Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

End-to-end differentiable learning of protein structure

biorxiv.org

Post reply on HN
Print view View original

21–28 of 28 posts

Re: End-to-end differentiable learning of protein structure

#21

Post by rytill »

RivieraKid wrote: post #12
cing wrote: post #6

Earlier quoted context omitted.

DeepMind and others are trying. "Hassabis said the company is now planning to apply an algorithm based on AlphaGo Zero to other domains with real-world applications, starting with protein folding." [1] https://www.bloomberg.com/news/articles/2017-10-18/deepmind-...

That doesn't make any sense unless I'm missing something, A0 is suited for a completely different problem than protein folding...

The AlphaZero algorithm (monte carlo tree search with value estimator trained by reinforcement learning) works on any environment you can simulate during play time, single player or not.

Re: End-to-end differentiable learning of protein structure

#22

Post by eutectic »

rytill wrote: post #21
RivieraKid wrote: post #12

Earlier quoted context omitted.

That doesn't make any sense unless I'm missing something, A0 is suited for a completely different problem than protein folding...

The AlphaZero algorithm (monte carlo tree search with value estimator trained by reinforcement learning) works on any environment you can simulate during play time, single player or not.

Any environment with finite action and state-spaces.

Re: End-to-end differentiable learning of protein structure

#25

Post by resiros »

superfx wrote: post #24
resiros wrote: post #23

Cool method! Are you planning to participate in the next CASP? Do you plan to open source the code?

Yes! Certainly on the source code, and hopefully on CASP13 too.

Thanks for the answer! I hope then to see you in CASP (and CAMEO too, it is a great tool to test/refine your method). I was discussing a paper with a co-worker of mine (we also work on psp, we work on RBO Aleph). We had a hard time pinpointing the thing that made your method finally work. You have mentioned in your blog post that you have been working on it for years now, and I guess a lot of other people had the idea of using deep learning for psp. But what was the insight that made it all work, using LSTM? or was it many small refinements and hacks?

Re: End-to-end differentiable learning of protein structure

#26

Post by superfx »

resiros wrote: post #25
superfx wrote: post #24

Earlier quoted context omitted.

Yes! Certainly on the source code, and hopefully on CASP13 too.

Thanks for the answer! I hope then to see you in CASP (and CAMEO too, it is a great tool to test/refine your method). I was discussing a paper with a co-worker of mine (we also work on psp, we work on RBO Aleph). We had a hard time pinpointing the thing that made your method finally work. You have mentioned in your blog post that you have been working on it for years now, and I guess a lot of other people had the ide…

I would say the biggest thing is obviously the architecture, coupling LSTMs with the geometric units that spit out the actual 3D structure that can then be directly optimized via the dRMSD loss function. That's the biggest point of distinction from everything else out there (no contact map prediction, etc.) So it really is about end-to-end differentiability IMO, which hasn't been done before.

As for why it took so long, it is and it is not fine-tuning. Getting RGNs to train _at all_ was a rather difficult process, and required a lot of finicking around. But since I got them working, I haven't actually spent all that much time fine-tuning them, and so I expect there to be a lot of low-hanging fruit in terms of optimizing performance (starting from the baseline I found.)

Re: End-to-end differentiable learning of protein structure

#27

Post by mathperson »

superfx wrote: post #8
mathperson wrote: post #7

Earlier quoted context omitted.

lol what-I was looking at this list of people who do this-in fact a lot of them ARE machine learning researchers...including some in my department!

I do think however that protein folding is very much understudied in the ML community, relative to say the big three of vision, NLP, and speech. The lack of standardized data sets and benchmarks, not to mention the need for domain knowledge, have made it difficult to get into the field

at the risk of offending NLPers/Vision/Speech I just think those tasks are 'easier' in a variety of ways.

Re: End-to-end differentiable learning of protein structure

#28

Post by rytill »

eutectic wrote: post #22
rytill wrote: post #21

Earlier quoted context omitted.

The AlphaZero algorithm (monte carlo tree search with value estimator trained by reinforcement learning) works on any environment you can simulate during play time, single player or not.

Any environment with finite action and state-spaces.

No, the key requirement which makes it difficult to use on real-world tasks is that you must be able to do a forward rollout of your environment in your decision-making process.
Post reply on HN