Open venkata-ramaswamy-cresset opened 2 months ago
Do you have any recommendations for the minimal resources required for neuralplexer in terms of CPU, RAM, GPU vs. size of input protein?
5.8Gb total capacity is certainly below the GPU RAM we have tested on. As a minimum we recommend testing on a system with at least 16GB CPU and GPU memory.
Thank you @zrqiao but even with 32 GB RAM, I could not generate the models for this example case and it returned FileNotFoundError for prot_all.pdb that has to be generated.
Command: neuralplexer-inference --task=batched_structure_sampling --input-receptor ./4dx5_ed.pdb --input-ligand MIY_model_ed.sdf --out-path ./testing_4dx5_MIY --model-checkpoint ./complex_structure_prediction.ckpt --n-samples 1 --chunk-size 1 --num-steps=1 --sampler=langevin_simulated_annealing
Output ` Block contact sampling: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 31/31 [01:06<00:00, 2.14s/it] Structure generation using langevin_simulated_annealing: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:21<00:00, 21.96s/it] /home/venkata/miniconda3/envs/neuralplexer_dev/lib/python3.10/site-packages/pytorch3d/ops/points_alignment.py:340: UserWarning: Excessively low rank of cross-correlation between aligned point clouds. corresponding_points_alignment cannot return a unique rotation. warnings.warn(
neuralplexer-inference 33
inference.py 672 main multi_pose_sampling(
inference.py 257 multi_pose_sampling write_pdb_models(
pipeline.py 971 write_pdb_models with open(out_path, "w") as of:
FileNotFoundError: 2 No such file or directory testing_4dx5_MIY/prot_all.pdb `
Any pointers on why I get this FileNotFoundError that Neuralplexer has to generate? It worked for smaller proteins but not this one.
These are the input files (added .txt extension to upload here) if needed to reproduce the error: 4dx5_ed.pdb.txt MIY_model_ed.sdf.txt
Description
I am trying to generate a protein-ligand complex for a small protein (536 residues) and got
torch.cuda.OutOfMemoryError
with--cuda
. To check if it works with CPUs only, I ran the same command without--cuda
and it worked fine. Further, if I reduce the chunk size and samples generated hoping that would reduce the memory consumption, (--n-samples 16 --chunk-size 4 --num-steps=40
to--n-samples 1 --chunk-size 1 --num-steps=1
), I don't gettorch.cuda.OutOfMemoryError
issue and the job runs successfully with--cuda
.Could you please advice what I could do to fix this issue.
Also, as another example, I tried building a protein-ligand complex for a bigger protein (~3000 residues) and NeuralPLexer crashed complaining about memory both with and without
--cuda
.Just in case it helps, my workstation has an "NVIDIA GeForce GTX 1660 Ti" GPU (6144 MB) card, 24 CPU cores and a RAM of 32 GB.
What I Did
The command I used for generating the complex for the small protein (536 residues):
The ligand sdf (txt extension added so I can upload here) with no coordinate info: 41X_model_ed.sdf.txt
I got this error:
The command for modelling the bigger protein:
The relevant input files are: MIY_model_ed.sdf.txt 4dx5_ed.pdb.txt
The output I got