Biophysical Neural Networks for Quantitative Analysis of Genomic Data

A quantitative and predictive understanding of gene regulation – how, when, and why specific genes are expressed as proteins - is crucial to fully deciphering cellular function which has profound consequences for biology, medicine, and biotechnology. Understanding and harnessing gene regulation is central to numerous applications in biofuels, biopharmaceuticals, and biotechnology. Moreover, many diseases including cancer, genetic disorders, and autoimmune conditions, result from misregulated gene expression. Gene regulation is driven by biophysical interactions between different regulatory proteins. These interactions can be quantified. But due to technical limitations, quantitative models of these interactions are only available for a handful of regulatory proteins. Advances in genomics and machine learning are providing the potential to address this shortcoming. The researchers have developed a novel framework, called BoltzNet, that combines the computational power of modern neural networks with the analytical power of biophysical models to translate genomic data directly into quantitative and predictive biophysical models of regulatory interactions. With this award, the researchers will extend BoltzNet to quantitatively model the regulatory interactions required to understand and predict gene regulation. The resources and algorithms developed in this research will have utility for molecular biologists and microbiologists seeking to quantitatively and mechanistically interpret genomic data, for synthetic biologists to predictively engineer new biotechnology applications, and for computational biologists seeking to develop interpretable and biophysically motivated neural networks. The project will extend BoltzNet to use ChIP-Seq and RNA-Seq genomic data to quantitatively and predictively model allosteric transcription factor (TF) binding to DNA, interactions between TFs, regulated binding of RNA polymerase to DNA, and ultimately transcriptional regulation. The resulting algorithms will enable the interpretation of genomic data in fundamental biophysical terms for diverse regulated promoters involving TF interactions; no such tool currently exists. The algorithms will also facilitate the quantitative design of new regulatory interactions for synthetic biology applications. The researchers will develop a publicly available BoltzNet web server to allow the scientific community to run BoltzNet algorithms on their own data sets. The web server will also host the results of analyzing publicly available bacterial ChIP-Seq data resulting in a compendium of quantitative biophysical models of transcriptional regulators as a resource for the scientific community. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.