Section: New Results

Pepsi-Dock: fast predictions of putative docking poses using accurate knowledge-based potentials functions to describe interactions between proteins

Participants : Emilie Neveu, Sergei Grudinin, David W. Ritchie, Petr Popov.

Many biological tasks involve finding proteins that can act as an inhibitor for a virus or a bacteria, for example. Such task requires knowledge on the structure of the complex to be formed. Protein Data Bank can help but only a small fraction of its proteins are complexes [16] . Therefore, computational docking predictions, being low-cost and easy to perform, are very attractive if they describe accurately the interactions between proteins while being fast to find which conformation will be the most probable. We have been developing a fast and accurate algorithm that combines the FFT-accelerated docking methods [67] with a precise knowledge-based potential functions [58] describing interactions between the atoms in the proteins .

Interactions between proteins follow complex and non-linear laws which computation is time-consuming. It is of common usage to start the predictions with a simple, approximated, expression of these interactions to then reduce the space search in order to use more complex laws. However we think it is important to use the most accurate free energy not to miss some important docking solutions. Thus, our aim is to integrate the very-detailed knowledge-based potentials into the Hex code and to take advantage of its exhaustive search, which is by now still the most efficient and reliable search algorithm [67] .

Last year, we adapted the machine learning process so that the knowledge-based potentials describing atom interactions can be translated into the polynomial basis used in Hex. The current evaluations of the knowledge-based scores takes more time than a shape+electrostatic representation but is still fast: exploring ${10}^9$ conformations of a complex takes on average 5-10 minutes on a regular laptop computer.

This year, we run cross-validation experiments and tested different data sets in order to improve the predictions. Using bound conformations of each proteins to make the predictions, we retrieve up to 70% correct complexes of about 200 complexes. Results show that the knowledge-based potentials, while being general, correctly predict the interactions. Even better results could be achieved without the limitations in the search range by the spherical sampling grid which lacks of precision far away of its origin. Because many complexes have separation distances greater than 30 Å, we are now working on a multi-centre definition of the potentials in order to correctly predict the structures of protein complexes starting from their unbound structures.