|
Ameet SoniAssociate ProfessorAssociate Dean of the Faculty for Diversity, Recruitment, and Retention Computer Science Department Swarthmore College phone: (610) 957-6288 office: 212 East Parrish email: 
|
|
|
Ameet SoniAssociate ProfessorAssociate Dean of the Faculty for Diversity, Recruitment, and Retention Computer Science Department Swarthmore College phone: (610) 957-6288 office: 212 East Parrish email:
|
Transcription factor binding: transcription factors govern the regulation of genes - that is, determine when a gene is on or off. Understanding which transcription factors bind to which areas of the genome and in which cells helps understand the function of target genes. Our lab utilizes deep neural networks to predict the binding affinity of specific transcription factors to a given portion of DNA to identify these relationships without the need for expensive in vivo experiments.
Brain image analysis: Recently, my group has been researching several different computational problems in the area of MRI braining imaging. Initial results in this area include improved approaches for performing image segmentation in brain images using probabilistic graphical models (conditional random fields). Ongoing work aims to apply deep learning approaches - including convolutional neural networks - to improve early diagnosis of Alzheimer's disease.
Statistical relational learning: Real world data is inherently noisy and relational (i.e., elements are dependent on one another). Traditional machine learning algorithms fail to account for these realities. In collaboration with Prof. Sriraam Natarajan at Indiana University, I have done work with Relational Dependency Networks to model problems in complex domains. Examples including relational extraction problems from text (e.g., identify the CEO of a company from a newswire article) and diagnosis of Parkinson's disease from medical records.
Protein Structure Prediction: Previously, I worked on ACMI (Automated Crystallographic Map Interpretation). The task of determining protein structures has been a central one to the biological community for several decades. The most popular method for producing protein structures is by interpreting an electron-density map - a three-dimensional image of a molecule produced through X-ray crystallography. This process, however, remains a resource- intensive and time-consuming task, stunting basic biological research. Thus, the main objective of the project is: The result of our group's efforts is ACMI, a probabilistic technique for determining protein structures. Prior to ACMI, techniques failed when trying to interpret low-quality images. With ACMI, crystallographers can now obtain complete and accurate structures from these difficult proteins instead of scrapping the project or dedicating months of effort.
Asterick's (*) indicate supervised students.