I have a task with two conditions and 4 ROIs. I want to run this DCM in an exploratory manner if possible. 

Do I have to specify all possible models (~100 000) and then do BMS or can I specify a full model (full A,B,C matrix) and then prun the connections by means of PEB? 

I tried the latter, but when doing a sanity check (spm_dcm_fmri_check) this full model explains only a very low percentage of variance in most of the subjects and the intrinsic and extrinsic connections have high uncertainty.