First of all you should apply a standard equilibration protocol for your system, which should include minimization of the solvent environment with protein heavy atoms constraint to their initial structure (taken from protein data bank for example), then again constraints simulation annealing (heating and cooling the system) with protein heavy atoms fixed. Then, the restraints NPT simulations using harmonic restraints of the protein heavy atoms at their initial structure for about several 100 ps by releasing slowly the restraints until the system is equilibrated to the density of around 1 gr/cm^3. All equilibration protocol may take up to 1 ns, depending also on the system size. Then, you should further do simulations in NPT without any restraints/constraints on the protein for 1 ns or so, and see if the system behave properly.

Also, it is important to explain the choice of the thermostat/barostat for your case. Nose-Hoover are the most frequent used one, but it also depend on the MD code you have used.

I suppose the phenomenon may be caused by the application of periodic boundary condition. When periodic boundary conditions are used, a system of infinite extent is modeled. It is composed of an infinite number of copies of the primary system cell. As you describe, one or two chain dissociate from the protein and come together after 10 to 15 frame, this may be caused by the virtual image of the neighbouring cell. Then, the molecules were wrapped into the initial cell, the chains came together.

As Lulu says, this is almost certainly a periodicity issue. I have done simulations of 2BEG for hundreds of ns and it never falls apart.

after simulation perform trjconv

trjconv -f initial.xtc -s md.tpr -o 1.xtc -pbc whole

trjconv -f 1.xtc -s md.tpr -o 2.xtc -pbc nojump

trjconv -f 2.xtc -s md.tpr -o 3.xtc -pbc mol -ur compact -center

trjconv -f 3.xtc -md.tpr -o final.xtc -fit rot+trans