this post was submitted on 24 Sep 2026
1115 points (95.9% liked)
Fuck AI
8326 readers
1134 users here now
"We did it, Patrick! We made a technological breakthrough!"
A place for all those who loathe AI to discuss things, post articles, and ridicule the AI hype. Proud supporter of working people. And proud booer of SXSW 2024.
AI, in this case, refers to LLMs, GPT technology, and anything listed as "AI" meant to increase market valuations.
founded 2 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
Protein structure prediction models have contributed a lot. They are delivering much better results than the previous state of the art method of homology modeling.
Predicting structure from sequence in large scale has been a huge problem before and those attention network based models enabled in-silico screenings that were simply not possible before. This had a huge impact on pharmacologic research and especially research on binding motifs.
That said, these models work best when target proteins are one among many and don't have very unique features. And things go downhill real fast when proteins form oligomeric structures or rely on structural features not yet resolved by empirical methods.
AF models oligomers just fine.
Not all proteins are structured.
Models are just models, cartoons until they are experimentally validated.
AI will generate all kinds of pathways and concepts and chemicals that are pointless without real experimental data. Done by real scientists. This is what Trump killed when he defunded the NCI via the NIH, from 1,100 funding announcements in 2024 to 11 in 2025. Apparently Americans don't care, so that funding may never be restored, but even if it were restored, the post doc fellows who do the work have left and are not coming back, labs have closed down and will not re-open. This was a pipeline of billions that fed a US pharma industry worth trillions, now quickly relocating to China.
Ad1, yeah, maybe your proteins and at superficial investigation. But my experience with the oligomeric interfaces, when having structures that are not publicly accessible and therefore can't be part of the training, was not so great.
Ad2, I am not talking about unstructured proteins. I am talking about structural features that have little to no precedence in public databases.
Ad3 Most models will be never validated by experimental structure elucidation. That's the whole point of those models. To be able to get some information without having to do an undoable amoun high effort experiments. Sure, the end result will be verified but model quality matters, it is the factor that determines the success rate of the in-silico part of the experiments, ie how many candidates have to be screened on average to find sufficiently improved candidates.
Attention network based models are fine but they are not the golden calf that some are selling it either. You put it yourself, models are just models. Quality is not homogoenous though. Like I said, models for monomers are generally more credible.