to improve cheap long-read accuracy through hybrid approaches or improve the alignment and assembly of short reads. Is that correct?

The resources I used were:

https://www.pacb.com/blog/understanding-accuracy-in-dna-sequencing/ https://spectrum.ieee.org/tech-talk/biomedical/diagnostics/99-9-percent-accurate-genome-sequencing and the paper recommended by @Maximilian Press

Answer 1

One that isn't on your list: cost.

• In the form of Oxford Nanopore data, we have extremely long reads of low accuracy. These are not too expensive but on their own they are probably not viable for many applications.
• In the form of Illumina data, we have extremely abundant, cheap, high-accuracy short reads, suitable for "counting" approaches. This is great for some variant calling approaches, and also for some kinds of orthogonal measures. On their own, they are not viable for a lot of applications, though.
• In the form of PacBio (specifically HiFi) data, we have long(ish) reads of high accuracy. These are the total package, and you can in principle use them pretty much on their own for any given application that I can imagine. The issue is that they remain quite expensive.

Here is a recent review as a reference that goes over these (I haven't looked at it in detail, but it seems to capture the tradeoffs accurately).

We have gotten extremely good algorithmically at using all of these various kinds of data, and at hacking them in various ways to complement each other, but we can't get around the fundamental constraints of contiguity (length) and accuracy.

The HiFiAsm assembler for example is at peak performance for the moment, and it combines PacBio HiFi data with e.g. Hi-C data (a hack of Illumina) to deliver pretty much whole diploid human genomes, which is really an astonishing technical achievement. It was also used to assemble the redwood genome, which is 10X the size of the human genome.

The issue is taking these technical possibilities and getting them to a scale and a technological refinement where we can generate this extremely high-quality data at-will and for all of the applications where we would like to use it.