How DNA-binding proteins find their target sites remains a fascinating question. Early work on the Lac repressor showed that proteins specifically bind faster than simple diffusion allows. This led to the idea that recognition could be accelerated by combining diffusion with sliding along DNA and hopping between neighboring strands. This, in turn, implied that proteins could interact with DNA in a distinct nonspecific manner. Experimental work has confirmed that proteins can indeed slide along DNA, although typical sliding distances vary from protein to protein. 4] Recent work also indicates that sliding follows the helical grooves of DNA. Crystallography of proteins bound to noncognate sites, NMR spectroscopy, and molecular simulations have all provided data on nonspecific binding, notably suggesting that proteins maintain similar orientations with respect to DNA in nonspecifically and specifically bound states (see, e.g., Refs. [7-9]). However, little is known about the transition between these states, although theoretical studies have suggested that a switching mechanism may exist, possibly involving a protein conformational change. To analyze this problem at the atomic level, we carried out a molecular dynamics (MD) study on the dissociation of a specific protein-DNA complex in water, starting from the bound conformation. We studied the sex-determining region Y (SRY) protein, which affects the gender selection in mammals and is linked to a number of gender-related pathologies. The SRY protein binds in the minor groove of DNA, optimally at an (A/T)AACAAT sequence, and opens the minor groove by partial intercalation of an isoleucine residue (Ile13) between two adjacent AT base pairs and induces local unwinding and bending of DNA away from the protein. Using a specially designed restraint for the minimal atomic distance between any pair of nonhydrogen atoms across the protein-DNA interface (d MIN ), we were able to control the dissociation of the SRY protein from a 14-base-pair (bp) DNA oligomer (5'-CCTGCA-CAAACACC-3') without biasing the conformational pathway. Using this approach, roughly 0.6 ms of umbrella sampling