Respiratory viruses hijack your social network to facilitate their spread1. Liquid droplets saturated with virus, released when you sneeze, have about a 6 metre radius (just an estimate). So the people you infect are most likely to be your close friends and relatives.

Pandemic avian influenza H5N1, swine H1N1 and SARS acquired the ability to jump to humans and cause damage with only a couple of changes to the viral genome. The right combination of proteins to facilitate their entry into cells as well as the presence of ‘virulence factors’ meant that virus could easily spread from host to host, and it is this spread that we find so difficult to control.

In response, the Chinese Academy of Sciences (CAS) created a dystopian sounding smartphone application called PEARL (Probing Entity Aggregation in Real Life).

PEARL works by using your short range, line of sight Bluetooth function to build up a map of your interactions with other volunteers. Bluetooth uses ultra high frequency waves, so has a transmission range like that of a droplet transmitted virus. When PEARL enabled volunteers come within this range of each other, their interaction is registered, and a close proximity interaction network, or CPI, can be built up and mapped:

Figure 1: CPI networks in two separate Chinese colleges (A and B). Each blue dot is a person, and the thickness of the yellow line depicts the length of time of the interaction. Adapted from 1.

The little blue dots are people, and their interactions (and their inter-interactions) are thick or thin lines depending on the efficiency –  the length, essentially – of the interaction. The two colleges are clearly different.

Real data is completely different to the models

In figure 2A from the paper (1), you can see that the ‘clustering’ versus ‘efficiency’ of the Chinese college (its called SCAU here – the red dots) differs completely from accepted population models in blue, yellow and green. These models are widely used in network science, and you have probable heard of ‘small world’. In addition there are scale-free and uniformly random – neither of which predict correctly the clustering or efficiency of the real social network in the college.

Figure 2: SF, SW and UR are the idealised social networks. They don’t fit with the real data collected at the Chinese College (SCAU). B shows that the Chinese Colleges SCAU and USTB differ also from an American College dormitory: USD and an American School: USHS as well as a French primary school: FRPS. Adapted from (1).

Next, they show that the real social networks they measured differed between Chinese, American and French schools and colleges (Figure 2B). As you can see, microcosms of behaviour exist, and result in alternate networks, even within the same type of situation (students at school).


Influenza epidemics are predicted better when applied to real social network data

So if in reality, social networks differ in microcosms between schools, do they better predict the spread of disease within that school? To answer that, a flu virus infection simulation was applied to the real social network data, and compared with actual flu virus data from the college in question (figure 3).

Figure 3: Real data from an influenza epidemic in the Chinese college (blue and yellow lines) match simulated data applied to the CPI networks acquired here (red line is mean average of green chart).

Figure 3: Real data from influenza epidemics in Chinese colleges (blue and yellow lines) match simulated data applied to the CPI networks acquired here using the bluetooth technology (red line is mean of green chart).

What you can hopefully see from this figure is that the three lines follow each other closely. The simulation based on the PEARL network correctly predicts both waves of viral spread.

Epidemics and CPIs

So, epidemic virus spread is better predicted when simulations are applied to close social networks, which can be collected using an app that volunteers can install on their phones.
What you can hopefully see from this figure is that the three lines follow each other closely. The simulation based on the PEARL network correctly predicts both waves of viral spread.

The implication different groups of people, probably depending on many different factors, occupy different social groups and would spread viruses in different ways.

If enough volunteers take part, the hope is that CAS, with PEARL, can make a huge, highly predictive CPI network – one that encompasses many islands of intra-connected interactions that add up to make a whole.

With this they can refine their models to reflect the spread of a real respiratory virus across many different real world demographics.

Global simulations require gigaflops of computing power

The plan now is to extend PEARL across China, collecting huge datasets of different networks over different durations, before running complex pandemic simulations on them. This map, on the PEARL website, shows how far they have got.

To cope with the gigaflops of computing power required for predictive modelling, CAS@home uses Berkeley’s BOINC (Berkely Open Infrastructure for Network Computing) project. BOINC, installed on your personal computer, hijacks your CPU (and GPU) when you are not using it to run complicated and data rich simulations. This is the same technology being used to simulate millions of different combinations of protein folding events, and to scan radio signals in the SETI project, as well as to predict climate changes based on constantly updated models.

With this kind of crowdsourced computing power, modelling becomes refined, and may provide us with the ability to predict disease spread and save many lives at once, by making the right public health moves at the right time.

If you are interested in running BOINC yourself – it can be downloaded here.


  1. Huang, C. et al. Insights into the transmission of respiratory infectious diseases through empirical human contact networks. Sci. Rep. 6, 31484 (2016).
  2. PEARL.


Written by Michael Shannon