Thoughts on biosecurity

There is a debate about whether biosecurity constitutes a major civilisation risk in the AI era. This is a reasonable overview of the current thinking. Basically, some people are worried that AI might assist terrorists in designing and producing novel pathogens, with higher transmissibility or immune escape, or even totally chiral to existing life. Others think that AI doesn't solve the true bottleneck which is actually the physical process to make pathogens. Either way, biosecurity is a big concern to effective altruists and philanthropic organisations who are spending large sums of money to address it.

Unlike cybersecurity, where compute can be thrown at both offence and defence in an arms race for more sophisticated systems, biosecurity is said to be asymmetric: terrorists can simply design any random new pathogen, but defenders have to be super targeted (specific surveillance or a specific vaccine, for example). People are therefore most excited about pathogen-agnostic technologies, for e.g. innate immune system engineering, far-UVC indoor irradiation, and wastewater metagenomic sequencing (MGS).

MGS seems to be a popular choice. The idea is you sample wastewater pooled from millions of people, then sequence DNA from absolutely everything you can find. Among the human, animal, food, environmental and other distractions, you will also find pathogen DNA. These sequences can suggest whether pathogens are currently present in the community. An amazingly successful example was the detection of a single measles case in Illinois among wastewater from over 1 million people. This effort was partially funded by Coefficient Giving (via SecureBio).

Among the limitations of MGS are varying lab protocols (both physical and computational), sensitivity and cost. However, in my view, the biggest challenge is more prosaic: even if we find a novel pathogen sequence in the community of over a million people, what are we going to do about it? We won't know if this signal is noise, whether it's a true positive but isolated case, where this case is, whether it's spreading.

A truly useful biosurveillance system will require distributed wastewater sampling, on the resolution of a building, e.g. an office, a school or a hospital. Only if this sampling could be made autonomous, joined up and scaled, would a public health authority have any realistic chance of responding in a useful way to such a signal. If we had such sequence data, we could use techniques such as phylogenetics and phylodynamics (such as in my PhD) to estimate the spread of any pathogen. AI would doubtlessly be necessary to manage this at a vast scale.