I work on the Nucleic Acid Observatory at SecureBio, but in this comment Iâm speaking only for myself.
Thanks for writing this, and thanks @mike_mclaren for sending it to me! I wish Iâd happened across it when youâd posted it [1] and could have given more timely feedback!
At a high level I think this is very reasonable, especially sections #1 and #3: itâs very important that you can turn âhere is a collection of highly suspect sequencing readsâ into a well-calibrated and trusted âhere is how concerned we should beâ, and itâs also very important that you can then turn that into âand hereâs what we will doâ. Iâve seen scattered work on both of these, but I think these are key areas for additional work.
On section #2, which is closer to what Iâve been thinking about, there were some places where I wanted to push back or add additional information. A few point-by-point replies:
I even found that disease surveillance is amongst the most well-funded issue priorities within biosecurity. âŚHowever, insofar as I used to think work on early detection is both one of the safest bets for GCBR mitigation and also not very neglected, I no longer think either of these to be true.
While there is a lot of spending on surveillance, this is overwhelmingly allocated to tracking known threats: how prevalent are SARS-CoV-2, Influenza, RSV, etc. Iâm not aware of anyone who has deployed a system capable of alerting on a novel pathogen. While (in my work at the NAO) Iâd like to change this, I do still think itâs quite neglected for now.
Wastewater surveillance data has the disadvantage of not necessarily being connected to a patient or the identifiable spread of a pathogen.
This is definitely a real drawback. On the other hand, once your wastewater-based initial-detection system flags a pathogen you have what you need to create a cheap PCR-based diagnostic, which you can then deploy widely to identify infected individuals (to see how theyâre doing, if they need treatment, and understand the effect the pathogen is having in their bodies, and also to limit spread). This could let you identify many more infected individuals than if you plan was for your initial-detection system to also double as your identification-of-individuals system. But the need for follow-up work ties into the rest of your post, which is that initial detection is only one piece, and if you donât have follow-ups ready to go youâll lose many valuable doubling periods fumbling the next steps.
For various surveillance modalities, lowering time-to-confirmation is an important part of lowering time-to-detection.
Strong agree!
Regularly screening 1% of the population every day is a lot. In the UK, this would be ~680,000 people per day, just over half the 1.3m people the NHS sees every day. As a very loose estimate, we can anchor on the costing provided in the proposed Threat Net architecture. Threat Net has the all-in variable cost per test at roughly $1500[10] in the US. Assuming 1 test = 1 sequencing run (e.g. so it also serves a diagnostic purpose), it would cost upwards of $372bn per year to test this many people[11]. Even assuming 10 samples can be multiplexed, ÂŁ37.2bn would be close to 20% of the UKâs healthcare budget. Waiting to multiplex many more samples than this would contravene the purpose of early detection.
While I agree that daily screening of 1% of the UKâs population would be very expensive, I think the main cost (unless youâre doing wastewater) is in collecting the samples. Thatâs just a massive endeavor! If you get sampling down to $1 person itâs still $250M/ây (68M people Ă 1% Ă 365) in sample collection alone. This range might be possible if you somehow just made it part of the culture (ex: every school child provides a sample every morning) but would be a huge lift.
But I think your estimates for the cost are otherwise too high. The main thing is that if you do pooled testing you can cost-effectively sample hundreds of thousands of people at once. This isnât âmultiplexingâ because youâre not adding barcodes (or something else) that would allow you to map sequencing reads back to specific sampled people. For example, say youâre doing pooled nasal swabs, and youâre sampling a large enough number of people that youâll likely get some high viral load folks. Then the average contribution of a sick person is a relative abundance of maybe 2% (rough analysis; @slg and @Will Bradshaw should have a much more thorough treatment of this out any day now). If you run a NovaSeq X 25B (which gets you 25B sequencing read pairs) on a massively pooled sample and 1:100,000 people are currently shedding, youâll get about 5k (2% á 10,000 Ă 25B) read pairs matching the virus, which is likely (again, hoping to get something out publicly about this soon!) enough to flag it. The list price for the flow cell for a 25B sequencing run is $16k, which is perhaps half the all-in cost. Which means if you did daily NovaSeq X 25B runs youâd be spending $10-$20M/ây ($16k Ă 2 Ă 365) on the sequencing.
This is all very ballpark, but mostly is to say that if youâre talking about directly sampling that many people the main cost is in the collection.
This is quite different if youâre using wastewater to avoid needing to sample that many people, but then your relative abundances are far lower and you need much deeper sequencing (expensive!) to get anywhere close to detection at 1:100,000.
populations that exhibit much higher levels of viral load in the early stages of an infection (e.g. due to weaker immune systems)
I was briefly quite excited about this direction, but a major issue here is that these populations are, because of their vulnerability, somewhat epidemiologically isolated: theyâre probably not close to the first people getting sick in anywhere.
Ideally, we should not be content with a 5% chance of a detection system failing us for a GCBR.
That seems too strong, at least in theory. We should evaluate the marginal detection effort in terms of how much it reduces risk for the cost. For example, a detection system that was very cheap to deploy and only flagged a pandemic 30% of the time could be well worth deploying even though it most likely doesnât help. And perhaps scaling it up to 90% probability of detection would remain cost effective, but getting it up to 99% would be massively more expensive for diminishing returns and so doesnât work out. On the other hand, political considerations might come into play: a system that on average doesnât work could be a hard sell even if itâs relatively cheap.
My first takeaway is a prima facie scepticism about the value of early detection for preventing GCBRs in particular compared to preventative measuresâwith the exception of stealthier pandemics.
