> yet it crashed spectacularly in Phase III clinical trials.
I don't think this necessarily means the approach is flawed: my understanding is that the value these companies/models try to deliver is that it brings down the search space of molecules you test in your funnel by several order of magnitude.
Some will still fail, but as long as you do better than how things used to be done before, you bring value.
Maybe it will just take one blockbuster success for Pharma companies to decide this approach is worthwhile.
> ... the value these companies/models try to deliver is that it brings down the search space of molecules you test in your funnel by several order of magnitude.
That’s the pitch, yes—but it rests on a crucial assumption: that the properties those models optimize (binding energy, predicted off-targets, etc.) are strongly correlated with what pharma ultimately cares about—clinical success. So far, that link is weak. That's great that I got 100 candidates instead of 1 or two, but if I have no way of knowing which one will be a better candidate for pre-clinical and clinical testing (which costs much more than the discovery stage, and hence, can't be done for more than a handful of molecules) the problem is only very partly solved.
Machine-learning screens are great at ranking compounds for tight binding or flagging obvious off-target liabilities, but most Phase II/III failures stem from issues the models don’t yet capture: PK/PD quirks, formulation hurdles, human safety signals, or simply the complex biology of disease. Narrowing the search space upstream is helpful, but it doesn’t tackle the real bottleneck—turning a “perfect” binder into a safe, effective therapy in people. Until the models address those downstream risks, the overall attrition curve won’t shift as much as the marketing decks suggest.
I wonder if this also applies to the new wave of materials science companies that have a somewhat similar model, i.e. come up with some material that has useful properties. See e.g. CuspAI and Dunia. CuspAI seems to be getting their data from partners, while Dunia is building their own lab.
> yet it crashed spectacularly in Phase III clinical trials.
I don't think this necessarily means the approach is flawed: my understanding is that the value these companies/models try to deliver is that it brings down the search space of molecules you test in your funnel by several order of magnitude.
Some will still fail, but as long as you do better than how things used to be done before, you bring value.
Maybe it will just take one blockbuster success for Pharma companies to decide this approach is worthwhile.
> ... the value these companies/models try to deliver is that it brings down the search space of molecules you test in your funnel by several order of magnitude.
That’s the pitch, yes—but it rests on a crucial assumption: that the properties those models optimize (binding energy, predicted off-targets, etc.) are strongly correlated with what pharma ultimately cares about—clinical success. So far, that link is weak. That's great that I got 100 candidates instead of 1 or two, but if I have no way of knowing which one will be a better candidate for pre-clinical and clinical testing (which costs much more than the discovery stage, and hence, can't be done for more than a handful of molecules) the problem is only very partly solved.
Machine-learning screens are great at ranking compounds for tight binding or flagging obvious off-target liabilities, but most Phase II/III failures stem from issues the models don’t yet capture: PK/PD quirks, formulation hurdles, human safety signals, or simply the complex biology of disease. Narrowing the search space upstream is helpful, but it doesn’t tackle the real bottleneck—turning a “perfect” binder into a safe, effective therapy in people. Until the models address those downstream risks, the overall attrition curve won’t shift as much as the marketing decks suggest.
Fair point!
I'm curious to see how this business model plays out.
Do you know of anyone trying to predict the later stage success/efficacy of molecules?
I wonder if this also applies to the new wave of materials science companies that have a somewhat similar model, i.e. come up with some material that has useful properties. See e.g. CuspAI and Dunia. CuspAI seems to be getting their data from partners, while Dunia is building their own lab.