Lead Generation

Can computer-based methods generate leads?  The answer is yes – but the computer needs some information as input to generate meaningful output.  If multiple series of active compounds are available, then derived models describing common features can be used to identify novel chemistries that possess these features.   Derived models can be in the form of 3D pharmacophores – arrangements of functional groups in space that are thought to give rise to activity – or more abstracted patterns of topological and chemical properties as they are distributed throughout the molecule.  Pharmacophores can be derived and understood by human beings.  They can be used to identify novel compounds in large libraries, real or virtual, of available compounds.    Topological and chemical property patterns are derived by learning methods.  They similarly can be used to identify novel compounds in large libraries.    The more general the pharmacophore or pattern, the more likely it is to retrieve candidates that are dissimilar to the compounds that were used to derive them.   Biopredict has developed learning methods as mentioned above that can select among and utilize thousands of molecular descriptors in the generation of a model for activity (see Technologies).

A different lead-generation method that has come into widespread use is that of docking screens.  A docking screen is a structure-based method that can only be used if an experimentally determined structure is available for the target protein or a homolog.  When only  homologs are available the method requires as a preliminary step that a homology model be built for the target protein.  Docking screens can be effectively performed using homology models when the overall sequence identity of the target to the nearest homolog whose structure is known is on the order of 35%  (lower in families where extensive numbers of structures are available).

If the 100 highest scoring compounds from a 100,000 compound virtual screen are immediately tested the hit rate is typically on the order of a few percent.  There is however a way to derive more useful information from the virtual screen and to drive the effective hit rate higher.  To do this we cluster multiple conformations of energetically reasonably docked  compounds based on their interactions with the target active site. This is done by constructing a description of these interactions called a “footprint”.  Footprints can then compared with those of known active compounds against the target class: where there is significant  footprint overlap the cluster has an increased likelihood of being active.   Selected clusters are examined and used to derive pharmacophores.  These pharmacophores are then used to identify additional compounds from our corporate library of purchasable compounds or from other sources for testing.  The philosophy behind this strategy is similar to that used for the interpretation of high throughput screens: each docked compound is treated as a separate experiment.  When multiple compounds vote for a particular mode of binding then that mode of binding has increased credibility.