search

Zuricho / ParallelFold

Modified version of Alphafold to divide CPU part (MSA and template searching) and GPU part. This can accelerate Alphafold when predicting multiple structures
https://parafold.sjtu.edu.cn
133 stars 45 forks source link

Is CPU acceleration failed? #26

Open yanchenmochen opened 2 years ago

yanchenmochen commented 2 years ago

Last day, I make some experiments in a Server to run the ./run_alphafold.sh -d /dataset/ -o result -p monomer -m model_2 -i input/T1061.fasta and I read the log, confused, the T1061 is 949AA. ` I0822 07:33:00.806264 140553952322112 jackhmmer.py:133] Launching subprocess "/opt/conda/bin/jackhmmer -o /dev/null -A /tmp/tmpxbrk9wt6/output.sE 0.0001 -E 0.0001 --cpu 8 -N 1 input/T1061.fasta /dataset//uniref90/uniref90.fasta" I0822 07:33:01.157015 140553952322112 utils.py:36] Started Jackhmmer (uniref90.fasta) query I0822 07:37:27.058227 140553952322112 utils.py:40] Finished Jackhmmer (uniref90.fasta) query in 265.901 seconds I0822 07:37:27.072012 140553952322112 jackhmmer.py:133] Launching subprocess "/opt/conda/bin/jackhmmer -o /dev/null -A /tmp/tmpnn6am537/output.sE 0.0001 -E 0.0001 --cpu 8 -N 1 input/T1061.fasta /dataset//mgnify/mgy_clusters_2018_12.fa" I0822 07:37:27.439405 140553952322112 utils.py:36] Started Jackhmmer (mgy_clusters_2018_12.fa) query I0822 07:42:42.192071 140553952322112 utils.py:40] Finished Jackhmmer (mgy_clusters_2018_12.fa) query in 314.752 seconds I0822 07:42:42.364272 140553952322112 hhsearch.py:85] Launching subprocess "/opt/conda/bin/hhsearch -i /tmp/tmpog4q4684/query.a3m -o /tmp/tmpog40/pdb70" I0822 07:42:42.712445 140553952322112 utils.py:36] Started HHsearch query I0822 07:44:18.199999 140553952322112 utils.py:40] Finished HHsearch query in 95.487 seconds I0822 07:44:18.555797 140553952322112 hhblits.py:128] Launching subprocess "/opt/conda/bin/hhblits -i input/T1061.fasta -cpu 4 -oa3m /tmp/tmpz9oq 1000000 -realign_max 100000 -maxfilt 100000 -min_prefilter_hits 1000 -d /dataset//bfd/bfd_metaclust_clu_complete_id30_c90_final_seq.sorted_optst30_2018_08" I0822 07:44:19.050110 140553952322112 utils.py:36] Started HHblits query

I0822 09:01:02.278290 140553952322112 utils.py:40] Finished HHblits query in 4603.228 seconds ` feature extraction spend time: 5305.185729026794 feature extraction Completed succesfully

I print the feature extraction time, find that , the 5305 is almost equals to the sum of each db search time, but according to the article, I think the feature extraction spend time should be almost equal to HHblits search, so can you explain the confusing problem?

Zuricho commented 2 years ago

5305 should be the sum of all MSA search time (actually include the PDB search, but that's fast), in the article, we state that the HHblits is the most time-consuming step, not meaning that the time for featuraization step should be equal to ones for HHblits.

Zuricho commented 2 years ago

If you mean the CPU accerlation, acutally we have not yet add that to this repo, I will upload that soon

yanchenmochen commented 2 years ago

If you mean the CPU accerlation, acutally we have not yet add that to this repo, I will upload that soon

oops, I means the step described from Sequential Tasks to Parallel Tasks on ParaFold. Threee independent sequential MSA searchs in parallel, in this way, we can reduce the duration consumed to complete MSA search。 @Zuricho image