Always a pleasure to organize this with @caityholmes.bsky.social @lauraamike.bsky.social Jay Vornhagen, Wen wen low, & this year joining us @tomstantonmicro.bsky.social & Juan Valencia.

Janitor saves the day again!
www.rdocumentation.org/packages/jan...

Lastly, we'd like to thank YOU, the Kaptive community, for guiding development, spotting bugs and collaborating with us!

But this is just the beginning, we have lots of exciting things in store for the future of Kaptive to make in silico serotyping even better!

#kaptive #klebsiella #acinetobacter

Kaptive 3 is now integrated within Kaptive-Web (kaptive-web.erc.monash.edu), PathogenWatch (pathogen.watch), the new Kleborate 3 framework (github.com/klebgenomics...) and Bactopia (bactopia.github.io/latest/).

Remember to cite us if you use Kaptive for your results, and watch out for "Untypeable"!

We know the command-line can be tricky, so we made the CLI much friendlier 🧑‍💻

For the code-savvy, there's also a Python API allowing Kaptive to be used within your own programs 🧱
All the information you need is in the documentation, which we update regularly: kaptive.readthedocs.io/en/latest/

Kaptive 3 is also much (much) faster than Kaptive 2, taking ~1 second per assembly 🏎️💨

This means that if you don't have a fancy HPC, then don't worry! You can still analyse thousands of your own assemblies on your laptop in a reasonable time! 💻

We then subsampled the corresponding short reads at decreasing depths and created sets of increasingly awful draft assemblies with loci broken over contigs and lots of genes missing.

Kaptive 3 was much more sensitive than Kaptive 2, and maintained accuracy even when the assemblies were awful! 💩

We put together a special dataset specifically designed to test Kaptive. We had completed hybrid assemblies for KpSC (preprints.scielo.org/index.php/sc...) and A. baumannii (RefSeq).

We identified the K- and O(C)-loci in each and visually confirmed each to determine a ground truth Kaptive call 🔎

So enter Kaptive 3, a complete overhaul of Kaptive with a new algorithm designed to handle fragmented loci.

We also refactored (and simplified) the confidence score to be more sensitive for broken loci and missing genes, allowing more Kaptive data to be used when the assembly may not be complete 💯

Because of how Kaptive 2 chose the best match locus, missing locus sequence resulted in a coverage bias for shorter loci in the database such, and could sometimes lead to inaccurate calls!

Ever seen a stray KL107 in your data that didn't make sense?

Yeah, that's why...

So in a nutshell, we traced Kaptive's issues with the Klebsiella K-locus all the way back to the gDNA, where:

The locus region is partially amplified ->
Low sequencing read coverage ->
region doesn't assemble well ->
Untypeable Kaptive call ->
Unusable data 🙅‍♀️

These genes have a very low GC compared to the rest of the Klebsiella chromosome, so we wondered if this was affecting how this part of the genome gets sequenced.

Turns out, these genes show decreased sequencing coverage when reads are prepped with Nextera XT, but not so much with Nextera Flex 🤯

The major drivers of low confidence were K-loci that were 1) broken over contigs 🚫 and 2) missing genes ➡️➡️, events that were mostly co-occurring

Turns out, the genes missing were usually those important for antigenic diversity, in this case the glycosyltransferases that dictate the CPS 🍬 structure

So you may have noticed your Klebsiella K-locus results from previous versions of Kaptive (v1-2) having lots of untypeable calls ("Low" + "None" confidence) with draft assemblies; we certainly did!

This meant that lots of useful seroepi data was unusable, so we started by finding out exactly why 🤔