| Summary: | Protein ligand docking is a valuable technique in the field of Structure Based Drug Design. One of its applications is to select one orfew candidates from a set of thousands or million compounds, often mmm as Virntal Screening. The choice of the correct scoring function is of paramount importance, and currently available scoring function are most often implemented to optimise reproduction of monn binding modes, or experimental free energies. The vast amount of information contained in the fact that most ligand do not, indeed bind to a given protein is discarded, building a considerable bias in these functions. The parameterisation method proposed in this work attempt to overcome said issue, by using a set ofdecoys among which the mown ligand is hidden. The parameters ofthe scoring function are chosen by a Genetic Algorithm so to maximize the monn ligand' ranking. The generality of the ftlnction is maintained by the constraint of simultaneously optimising several of these sets: the influence of the training set's size on the quality and robustness of the flUlction is assessed. Also, the effect of using different treference poses for the mown ligand, either experimental or re-generated by docking, is investigated. Finally, the immediate. applicability of the method is shown by reparameterising commercially available scoring functions (GOWSCORE, CHEMSCORE); a final attempt is made at creating scoringfimctions tailored to specific proteinfamilies', interpreting the results in tenllS ofwhat is mmm of their ligan-protein interactions. TIle results show that such a function is able to discriminate between mown actives and decoys, and would therefore be a valuable addition to the computational chemist bag of tricks to select those molecules able to tum into successfill drugs.
|
|---|
