How a molecule interacts within a living organism is incredibly complicated. Even the simplest molecules can interact with over 25,000 protein products from genes (not including their splice variants) as well as a wide variety of lipids, sugars, nucleic acids, and their combinations distributed throughout the body. In a typical drug discovery effort, a narrow but important range of proteins are tested for their interactions with a drug candidate molecule. A reality that, the more you think about, is disappointing. One of the great hopes in drug discovery is that machine learning can help bridge the gap between a drug’s effects and its biophysical interactions.
Predicting a molecule’s biological binding can be performed via multiple philosophies and computational methods. The method I want to focus on is via taking known ligands of proteins to “train” an algorithm and then use the algorithm to predict targets of ligands with unknown or unverified targets. It is a pretty clever strategy, and if it works, a truly invaluable tool that can help bridge the gaps in human knowledge of so many molecules. In any case, using ligand-based methods for target prediction is, in my opinion, more attractive in many aspects than structure-based methods since a greater emphasis is placed on chemical moieties that are well known to interact with specific proteins. Non-specific interactions theoretically have lower propensity to skew results when using ligand-based prediction, a major issue which can be traced back to the “scores” derived from force fields which are needed when applying structure-based methods.
Two of the most important biological interaction prediction software include the SwissTargetPrediction developed by the Swiss Institute of Bioinformatics and the Similarity Ensemble Approach developed by professor Brian Shoichet’s lab at UCSF. Since the programs have different methodologies, they should have slightly different perspectives on predicted biological properties. The value of these predictions, assuming a decent accuracy, is multifold. Firstly, these predictions would enable a focused in vitro screening of potentially favorable, or toxic, interactions between a small molecule and a target. A much larger investment of time and money is required for biological experiments, so any process or algorithm which can optimize the money spent on such experiments is highly valuable. Another more general benefit of biological target prediction is to understand how the medicines, nutraceuticals, and natural products so many people consume really work. My plan is to use these webservers to help understand many commonly used drugs, nutraceuticals, and natural medicines, and then develop potential derivatives of these natural medicines with enhanced efficacy. One could call this new subfield of drug discovery “Computational Pharmacology.”