131–140 of 170 posts
This isn't a criticism - I'm just curious to hear people's thoughts on this. When I look at this code, one of my initial reactions is that it does not seem to be very thoroughly tested. Sure, certain modules have been tested (e.g. `model.quat_affine`) but it's not clear how completely. Meanwhile, other modules, for example `model.folding`, have not been tested at all, despite containing large amounts of complex logic…
Alphafold 2 is very very cool, but we need a little dose of reality. It's still a bit away from really solving protein folding as it was marketed. For example, multi-complex proteins are not well predicted yet and these are really important in many biological processes and drug design: https://occamstypewriter.org/scurry/2020/12/02/no-deepmind-h... A disturbing thing is that the architecture is much less novel than I…
Honest question: since AlphaFold doesn't really _solve_ the protein folding problem (it's NP-complete after all), but only _approximates_ solutions very well, what are the real impacts of this? Isn't a good approximation of a protein enough to cause unexpected problems? How do we know that an approximate structure will perform the same as the correct solution?
I dont know much about protein folding, but for most things in life,exact solutions to NPC problems usually aren't needed for non-contrived problems. In many cases, approximations are good enough. Besides, this is real life - if predictions and real life match, that's great. If they don't, well you know you went wrong somewhere.
Joke, which I think is from Sean Eddy (hammer).
Bioinformatics approaches a Computer Scientist for help with a hard problem. CS agrees to help. Year later CS comes back very excitedly. "your problem is not hard it is NP-hard!". Bioinformatics nods, and says "I still got to solve it" and continues finding ever faster and better approximations ;)
Also problem space is both bounded (you don't have infinite length proteins) and f'd up in reality. e.g protein hijacking and re-conformation in the face of an infectious agent.
Distribution of this 2 TB file seems like a good use of torrent…
On the other hand outbound network traffic from an university is "free". So the benefit is absolutely minimal from a hosting perspective.
It was tried (https://journals.plos.org/plosone/article?id=10.1371/journal...) but it is gone the way of the dodo for the above reasons.
The model parameters are only available for non-commercial use. That's a shame, as I presume there might be a lot of medical startups that would benefit from having this kind protein-folding tech available.
Unless I'm mistaken, you could train the model yourself, starting with a random set of values. In time, your error rates would be low enough to have a new set of parameters which you could use however you like.
If you're a a medical startup, having an off-the-shelf prediction model you can just start using for all your protein folding needs is a very different proposition from having to train one yourself from scratch.
That said, hopefully other researchers and institutions will take Google's research and produce an equivalently powerful model but with a more commercially-friendly open-source license. From some comments in this thread, it sounds like that's already happening, in fact.
Earlier quoted context omitted.
> A disturbing thing is that the architecture is much less novel than I originally thought it would be, so this shows perhaps one of the major difficulties was having the resources to try different things on a massive set of multiple alignments. This is something an industrial lab like DeepMind excels at. Whereas universities tend to suck at anything that requires a directed effort of more than a handful of people. Y…
That's the case with basically everything DeepMind does. They have a very good PR department which hypes up everything they do while conveniently ignoring that basically nothing of any practical consequence has come of their endeavors. But I do think it's important that these companies exist now so we can see what not to try going forward.
Earlier quoted context omitted.
there are research groups this would be useful for but structures are not on the critical path to drug discovery or approval.
Out of (probably overoptimistic :) ) curiosity, what do you see are the critical paths?
Still, even if structure determination is not on the "critical path", it IS a big barrier that has (started) to fall now.
> The AlphaFold parameters are made available for non-commercial use only, under the terms of the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. You can find details at: https://creativecommons.org/licenses/by-nc/4.0/legalcode Does CC BY-NC actually do this? As far as I can tell it only really talks about sharing/reproducing, not using. Or is the only thing prohibiting other comm…
If you took their parameters, then trained it for while on a different set of data, it would vary from the original. I wonder how much compute would be required to make the offset far enough to hold up from scrutiny, and in court. Alternatively, you could manually change the network model, add a few hidden layers, etc... modifying the parameters in step, and result in a new model and new parameters. Some training to…
Does anyone on HN work in bio or drug discovery? Could you give an overview of how people can leverage this (or how you might?). From reading around about it, it sounds like there's often a need to find a certain type of molecule to activate/inhibit another based on shape and the ability to programmatically solve for this makes the searching way easier. Is this too oversimplified/wrong? How will this be used in pract…
> Could you give an overview of how people can leverage this (or how you might?). Short answer: nobody knows. Traditionally, protein folding is a solution in search of a problem, but that's largely because the predictions were...unusably bad. This was always more of a super-difficult validation problem for the force fields and simulation methods, which could then be used for other problems of greater value (such as r…
There are some examples of this issue in the AlphaFold blog, some protein loops that they thought were mispredicted but it turned out they were part of an energy degeneracy so the natural state fluctuated pretty wildly, so if you can't simulate this properly it matters less how accurate the incoming structure is (to a certain degree of course).
This isn't a criticism - I'm just curious to hear people's thoughts on this. When I look at this code, one of my initial reactions is that it does not seem to be very thoroughly tested. Sure, certain modules have been tested (e.g. `model.quat_affine`) but it's not clear how completely. Meanwhile, other modules, for example `model.folding`, have not been tested at all, despite containing large amounts of complex logic…
My understanding is that it has been manually tested. I.e. it has produced correct results to previously intractable problems. I'm not sure how much automated testing would add at that point.