Some areas of the text also seem formatted in a way less like a research paper and more like, say, a chatbot output..

um, 𝘢𝘤𝘵𝘶𝘢𝘭𝘭𝘺, we call them orthopicobirnaviruses now

Is it something like "by" -> "bystander" -> something like "onlooker" or "watcher" -> the colloquial "peeping Tom" -> further transmogrification?

I have strong memories of an automated plagiarism of a press article about my research into bat virus distributions suggesting we analysed "illnesses of cricket bats"

🏏🤒

Hope we will have many attendees from @royalstatsoc.bsky.social @unisouthampton.bsky.social !

@besaldernetwork.bsky.social
@qedmathsnetwork.bsky.social
@lgbtmath.org
@outtoinnovate.bsky.social
@lgbtqstem.bsky.social

Please let me know any other networks we should share this with!

You can when smartass titles become a tickbox in the CRediT author taxonomy 😁

This is the 2nd preprint to come from collaboration between @thepandemicinst.bsky.social @livuniresearch.bsky.social & CSL Seqirus, supported by @baylism.bsky.social & Joaquin Mould.

We hope this work can help risk assess new bird flu strains and flag key mutations in the wild!

#preprint #avianflu

This stack is able to correctly predict zoonotic potential of sequences in entirely unseen subtypes with AUC=0.95 and F1=0.90, a level of generalisability that is not often seen for machine learning host predictors.

Interestingly, it flags some duck H4 viruses from Americas as having distinct risk.

Training on 12 feature sets over each of 8 segments, we find protein properties are usually best at estimating zoonotic potential from a single segment.

But what about whole genomes? We can combine the best models in a single trained meta-learner (or "stack"), that draws on info from all of them!

We extracted ~19000 influenza sequences from birds and ~600 zoonotic sequences from humans (only non-seasonal subtypes).

Before training, we remove redundancy by grouping similar sequences into clusters. This is important to reduce bias, as most come from just a few subtypes like H7N9 and H5N1.

Lots of ML models can predict human spillover. However for influenza this task is harder because of a) genome segmentation, and b) strong signal within subtype or lineage.

We planned a model training architecture to handle this, ensuring predictions are rooted in virus biology, not shared ancestry.