Google Deepmind Solves All Genetics Problems

It is game over for geneticists and genome scientists. All genetics problems are solved.

The Brawndo (Alphafold) guys, who got Nobel prize for accurately predicting the structure of every single protein are back again. This time they solved all genetics problems !!

Nature reports - (DeepMind’s new genome ‘atlas’ charts effects of all nine billion human gene mutations) -

AI model AlphaGenome forecasts the consequences of altering every single DNA letter in the human genome.

Yes, you read that right. Now we know the consequences of altering every single DNA letter in the human genome, thanks to the Brawndo Nobel laureates. I am not eager to spend $32.99 to read the rest of the Nature article.

Not so fast, says Johns Hopkins professor Steven Salzberg -

I have some thoughts about this: a thread. “DeepMind’s new genome ‘atlas’ charts effects of all nine billion human gene mutations” https://nature.com/articles/d41586-026-02835-4 TL;DR: AI slop is taking over genomics 1/N

For each of these 9 billion changes, they report its “impact” as something they called AVI, for AlphaGenome Variant Impact.” What does this even mean? Well, I’ve spent a few hours now with the paper and … 3/N

As far as I can tell, the AVI score is some kind of estimate of how “important” a DNA base is. But there are some gargantuan problems here. First, there isn’t enough biological data in the universe to train a model on these 9 billion mutations. Most have never been observed 4/N

… and even for the mutations that have been observed (a small fraction of the 9 billion), we only have imprecise biological measurements of their impact. So the claim that these AVI scores are useful is wildly implausible on its face. But I read the paper… 5/N

… trying to find something that would convince me that at least a subset of the scores might be useful. Instead, I found a enormous number of claims, mostly supported by comparing AlphaGenome to other computational methods, usually with very thin or no explanation 6/N

But how about some actual biology? It’s there, but only in an anecdote about how they helped to resolve a rare disease case. Hmm, what about that? Well, first they ranked variants in 814 “unsolved” cases. (No explanation of how these variants were initially collected.) 7/N

… the top-ranked AVI score for one patient was in an intron of a gene called DNM1. Turns out this variant was reported in 2 previously published cases. They did some cell-line expts to show it changes splicing in DNM1. So it’s plausible! But there’s so much missing here 8/N

But back to the main result: the 9 billion AVI scores. What the heck does one do with this? They didn’t prove that any of them are meaningful (okay, maybe 0.0001% of them), so this is just more AI slop to me, because look… 10/N