Date Thesis Awarded

5-2022

Access Type

Honors Thesis -- Access Restricted On-Campus Only

Degree Name

Bachelors of Science (BS)

Department

Biology

Advisor

Lizabeth Allison

Committee Members

Daniel Vasiliu

Randolph Chambers

M. Drew LaMar

Abstract

The thyroid hormone receptor α1 (TRα1) is a key protein for development and healthy growth of individuals that regulates gene expression in response to thyroid hormone. Mutations within TRα1 can cause Resistance to Thyroid Hormone (RTH) syndrome leading to potential physical setbacks as they play a key role in many aspects of growth. In addition, RTH missense mutants of TRα1 that have arginine replaced with either cysteine or histidine at position 384 have a more cytoplasmic localization than wild-type TRα1. TRα1 shuttles between the nucleus and cytoplasm, and prior studies suggest that post-translational modification by acetylation may affect localization. In the first part of this thesis, an acetyltransferase inhibitor was used to investigate whether blocking acetylation of these RTH missense mutants would promote greater nuclear localization. Expression plasmids for these proteins fused with GFP were transfected into HeLa cells. The intracellular localization patterns of wild-type and mutant GFP-TRα1 were measured using a Keyence BZ-X Series fluorescence microscope. The distribution of fluorescent pixels within the nucleus was compared to pixels illuminated within the cytoplasm as a ratio (N/C). Results showed that there was a significant increase in nuclear localization for the mutants with the addition of the acetyltransferase inhibitor.

The second part of this thesis worked with collected lab photos to engage in modeling with Convolutional Neural Networks (CNNs). All of the photos were cropped and preprocessed using Li thresholding methods to reduce noise within the data. The photos were divided into different sets based on classes and image types. The networks utilized were DenseNet, AlexNet, and StyleGAN. The first two networks classified each of the proteins based on the image data provided. The accuracy of the models decreased with an increase in the number of classes. The most successful network was DenseNet, likely due to its pre-trained function. StyleGAN was able to successfully create artificial cells; however, there was a lot of over-saturation with the pixels. The results show that TRα1 fluorescence images have potential to be implemented within classification networks.

Download

Share Feedback

On-Campus Access Only

Share

COinS