Discovering new drugs is an expensive and challenging process. Any new drug should not only produce the desired response to the disease but should do so with minimal side effects and be superior to the existing drugs in the market. The goal of this project is to develop effective and efficient algorithms for analyzing chemical compound databases and identifying biologically active compounds.
One of the key steps in the drug design process is the identification of the chemical compounds (hit compounds) that display the desired and reproducible behavior against the specific biomolecular target and represents a significant hurdle in the early stages of drug discovery. The 1990s saw the widespread adoption of high-throughput screening (HTS), which use highly automated techniques to conduct the biological assays and can be used to screen a large number of compounds. Although the number of compounds that can be evaluated by these methods is very large, these numbers are small in comparison to the millions of drug-like compounds that exist or can be synthesized by combinatorial chemistry methods. Moreover, in most cases it is hard to find all desirable properties in a single compound and medicinal chemists are interested in not just identifying the hits but studying what part of the chemical compound leads to desirable behavior, so that new compounds can be rationally synthesized (lead development).
Computational techniques that build models to correctly assign chemical compounds to various classes of interest can address these limitations, have extensive applications in pharmaceutical research, and they are used extensively to replace or supplement HTS-based approaches.
Our most recent research is currently concentrated on the following areas:
Develop computationally efficient algorithms to mine large databases of molecular graphs and identify key substructures present in active (inactive) compounds.
Develop sophisticated feature selection and generation algorithms that combine multiple criteria to identify and synthesize a set of substructure-based features that simultaneously simplify the representation of the original compounds while retaining and exposing their key features.
Develop kernel-based clustering and classification approaches that take into account the relationships between these substructures at different levels of granularity and complexity.
The research over the years has been funded by a number of Federal agencies including NIH (primary agency), ARL, and NSF.
Our research thus far has resulted in the development of computationally efficient algorithms to find frequent substructures in molecular graphs (either topological or geometric). The topological version of this algorithm, called FSG, is currently available as part of our pattern discovery toolkit PAFI, which can be downloaded and used for educational and research purposes.
Another recent development is the AFGen program that operates on a database of chemical compounds and generates their descriptor-based representation by considering all bounded length acyclic fragments that they contain. These descriptors are quite effective in capturing the structural characteristics of chemical compounds. Experiments in the context of SVM-based classification and ranked-retrieval show that these descriptors consistently and statistically outperform previously developed schemes based on the widely used fingerprint- and Maccs keys-based descriptors, as well as recently introduced descriptors obtained by mining and analyzing the structure of the molecular graphs.