This is very insightful, Mike. There is so much exhuberance about what I call “single node innovations” that get excited about one step but forget everything else in the system and lose sight of the big picture. Thank you!
Hi Mike. I think the economic argument is that if it costs, say, $30m* to deliver one single Development Candidate (DC) and you need 20 of these to go through the Development process to generate 1 NME (i.e. assuming a 5% success rate) then Discovery is costing $600m per NME. If AI is genuinely helping to reduce the $30m/DC cost then it can, in theory, have a meaningful effect on the cost of the end-to-end process. But I’m not sure we know if that is genuinely the case yet.
Thanks James - I can see the maths on that, but my challenge is that figures like the 5% hide so many poor decisions, and are so poorly calculated (for example, if it were 3% or 7%, would anyone know, within one pipeline?)... Plus, the real cost of Development doesn't sit in the $600m per NME, but in the late stage 'wrong' decisions... If I could see an effect from 'better' DCs later in lifecycle, I'd buy in a little harder...
Hi Mike. No doubt the 5% hides a lot of poor decisions, with a large part of it being driven by the low predictive validity of the preclinical translational models (as Jack Scannell has demonstrated). It will be fascinating to see if Discovery AI companies can get any sort of edge in this regard (see, for example, Exscientia’s F-1 filing where they claim their approach will ultimately improve PoS), although it will take many years before we know. In terms of the calculation methodology of PoS, have you looked at Andrew Lo’s work as his approach is arguably purer?
Absolutely (Jack is part of my team at Protodigm 🙏🏻)… The problem with all of the PoS methods is the bundling - as I wrote on path dependence, there’s a risk that overestimates early in process get rolled into false confidence later… I see huge opportunity for AI, but I’m not sure that it is being pointed at the right problem yet
This is very insightful, Mike. There is so much exhuberance about what I call “single node innovations” that get excited about one step but forget everything else in the system and lose sight of the big picture. Thank you!
Hi Mike. I think the economic argument is that if it costs, say, $30m* to deliver one single Development Candidate (DC) and you need 20 of these to go through the Development process to generate 1 NME (i.e. assuming a 5% success rate) then Discovery is costing $600m per NME. If AI is genuinely helping to reduce the $30m/DC cost then it can, in theory, have a meaningful effect on the cost of the end-to-end process. But I’m not sure we know if that is genuinely the case yet.
Thanks James - I can see the maths on that, but my challenge is that figures like the 5% hide so many poor decisions, and are so poorly calculated (for example, if it were 3% or 7%, would anyone know, within one pipeline?)... Plus, the real cost of Development doesn't sit in the $600m per NME, but in the late stage 'wrong' decisions... If I could see an effect from 'better' DCs later in lifecycle, I'd buy in a little harder...
Hi Mike. No doubt the 5% hides a lot of poor decisions, with a large part of it being driven by the low predictive validity of the preclinical translational models (as Jack Scannell has demonstrated). It will be fascinating to see if Discovery AI companies can get any sort of edge in this regard (see, for example, Exscientia’s F-1 filing where they claim their approach will ultimately improve PoS), although it will take many years before we know. In terms of the calculation methodology of PoS, have you looked at Andrew Lo’s work as his approach is arguably purer?
Absolutely (Jack is part of my team at Protodigm 🙏🏻)… The problem with all of the PoS methods is the bundling - as I wrote on path dependence, there’s a risk that overestimates early in process get rolled into false confidence later… I see huge opportunity for AI, but I’m not sure that it is being pointed at the right problem yet
*this is not a precise figure, but one based on some reasonable data I can explain privately if interested