Hi, If you use GPU to accelerate predict process, what is the acceleration effect? I used Tesla T4, tensorflow-1,14-gpu，tested the consumption time of sess.run() in predict.py. The acceleration effect is only 1.5 times that of the CPU. Is this normal? How many times the acceleration should be achieved? Looking forward to your reply, thanks.

I want to test the performance of trRosetta on a bunch of proteins I have, I got the .npz files from predict.py but when I tried to use trRosetta to output the predicted structure I got following error: PyRosetta-4 2020 [Rosetta PyRosetta4.Release.python36.ubuntu 2020.50+release.1295438cd4bd2be39c9dbbfab8db669ab62415ab 2020-12-12T00:30:01] retrieved from: http://www.pyrosetta.org (C) Copyright Rosetta Commons Member Institutions. Created in JHU by Sergey Lyskov and PyRosetta Team. temp folder: /dev/shm/64q8sw5m dist restraints: 8485 omega restraints: 8437 theta restraints: 17282 phi restraints: 17292 mutation: G26A mutation: G73A mutation: G81A mutation: G89A mutation: G95A mutation: G117A mutation: G150A mutation: G151A

ERROR: Pose::set_phi( Size const seqpos , Real const setting ): residue seqpos is not part of a protein, peptoid, or carbohydrate! ERROR:: Exit from: /home/benchmark/rosetta/source/src/core/pose/Pose.cc line: 1069 Traceback (most recent call last): File "/home/zyxue/Record/trRosetta/trRosetta.py", line 212, in main() File "/home/zyxue/Record/trRosetta/trRosetta.py", line 110, in main set_random_dihedral(pose) File "/home/zyxue/Record/trRosetta/utils_ros.py", line 152, in set_random_dihedral pose.set_phi(i,phi) RuntimeError:

File: /home/benchmark/rosetta/source/src/core/pose/Pose.cc:1069 [ ERROR ] UtilityExitException ERROR: Pose::set_phi( Size const seqpos , Real const setting ): residue seqpos is not part of a protein, peptoid, or carbohydrate!

It seems like there's some residues that are not protein but that's impossible since the sequence I used was directly generated from pdb structure and the sequence alignment and first step prediction of constraints went well, I wonder what does this error mean and if there's a way to fix it.

Hi, Ivan Can you relax the training code which contains data processing such as subsample and 300 residues limit? This confuse me a long time. Thanks!

I'm trying to run trRosetta on about 3000 different proteins. For each protein I have generated an MSA file with HHblits.

For the first few proteins everything goes smoothly, and it predicts the distogram/angles in about 1-3 minutes. However, at some point it tries to predict on one of the larger MSA's. The particular protein that it stalls on has a length of 1600 amino acids, and roughly 3500 aligned sequences.

So far Rosetta has been running for about 1 hour on this sequence, and has printed the following:

2020-08-06 15:38:16.462083: W tensorflow/core/framework/allocator.cc:124] Allocation of 4498921476 exceeds 10% of system memory.
2020-08-06 15:38:24.750222: W tensorflow/core/framework/allocator.cc:124] Allocation of 4498921476 exceeds 10% of system memory.
2020-08-06 16:46:23.006386: W tensorflow/core/framework/allocator.cc:124] Allocation of 4498921476 exceeds 10% of system memory.
2020-08-06 16:46:23.006592: W tensorflow/core/framework/allocator.cc:124] Allocation of 4080654400 exceeds 10% of system memory.
2020-08-06 16:46:30.095595: W tensorflow/core/framework/allocator.cc:124] Allocation of 4080654400 exceeds 10% of system memory.
./data/model2019_07/model.xaa - done

I'm guessing this is some sort of issue where the whole thing isn't loaded into memory, but rather ends up in the swap file or something? which would probably explain why it ends up taking such an awful long time. However I have 32 GB of ram on this machine and it isn't really using more 11 GB for this, and furthermore everything seems to be done on the CPU as opposed to the GPU, is that normal for this code?

Finally, one thing I have been wondering about, it seems like the model is loaded in everytime as the code is running right now, which takes quite a while and seems like a waste, I would imagine the model should only need to be loaded in once, and then just run on each of the MSA's to predict the distograms/angles, however it seems like the model being loaded depends on the specific MSA used as input, which would make it less trivial I guess?

Hi Ivan,

Thanks for sharing the source code, it's a great job.

For the distance histogram prediction, as the paper says: "The distance range (2 to 20 Å) is binned into 36 equally spaced segments, 0.5 Å each, plus one bin indicating that residues are not in contact." So the output shape of the network is LL37, and the last bin corresponds to the >20 Å distance.

But when converting the distance distribution to energy potential, what the article says is "For the distance distribution, the probability value for the last bin, i.e., [19.5, 20], is used as a reference state ... ", which is contradictory to the definition of the last bin (>20 Å but not [19.5, 20]).

Is it my misunderstanding or is there something wrong?

Thank you for providing this excellent resource!

Can I ask a question about a minor detail: how did you define d, omega, theta and phi when one or both of the residues in a pair is a glycine? What is the network predicting in these cases?

I could not find this in your paper or supplementary info - sorry if I have missed it!

Many thanks!

Hi there.

I want generate structure models from trRosetta. So, I typed below. python trRosetta/trRosetta.py trRosetta/example/T1008.npz trRosetta/example/T1008.fasta model.pdb

But, the output is it

temp folder: /dev/shm/3rsb8fq5 dist restraints: 2877 omega restraints: 2872 theta restraints: 5835 OSError: [Errno 28] No space left on device

During handling of the above exception, another exception occurred:

Traceback (most recent call last): File "trRosetta/trRosetta.py", line 212, in main() File "trRosetta/trRosetta.py", line 58, in main rst = gen_rst(npz,tmpdir,params) File "/tf/trRosetta/utils_ros.py", line 137, in gen_rst f.close() OSError: [Errno 28] No space left on device

my shm folder has inode=7.9M, IUse%=1%, size=64M How can I fix it?

Hi --

I'm interested in understanding the computational profile of these kinds of structure prediction tasks.

As I understand it, there are three broad steps to go from input to 3d structure: a) "feature engineering" via MSAs, etc to form LxLx526 input tensor b) forward pass through the CNN to predict distances/angles c) structure determination to realize an actual 3d structure

Are you able to comment on the relative cost in terms of wall-clock time / compute required / etc of those three steps? Is one step substantially more costly than the others? Do you have a sense of how long it would take to go from raw input -> 3d structure for a single example?

(I'm primarily concerned w/ inference-time costs here -- I see in your paper that training the CNN takes ~ 9 days on a single GPU.)

Thanks!

Hello!

I have a naive question. I don't really understand the following process in the README file:

Using the generated MSA, predict the distance and orientations by running the scripts at: https://github.com/gjoni/trRosetta

I have an MSA in a3m format. Which scripts shall I use to predict distance and orientations ?

Thanks in advance for any help.

Best,

Anupam

Hi there,

When I tried to run the predict.py I get the following error.

File "/mnt/vdf/pepBuilder/trRosetta/trRosetta/network/utils.py", line 22, in parse_a3m msa = np.array([list(s) for s in seqs], dtype='|S1').view(np.uint8) ValueError: setting an array element with a sequence

Should this line 22 be changed like the following

msa = np.array([list(seqs) for s in seqs], dtype='|S1').view(np.uint8)

Do you have an explanation for it.

Thanks in advance.

Kapila