CAREER: Optimizing Learning Models for Interpretation of Heterogeneous Biological Data

Virginia Commonwealth University, Richmond VA

Investigators

Abstract

The gap between techniques for rapid gathering of data and the ability to navigate the resulting large, heterogeneous datasets and formulate novel hypotheses is widening. In molecular biology, this limits basic discoveries and slows the translation of results from laboratory to clinic. Bridging this gap will necessarily involve intelligent algorithms. Yet, despite significant advances, machine-learning algorithms struggle with taking the calculated risk that underpins promising hypotheses. A machine can generate de novo a number of hypotheses that match experimental data, but are they plausible given our understanding of biology? An algorithm can use the data to score which of the known biological pathways is dysregulated in the disease, but can it propose a novel explanation of the observations? This project will address the limitation of current methods through novel algorithms that will be able to integrate existing knowledge into data-driven analysis and modeling. Following rigorous validation involving real and synthetic datasets, the proposed algorithms will form a transformative computational tool set for biologists, greatly enhancing the capabilities for understanding normal and pathological behavior of molecular systems involving heterogeneous elements interacting in complex ways. The work on the algorithms will involve undergraduate and graduate students. Through participation in multidisciplinary research, they will learn how to overcome barriers in cross-discipline communication, which is crucial in the world where computer methods permeate many other areas. Students will also benefit from a new course on Graph Theory and Machine Learning. The PI will make the learning modules from the course available online to instructors at other institutions. Also, annually, the PI will organize a high-school programming contest that will serve as a hands-on demonstration of how algorithms can help solve societal and scientific problems. All activities at high-school, undergraduate and graduate levels will have emphasis on supporting women and other underrepresented groups in their exploration of computer science. The project's main objective of creating algorithms for training classifiers that are accurate and interpretable will be achieved through graph-regularized machine learning that ties together ensembles, submodular set functions, and techniques for non-smooth function optimization. Submodular regularizers have recently received attention as a powerful way for equipping linear models with information about structures in the feature space, but extending the approach to non-linear classifiers is hampered by the nonlinearity of the relationship between measured features and the predicted outcome. This research will move past this obstacle by designing innovative ensemble-based methods that incorporate submodular regularizers, thus leading to improved accuracy, reduced overtraining and increased interpretability of models. These novel methods will be applicable to any classification problem that involves features connected by a network. Such problems are not confined to biology; the methods will be of broader use, for example to image analysis or text categorization. For biological applications, which will be the main focus of the work, a unified meta-network that links molecular elements at genetic, epigenetic, transcriptomic, proteomic and metabolomics levels will be constructed, and a new approach for dealing with missing data will be designed using concepts from spectral graph theory.