Step 1. generate datafile with n sequences

head -n*2 original.fsa > small_sample.fsa

Step 2. run preprocess

./preprocessPaCE 5k.fsa 1

Step 3. run GItitle.pl

./GItitle.pl 5k.fsa.PaCE > 5k.GI.fsa

Step 4. run PaCE

#!/bin/csh #@ job_type = parallel #@ class = LONG #@ account_no = NONE #@ node = 2 #@ tasks_per_node = 4 #@ checkpoint = no #@ wall_clock_limit = 05:00:00 #@ error = $(Executable).$(Cluster).err #@ output = $(Executable).$(Cluster).out #@ environment = COPY_ALL #@ queue

llmachinelist mpirun -v -np 8 -machinefile /tmp/machinelist.$LOADL_STEP_ID ./PaCE_v9 /N/gpfs/cap3/leesangm/HumanMRNA/split1mil/5k.GI.PaCE 5000 ./Phase.cfg

llsubmit submitTest.sh

Step 5. split PaCE output

./PaCEclusterFasta.pl test.fsa estClust.5000.3.PaCE

##### Notes

1. Pace: How to run

mpirun -v -np 8 ./PaCE /N/gpfs/cap3/leesangm/HumanMRNA/example5K/5k.GI.fsa 5000 ./PaCE.cfg