The function for much of the 3 billion letters in the human genome remain to be understood. Advances in DNA sequencing technology have generated enormous amount of data, yet we don’t have the tool to extract rules of how the genome works. Deep learning holds great potential in decoding the genome, in particular due to the digital nature of DNA sequences and the ability to handle large data sets. However, like many other applications, the interpretability of deep learning models hampers its ability to help understand the genome. We are developing deep learning architectures embedded with the principles of gene regulation and we will be leveraging billions of existing measurements of gene activity to learn a mechanistic model of gene regulation in human cells.

Continue reading

This project will be focused on creating a deep learning framework for tracking individual molecules and proteins as they move within a cell under various conditions. Using total internal reflection (TIRF) microscopy, we have accumulated more than 10 million trajectories over dozens of experimental preparations with differences in both the imaging approaches as well as the biological context. In our experiments we have captured particles under a wide variety of conditions including increased protein expression level, and a range of drug concentrations. Our biggest challenge is being able to stably track the movement of a particle as it passes by other particles or groups of particles, and to do this in a way that generalizes over novel conditions. The Data Science Institute Scholar chosen for this project would work with scientists in the Javitch laboratory and others across the Columbia campus to conceive of an approach for efficiently and effectively tracking particles. The resulting work would be of great interest to an increasing number of scientists working in this field who currently rely on methods based on feature engineering that are often inaccurate or inflexible compared to modern deep learning methods.

Continue reading

Big data with temporal dependence brings unique challenges in effective prediction and data analysis. The complex high-dimensional interactions between observations in such data brings unique challenges which standard off-the-shelf machine learning algorithms cannot handle. Even basic tasks of clustering, visualization and identification of recurring patterns are difficult.

Continue reading

A central issue facing systems neuroscience is defining the rich naturalistic behavioral repertoire that mice engage in under psychiatrically relevant situations. Recent advances in deep learning (e.g., DeepLabCut) have made frame by frame detailed pose estimation possible. However, this dense behavioral data requires new techniques for defining the ethogram (full description of behavior). To date, researchers have used frequency based time series approaches to tackle this problem, with significant limitations. An alternative approach would be to take advantage of new topology methods (persistent homology and directed algebraic topology) to characterize the shapes formed by mouse limb trajectories. Such an approach would have broad application in systems neuroscience. For this project, the student will use machine learning to label animal body parts, then topology to characterize the ethogram and compare the results to existing approaches.

Continue reading

The function for much of the 3 billion letters in the human genome remain to be understood. Advances in DNA sequencing technology have generated enormous amount of data, yet we don’t have the tool to extract rules of how the genome works. Deep learning holds great potential in decoding the genome, in particular due to the digital nature of DNA sequences and the ability to handle large data sets. However, like many other applications, the interpretability of deep learning models hampers its ability to help understand the genome. We are developing deep learning architectures embedded with the principles of gene regulation and we will be leveraging millions of existing whole genome measurements of gene activity to learn a mechanistic model of gene regulation in human cells.

Continue reading

Galaxies in our universe form hierarchically, continuously merging and absorbing smaller galaxies over cosmic time. In this project we aim to identify the most important features of, as well as generate efficient new features from, the merger histories of galaxies. Namely, features that predict (or physically speaking, determine) the properties of galaxies, e.g. their shape or color. This will be done using the results from a large cosmological simulation, IllustrisTNG (www.tng-project.org). We will begin with identifying ways to represent the rich information in the merger history. We will then compare various ML methods oriented towards feature selection or importance analysis: random forests or gradient boosted trees, L1SVM, neural networks (through analysis of e.g. saliency maps). More advanced models can also be applied, such as neural network models designed for feature selection. Finally, we wish to apply / develop methods that can build ‘interpretable’ new features by constructing them as algebraic formulas from original input features (inspired by e.g. https://science.sciencemag.org/content/324/5923/81). The overarching goal is to understand better what in the merger history is most crucial in determining a galaxy’s present-day properties, an answer to which can be widely applicable to problems in galaxy formation.

Continue reading

The visual cortex has a distinctive deep hierarchical organization as a result of ontogenetic and phylogenetic optimization. It is unclear what the factors are that shape this particular hierarchical organization. One factor is the compositional and hierarchical nature of our world’s appearance, which may be optimally processed by a hierarchical visual system. Another factor is the need for space and energy efficiency, which constrains the number of neurons and connections. The project will employ computational modeling to understand the contribution of these constraints to shaping the combination of breadth, depth, and skipping connections employed by primate visual cortex.

Continue reading