Computational methods are useful to identify favorable structures of transmembrane (TM) helix oligomers when experimental data are not available or when they cannot help to interpret helix–helix association. We report here a global search method using molecular dynamics (MD) simulations to predict the structures of transmembrane homo and heterodimers. The present approach is based only on sequence information without any experimental data and is first applied to glycophorin A to validate the protocol and to the HER2-HER3 heterodimer receptor. The method successfully reproduces the experimental structures of the TM domain of glycophorin A (GpATM) with a root mean square deviation of 1.5 Å. The search protocol identifies three energetically stable models of the TM domain of HER2-HER3 receptor with favorable helix–helix arrangement, including right-handed and left-handed coiled-coils. The predicted TM structures exhibit the GxxxG-like motif at the dimer interface which is presumed to drive receptor oligomerization. We demonstrate that native structures of TM domain can be predicted without quantitative experimental data. This search protocol could help to predict structures of the TM domain of HER heterodimer family.