For those interested, it appears as though David Baker (who dedicated his life to protein folding) has also turned to deep learning. His lab recently published https://www.biorxiv.org/content/10.1101/846279v1, which seems to outperform Alphafold with a very concise architecture. Code and model is at https://github.com/gjoni/trRosetta
Do you know if he turned to deep learning after AlphaFold's stunning performance at CASP13 in 2018? I haven't heard anything from that particular niche since the "AlphaFold @ CASP13: “What just happened?”" blog post: https://moalquraishi.wordpress.com/2018/12/09/alphafold-casp...
The author of the article you linked also has this repo, to welcome the public to start training https://github.com/aqlaboratory/proteinnet , following in the same veins as Imagenet
It absolutely did. Both AlphaFold and RaptorX (another deep learning approach from Jinbo Xu's lab) were mentioned in that paper as the main motivation of their architecture
A colleague of mine in chemistry gave his thoughts on AI for protein folding recently: “things keep getting better, but they’re nowhere close to being good”.
I think part of the issue at play here is the cost of confirming success, not simply the cost of generating a plausible solution. In most domains of AI that have shown success, the cost of confirmation is trivial (look at the image and check the label) whereas the cost of generating a plausible solution was high.
Like many AI fields, I believe that the real breakthrough will not be a direct approach, but an approach the solves the most pressing barrier to using AI in the first place.
From my perspective, in the life-sciences space, it’s certainly the lack of high quality data. To be more specific, there’s tons of data, but it’s sloppy, lossy, and non standardized.
Many of those discussing the promise of near future AI are either academics or new to the field. The real data landscape is so much worse than they would imagine it is.
The cost of confirming success. If we can decrease the cost of going from theoretical protein to physical test, then I think the AI part will be quickly solved.
I think the point here is slightly missed. Protein folding simulations are accurate ok hydro all models where brute force is used to explore the parameter space. The AI models could narrow that space and then employ the stricter models. Therefore I don’t think the cost here is great.
A bit off topic, but what the hell does the word "Alpha" mean in their marketing? Are they just going to call every AI they create from now on "Alpha_"
AlphaStar didn't really use any techniques from AlphaGo either, so the answer is 'yes'. Anything sufficiently whizzy will get the 'Alpha' prefix for branding.
