First we did not "fail" in drug development, our collaboration with GSK yielded substantially more programs than we had anticipated when we started. The verdict on whether the strategy is successful or not will not be clear for another 5-10 years because being better at picking targets doesn't shorten the timeline for bringing a drug to market, it mostly means that you should have a modestly higher proportion of programs that end up being successful.
Second we did not start with a premise that we could or needed to do a great job of predicting disease from genetics alone, that is not required to identify good targets. You're correct we don't have huge genomics datasets, we largely use the same genomics datasets others use. Identifying targets with genetics in the simplest sense requires using genetics to identify an association at a genomic location, and then using genomics for functional interpretation of that association. And having the largest database of genetic associations enables the first step of that.
I'm not sure what you mean by "the underlying relationship between the genome and phenotypes... is a highly nonlinear function". Simple additive models account for most of the heritability of complex phenotypes.
I was a team from Google that evaluated 23&Me's data and technology many years ago for a business deal. We already talked about this with Anne and she confirmed what I said above. It might have been before you were hired- but I'm pretty sure we talked about building a variant store and a transpose service? I stand by my statements (note: I work at a competitor of GSK and I know all about these deals). It's not correct that being better at picking targets doesn't shorten the timeline, either- at least in the opinion of the scientists at my company.
Additive models don't really account for most of the heritability of complex phenotypes. They're what have worked best and been published so far. Complex phenotypes are nonlinear because the generative processes in biology have feedback, homeostasis, enormous numbers of individual elements... etc...
By the way, does 23&Me still put tongue-twisting as a heritable trait on its reports? https://blog.23andme.com/articles/tackling-tongue-curling-th... It's not.
You may be correct that better targets are also be faster to market. We would also hope so: better targets might be more straightforward to validate in the lab and may enable smaller trial sizes, for instance. Timelines are still long and this doesn't alter my statement that the success or failure of our target selection strategy won't be known for some years.
Do you have any evidence for additivity not accounting for most heritability? I'll give one cite (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2265475/). The fact that biology is very complex and non-linear is not inconsistent with most genetic effects being small and approximately additive.
We have never had a trait report on tongue curl, the 15-year-old blog post you linked to says it isn't a Mendelian trait, so I'm not sure where you were going with that. I can tell you that it is somewhat heritable but complex.
Environment is important but we could still have huge improvements in medical care using genomics. It's easy to obtain and still has a very strong relationship to disease, and is a problem best solved by deep learning. I've watched people tilt against this windmill for 25+ years and it's kind of funny just how bad our labelling of diseases is.