I also donât expect work focused on initial-detection to help much outside of stealth pandemics. I see my work at the NAO as almost entirely valuable in proportion to the likelihood that someone creates (or perhaps would create, if not deterred by the likelihood of it being flagged) a stealth pandemic.
That said, there is one other situation Iâve seen bounced around where I think this kind of monitoring could be valuable outside of stealth scenarios: when we know thereâs something happening in a region inside a country we have very poor relations with. In that case if we had a system of ongoing monitoring we could pivot it to populations enriched for contact with that region and (ideally) ramp it up.
[1] A few weeks ago I happened to go through and changed my settings to give a big boost to any biosecurity-tagged posts, so I think if this came through again Iâd see it.
I work on the Nucleic Acid Observatory at SecureBio, but in this comment Iâm speaking only for myself.
Thanks for writing this, and thanks @mike_mclaren for sending it to me! I wish Iâd happened across it when youâd posted it [1] and could have given more timely feedback!
At a high level I think this is very reasonable, especially sections #1 and #3: itâs very important that you can turn âhere is a collection of highly suspect sequencing readsâ into a well-calibrated and trusted âhere is how concerned we should beâ, and itâs also very important that you can then turn that into âand hereâs what we will doâ. Iâve seen scattered work on both of these, but I think these are key areas for additional work.
On section #2, which is closer to what Iâve been thinking about, there were some places where I wanted to push back or add additional information. A few point-by-point replies:
While there is a lot of spending on surveillance, this is overwhelmingly allocated to tracking known threats: how prevalent are SARS-CoV-2, Influenza, RSV, etc. Iâm not aware of anyone who has deployed a system capable of alerting on a novel pathogen. While (in my work at the NAO) Iâd like to change this, I do still think itâs quite neglected for now.
This is definitely a real drawback. On the other hand, once your wastewater-based initial-detection system flags a pathogen you have what you need to create a cheap PCR-based diagnostic, which you can then deploy widely to identify infected individuals (to see how theyâre doing, if they need treatment, and understand the effect the pathogen is having in their bodies, and also to limit spread). This could let you identify many more infected individuals than if you plan was for your initial-detection system to also double as your identification-of-individuals system. But the need for follow-up work ties into the rest of your post, which is that initial detection is only one piece, and if you donât have follow-ups ready to go youâll lose many valuable doubling periods fumbling the next steps.
Strong agree!
While I agree that daily screening of 1% of the UKâs population would be very expensive, I think the main cost (unless youâre doing wastewater) is in collecting the samples. Thatâs just a massive endeavor! If you get sampling down to $1 person itâs still $250M/ây (68M people Ă 1% Ă 365) in sample collection alone. This range might be possible if you somehow just made it part of the culture (ex: every school child provides a sample every morning) but would be a huge lift.
But I think your estimates for the cost are otherwise too high. The main thing is that if you do pooled testing you can cost-effectively sample hundreds of thousands of people at once. This isnât âmultiplexingâ because youâre not adding barcodes (or something else) that would allow you to map sequencing reads back to specific sampled people. For example, say youâre doing pooled nasal swabs, and youâre sampling a large enough number of people that youâll likely get some high viral load folks. Then the average contribution of a sick person is a relative abundance of maybe 2% (rough analysis; @slg and @Will Bradshaw should have a much more thorough treatment of this out any day now). If you run a NovaSeq X 25B (which gets you 25B sequencing read pairs) on a massively pooled sample and 1:100,000 people are currently shedding, youâll get about 5k (2% á 10,000 Ă 25B) read pairs matching the virus, which is likely (again, hoping to get something out publicly about this soon!) enough to flag it. The list price for the flow cell for a 25B sequencing run is $16k, which is perhaps half the all-in cost. Which means if you did daily NovaSeq X 25B runs youâd be spending $10-$20M/ây ($16k Ă 2 Ă 365) on the sequencing.
This is all very ballpark, but mostly is to say that if youâre talking about directly sampling that many people the main cost is in the collection.
This is quite different if youâre using wastewater to avoid needing to sample that many people, but then your relative abundances are far lower and you need much deeper sequencing (expensive!) to get anywhere close to detection at 1:100,000.
I was briefly quite excited about this direction, but a major issue here is that these populations are, because of their vulnerability, somewhat epidemiologically isolated: theyâre probably not close to the first people getting sick in anywhere.
That seems too strong, at least in theory. We should evaluate the marginal detection effort in terms of how much it reduces risk for the cost. For example, a detection system that was very cheap to deploy and only flagged a pandemic 30% of the time could be well worth deploying even though it most likely doesnât help. And perhaps scaling it up to 90% probability of detection would remain cost effective, but getting it up to 99% would be massively more expensive for diminishing returns and so doesnât work out. On the other hand, political considerations might come into play: a system that on average doesnât work could be a hard sell even if itâs relatively cheap.
I also donât expect work focused on initial-detection to help much outside of stealth pandemics. I see my work at the NAO as almost entirely valuable in proportion to the likelihood that someone creates (or perhaps would create, if not deterred by the likelihood of it being flagged) a stealth pandemic.
That said, there is one other situation Iâve seen bounced around where I think this kind of monitoring could be valuable outside of stealth scenarios: when we know thereâs something happening in a region inside a country we have very poor relations with. In that case if we had a system of ongoing monitoring we could pivot it to populations enriched for contact with that region and (ideally) ramp it up.
[1] A few weeks ago I happened to go through and changed my settings to give a big boost to any biosecurity-tagged posts, so I think if this came through again Iâd see it.