Computational biologist, data scientist, occasional blogger | https://blog.genesmindsmachines.com/ | https://clauswilke.com/ | Opinions are my own and do not represent UT Austin.
Claus Wilke
Loading...
I'm in favor of limits, but I think they should be imposed on dollar amounts, not on number of grants. Why should somebody with three small grants get penalized over somebody else with one large one?
Claus Wilke
Texas wildflowers.
The fundamental problem is that zero-shot predictions know nothing about the phenotype of interest, and different phenotypes don't have to correlate. Therefore, zero-shot prediction has a fundamental problem that no model improvements can overcome. 2/
When ranking models by zero-shot prediction performance, we see that most models perform roughly the same and there is much more variation within models among datasets than among models. Also, the phenotype assessed matters for the ranking. 3/
The actual challenge in slide preparation is how to structure the story and what evidence to use to support it. AI cannot help with either. We're failing students by not teaching what matters in slide preparation.
Overall, models learn some type of generalized survival landscape (which mutations kill the protein, which do not), but they can't do much beyond this. Large-scale benchmarks obscure the fact that model predictions are often poor for specific tasks, such as picking beneficial mutations. 5/5
New paper about zero-shot predictions of mutational effects. There's a lot of misunderstanding about what such predictions can and cannot do. In brief, they're good at identifying deleterious mutations but bad at predicting fitness improvements. 1/
www.biorxiv.org/content/10.6...
Here's a starting point for how to design effective slides. It's a two-for-one: not only is the template format I describe incredibly effective, it's also super easy to make slides in this format. I normally only need a minute or two for each slide.
blog.genesmindsmachines.com/p/slides-tha...
Model performance is usually assessed by Spearman correlations. Those correlations tend to be quite good when considering all mutations, but tend towards zero for the fittest mutations. Again, models aren't good at picking beneficial mutations. 4/
Can somebody please do this for a seahorse emoji so LLMs can finally rest?
