| | | | Given structure-property data sets the Discovery Bus automatically generates multiple QSAR models for each property and updates these as new data or methods become available. This creates a shifting landscape of QSAR models for multiple properties which can be used to guide the selection of novel chemical structures that satisfy the Research Target Profile (RTP) definition of a new drug. Forager has been developed to search for non-dominated solutions to an RTP within a complex descriptor space, where the search heuristics are provided by multiple QSAR models.
The objective for Forager is the rapid and complete identification of non-dominated solutions in Chemical Descriptor Space for multiple properties estimated by QSAR models.
Forager uses a modified Particle Swarm Optimisation (PSO) algorithm to search descriptor space for non-dominated solutions. The descriptor space is the union of descriptors in QSAR models used to estimate the properties of interest. The PSO is modified by allowing “herding” of particles into sub-groups that search together. A second modification is the variation of particle speed depending on recent success in identifying non-dominated solutions.
A Particle Swarm Optimisation (PSO) is a form of swarm intelligence. When one particle detects a desirable path the rest of the swarm will be able to follow quickly even if they are on the opposite side of the swarm. Particles are influenced by the rest of the swarm but also explore independently.
The swarm traverses reverse QSAR space as particles have a position and velocity in multi-dimensional descriptors space created from union of descriptors used by QSAR models. Movement is influenced by memory of their own best position and knowledge of the swarm's best. Particles communicate good positions to each other and adjust their own position and velocity based on these good positions defined in two ways, a global best updated when a new non-dominated position is found by any particle in the swarm and a neighbourhood best where each particle only communicates with a sub-set of the swarm about non-dominated solutions Since there is not one best global result, other techniques allow for global movement include separation (steering to avoid crowding), alignment (steer towards the average heading of local particles, and cohesion (steer to move toward the average position of local particles). These rules create 3 vectors which are then weighted and added to the vector for moving towards a local best, producing the finished movement vector. The relative weightings of these vectors is determined at the start of the simulation.
As particles move around the search space it is possible for them to vary their speed within lower and upper bounds. All particles are created with random speeds between the upper and lower bounds. If a particle doesn't find a new non-dominated solution, it’s speed will increase to cover more area. Alternatively, the particle slows to investigate the area in more detail. The change in speed depends on whether the particle found a personal best value or a global best. By varying the speed the particle can move quickly to cover large areas of search space, while exploring space more thoroughly once the right area has been found.
Because Forager contains a large number of arbitrary variables we have also written a program to optimize those variables using a conventional genetic algorithm. Work is continuing in this area.
Conclusions from our first simple study suggestst the following. Forager rapidly identifies non-dominated solutions in descriptor using 2 QSAR Linear Models for HIV Protease Inhibition and Solubility. Even though constrained to stay within descriptor ranges for drug-like compounds, optimisation goes beyond normal property ranges Fully automated operation and updating using the “Discovery Bus” Easily extended to optimise more than 2 properties . Easily extended to more complex QSAR models
|
| |
|
