-I see your ligand is ATP!? If so, I assume the binding site should contain a fair amount of charged resides, and then you could use a higher intdiel, (in the input file this variable is commented and set as default=1)... try with 5 and see how it goes...
-Also, consider running the calculation for a chunk of the trajectory where the system is relatively stable, let's say from the frame 500 to 1000 (you can either check the energy or the rmsd)...
-Finally, try with a more recent GB model igb=8
hope this helps!