Building a causal model of gene regulation is an infinitely harder problem that will take decades of efforts with new data modalities, well thought out experimental designs & active iterative feedback loops between data generation to push models out of zones they struggle in. 4/
Perturbation prediction with well defined goals & scope is a valuable & valid task. Should be solved by hook or by crook - whatever works 3/
Some users welcome Kundaje's pragmatic realignment of gene perturbation prediction goals as productive, while others criticize hype around parameter scaling and urge more honest experimental design over pattern matching.
Perturbation prediction with well defined goals & scope is a valuable & valid task. Should be solved by hook or by crook - whatever works 3/
Also will once again iterate that one can develop simple approaches that can predict perturbations that are in distribution very well without ever learning a causal model of gene regulation. 2/
We need to continue to be honest about what our models are actually capable of and where they fail so we keep moving forward. Less hype & blind faith in parameters & arbitrary scaling. More evidence based strategies & honest reflection. 6/6
The earlier we come to this conclusion the more productive our paths will be. It's great to see some realignment & recalibration of tasks, goals & critical thinking. 5/
Building a causal model of gene regulation is an infinitely harder problem that will take decades of efforts with new data modalities, well thought out experimental designs & active iterative feedback loops between data generation to push models out of zones they struggle in. 4/
@anshulkundaje We need experimental design driving the question, not parameter count. And let's be real, biotech keeps stumbling over the same gap between prediction and causal insight.
@anshulkundaje And let's be real, predicting in-distribution perturbations isn't biology. We celebrate pattern matching and call it mechanism. Patients deserve better than that.
@anshulkundaje We need decades of iterative work to build causal models of gene regulation. And let's be real, funding for that kind of work is still woefully inadequate. Good to see rigor prioritized.
Building a causal model of gene regulation is an infinitely harder problem that will take decades of efforts with new data modalities, well thought out experimental designs & active iterative feedback loops between data generation to push models out of zones they struggle in. 4/
Perturbation prediction with well defined goals & scope is a valuable & valid task. Should be solved by hook or by crook - whatever works 3/