Cellular behavior emerges from a complex network of chemical interactions, the details of which remain largely unknown. Developing effective therapies requires a quantitative understanding of these fundamental processes and how they can be safely manipulated to thwart disease. Our lab focused on constructing mathematical models to assist in diagnosis and treatment of neurodegenerative disease and cancer.
Many complex diseases can only be successfully treated with multi-drug therapeutic strategies. Identifying the proper drugs, doses and scheduling for an effective combination therapy is hindered by the sheer number of permissible combinations. Our lab developed computational methods for predicting the effects of combination treatments in phenotypic screens. Working with researchers at the University of Pittsburgh Drug Discovery Institute, we applied these methods to drug discovery in cancer and neurodegenerative diseases including Alzheimer's and Huntington's.
The molecular mechanisms of many modern diseases are poorly understood, and a critical bottleneck in conventional early-stage drug discovery is identifying molecular targets for drugs. Phenotypic screens allow investigators to find compounds that effectively treat cell-based models of diseases, but the screens themselves don't identify mechanisms. We developed novel computational approaches to uncover targets and pathways that are relevant to disease onset and progression. We applied our methods, which combine statistical inference with mechanistic modeling, to better understand druggable molecular systems in breast cancer, Huntington's disease, and traumatic brain injury.
Typical hit rates for monotherapies in phenotypic screens are around 5%, and the success rate plummets exponentially when screening combinations of compounds. Running large screens requires time and resources that may not be widely available to researchers. Our lab developed methods to actively incorporate the results of ongoing screens into the computation as they are acquired. This active learning approach enables screeners to computationally predict the results of the next round of experiments, perform those experiments with the greatest expected pay-off, and fold the results back into the software for the next round of predictions. Our methods were constructed to extract maximal information from limited phenotypic screens and can be readily applied to diverse diseases, assays, and data types.
An important part of drug discovery is determining safe dose regimens for a heterogeneous patient population. Increasingly, this means personalizing medications and doses to provide optimal effects in individual patients. Working with industry collaborators, the lab developed mathematical models of the immune response to cancer treatment. Our goal was to quantitatively predict how individual patients will react to a variety of dosing regimens, in order to optimize the efficacy and minimize the side effects of cancer therapy.