CRII: III: A Scalable Probabilistic Model Selection Method for Deep Learning in Gene-Protein Network Inference and Integration

Rochester Institute Of Tech, Rochester NY

Investigators

Abstract

Detailed characterization of interactions between individual genes and proteins has been one of the focuses of biological research, since these interactions control cellular processes involving complicated cascades of biochemical reactions and signaling pathways. Dysregulation of any component of these pathways can lead to a broad spectrum of human pathologies including cancer, cardiovascular disease, neurodegenerative conditions, and metabolic diseases. Deep learning techniques, which is attributed to the most recent advances in artificial intelligence and machine learning, have emerged and many studies have tried to construct and analyze gene/protein interactions at a genome/proteome-wide scale to describe their global characteristics. However, the heterogeneous nature of the biological data and their continuous evolution pose a unique challenge to the architecture design of deep neural networks. The state-of-the-art solutions heavily rely on expertise, heuristics, and experimentation, and are time-consuming and not scalable. This research proposes to enable deep neural networks to automatically go through a qualitative growth to accommodate richer information from new heterogenous data as they are accumulating. This project will provide powerful novel computational tools to discover and target gene-protein interactions driving regulatory diseases such as cancer, diabetes and neurological disorders, while gaining insight into the fundamental principles behind cellular information processing. This project will focus on developing scalable probabilistic model selection approaches to infer deep neural network architectures for gene-protein interaction network inference and integration. The project will design efficient inference algorithms for the proposed model selection approach to enable its translation into a deployable tool for use by biologists. The researchers will develop novel principled model selection methods to infer the most plausible architectures of deep neural networks warranted by the heterogenous biological data. By modeling the hypothesis space of neural network architectures as stochastic processes, the proposed method enables neural network architectures to evolve according to the biological data; to design and implement efficient techniques to make use of the inference computationally tractable. The project proposes to evaluate the marginal likelihoods for preference to the alternatives approximately based on variational methods. The investigator will compare the performance of the deep neural networks whose architectures are learned with the proposed model selection method with the state-of-the-art methods on networks and functional annotations of eukaryotic organism S. cerevisiae and human cells by treating protein function prediction as a multi-label classification problem, and measuring the performance with two complementary approaches: cross-validation and temporal holdout validation. 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.