Dr Liam Brierley

@liambrierley.bsky.social

437 Followers 600 Following

Virologist, statistician, and science presenter. Runs @vibelab.co.uk Research Fellow at @cvrinfo.bsky.social Ambassador for @royalstatsoc.bsky.social Five parts emerging virus epi, two parts R/compsci, ten parts caffeine. he/him 🏳️‍🌈♾

A dotplot showing predicted zoonotic probability for all test sequences, which represent 13 subtypes in total. There are blue bird-only sequences toward the left, which generally cluster below the threshold for zoonotic prediction, and red confirmed zoonotic sequences on the right which show subtype specific patterns. Human zoonotic H5N1 sequences are predicted with consistent confidence, all just over the threshold to be considered zoonotic, while H7N9 shows much more variability with a fraction of sequences being misclassified as not zoonotic, and a fraction being classified correctly with extremely strong confidence. Of note is that most sequences of rare spillover events (<10 sequences available, like H10N8 and H3N8) were correctly classified with middling confidence. Highlighted and annotated are also three bird origin sequences that are distinctly higher in predicted zoonotic risk than all other sequences of their same subtypes. They are: H4N6 A/American black duck/New Brunswick/00499/2010, H4N6 A/yellow-billed teal/Argentina/CIP051-91/2011, and H4N8 A/American black duck/New Brunswick/02375/2007.
A heatmap showing how well each modelled combination of feature set and gene/protein predicted zoonotic potential. Protein features showed slightly higher AUC across the board, while the strongest gene/proteins for zoonotic prediction were PB2, PB1, HA, NP, and NS1. We tried five different algorithms, and each was kept at least once, but the best performing algorithms were usually RF (random forests) or XGB (XGBoost).
Histogram over time of the most common avian flu subtypes in our dataset by sequence frequency. Most zoonotic sequences are from H5N1 and H7N9, and peak around 2005-2015. Avian sequences are more mixed and feature many more subtypes including H9N2, H3N8, H10N7, and H5N1, all at highly variable frequencies over time.
Model training diagram emphasising we hold out an entire subtype of avian influenza during training in order to make unbiased predictions on it, and that we use multiple algorithms, multiple feature sets, and all eight segments to train models