Thread by @ewanbirney

*trumpets* A new preprint by colleagues in @PHE_uk from @isaperena's group and myself (my first infectious epidemiology paper!) on single source transmission of COVID19 using viral genotyping to understand relative risk of transmission settings. papers.ssrn.com/sol3/papers.cf…

Background; we have known for a long time that there is overdispersion of SARS-CoV-2 transmission; some estimates are that 20% of settings/events account for 80% of transmission. Understanding where these transmission events occur is important for non-pharmaceutical interventions

Furthermore, if we can be confident of spotting these individual small-scale super-spreading events and inform other individuals who are at risk of infection at the same time we can highlight people who are at the higher risk for infection, eg, asking them to get a test.

(there is an incubation time twist to this which thankfully works in our favour in the way the data is collected; this "backtracing with action" is one of the planks of the successful Japanese response to SARS-CoV-2)

The English Test and Trace system since August 2020 has capture "backwards tracing" attempting to highlight points of infection, but as one can imagine there are a large number of places people have been 5 to 7 days ago from a positive test.

Which ones are the true sites of infection? Here work by Cong Cheng leverages the by-chance widespread viral genotype difference of the Alpha strain vs other strains; in many RT-PCR assays one of the 3 assays failed for the Alpha strain genome, the S-gene target failure (SGTF)

When a single source transmission happens in a setting one expects the recipients to be either all-Alpha or all-not-Alpha, and so concordant SGTF; in contrast, if recipients have a mixture of SGTF status (discordant) then this is incompatible with single source transmission.

As the Alpha variant rose from low levels to complete levels in late 2020, we can choose time windows when this concordance/discordance would be discriminatory (we go for between 20-80% Alpha variant; 50-50 is the most discriminatory time).

We can now look at all possible transmission points where 2 or more people are infected; without knowing the person who infected them (or indeed whether this was the right transmission event) we can calculate the odds ratio of seeing concordant transmission by chance

The results are striking. Some settings have over a 50 fold higher odds ratio for concordance than expected.

Supportive of a biological / transmission phenomena, the odds ratio of concordance *goes up* as the cluster size of people potentially infected at these events - this is absolutely the opposite effect if this was being driven by random events.

Furthermore if we contrast co-occurrence of positive people in the same postcode in the same 24 hours (Type 2 in this figure) from co-occurrence on the 2 flanking days (Type 3 groups) we see a collapse of this odds ratio.

All of this is evidence that we have strong statistical signal that we are detecting single-source transmission. Now... having I hope persuaded you this is valid, let's dig into a bit of the detail.

The first thing to realise is that this is *not* proportion of transmission by setting. For example, shops have many reported potential transmissions, but with the lowest single source; personal services (beauticians, hairdressers) the opposite way around.

These number by themselves can't tell you whether the amount of transmission attributable to shops is going to be more or less than beauticians/hairdressers. A similar issue is around schools. Because the recall of school attendance is so good and consistent we get strong results

However, this might be school "setting", ie the social structures around schools (for example, attending casual gatherings in houses) as it is about the school setting itself, though we note that teacher<=>pupil concordance is similar to pupil<=>pupil.

The recall bias and consistency of population in schools (contrast to shops) is a real headache though here, and for me you have to put this school data into a different mental category from the rest.

(the other biases is that England was in a variety of NPIs stages - the dreaded "tiers" - and so not all settings were being sampled; eg, at the time in many places pubs and restaurants were closed)

Stepping all the way back, what is striking to me is just how risky for single source transmission visiting friends and staying with friends are. No surprise, and consistent with previous studies, but it stresses that the some of our most "social" everyday interactions

This work has also given the operational side of Test and Trace more confidence I believe in triggering QR check in warnings for risky events. Here the fact that backtracing happens by definition on the second positive from a site means that these triggers can recommend tests

ie, there is no need to wait for incubation because other people at risk would have been infected at the same -5/-7 days back.

Finally this shows one of the insights that can be gained by combining viral genotype data with contact tracing data, and I am looking forward to doing more (or cheerleading for more) of this sort of analysis using genomics as well as genotyping.

It's been a great collaboration between @PHE_uk and myself (I feel... a bit of an impostor, but I did cheerlead and agitate for the power of this analysis) and given me a massive step up in understanding of PHE data and operational aspects. Respect to the @PHE_uk crew!