Section: New Results

Flex-Dock: towards flexible docking predictions using metaheuristics optimisation methods

Participants : Emilie Neveu, Sergei Grudinin, Alexandre Hoffman, Angelo Migliosi, Xavier Besseron, Grégoire Danoy, Pascal Bouvry.

Docking numerical methods are used to predict the preferred location of one molecule with respect to the second when bound to each other. This is particularly useful for the design of drugs that inhibit the effects of viruses or bacteria. However proteins change their conformation upon binding and searching for flexible conformations involves enormous degrees of freedom and complex physics. Thus, the prediction of realistic interactions with full flexibility of the two partners is an intractable global optimisation problem.

There are currently several algorithms that produce high quality predictions of molecular complexes [43] . But very few manages to deal with the flexibility of the proteins. A common method is to refine the most probable predicted rigid complexes with a scoring allowing for flexibility [81] . Here, we want to tackle flexibility and sampling all together. Exhaustive search methods, which were by now the most accurate optimisation method for relatively small molecules [53] will be too time-consuming when it comes to large proteins. There is a strong need to explore and define new optimisation algorithms such as metaheuristic ones that can deal with several local minima and a large minima and a large search space. The main goal of this project is to define the problem and find for the optimisation method that will potentially give better results than the actual reference, SwarmDock [54] .

We worked on a first comparison of several evolutionary-based algorithms (Genetic Algorithm [40] , Differential Evolution [76] , Particle Swarm Optimisation [46] ) using rigid proteins only and on the use of multi-objective algorithms when the proteins are flexible.

To take into account flexibility, we approximate large-scale deformations of each proteins using an elastic network model combined with a low-frequency approximation called normal mode analysis such as in [81] or in [54] . Combined with the rigid transformation between the two proteins, it defines a complete while reduced set of degrees of freedom to search for.

The scoring function has to discriminate correct conformations from impossible ones. Our scoring is the main difference with SwarmDock. It takes into account the energy gained by docking using the precise knowledge-based potentials derived in [58] , whereas only a simple physics-based energy is used in SwarmDock. We also want to explore another scoring that will also add the energetic cost of each moves of the proteins. To do so, we started to develop multi-objective algorithms. Combined with a Pareto Front analysis, this will help us to validate the scoring and to compare different evolutionary-based algorithms.Tests will be directly made on the Protein-Protein Benchmark [42] so that we can compare with other docking methods.