QSAR/QSPR modeling for development of apatasensor and radio therapy dosimeter and 3D CNNEDPL system for ligand binding surface generation

Abstract

Another computational study in dosimetry response was also performed to address deficiencies in current technologies. Existing dosimeters in clinical use in radiation therapy have the drawbacks of high cost, loss of readout and lack of sensitivity. We developed a computationally enabled facile, robust and biocompatible dosimeter with amino acid capping gold nanoparticles in the hydrogel that would be sensitive in the low range of radiation doses. The resulting QSAR models were employed in the parallel screening of amino acids to identify the physico-chemical properties required for high sensitivity in the formation of gold nanoparticles and determine the lead component of a hydrogel nanosensor. The engineered hydrogel nanosensor shows high potential for translation because of its ease of fabrication, operation and readout in addition to being biocompatible making it a superior replacement to succeed existing commercial dosimeters in the clinic.

Over the past years, there has been an exponential increase in available chemical and biological data, which both compels and promotes the need for cheminformatics studies to extract information and build models to shorten the process of design and development of new molecules or materials. In this thesis, we applied traditional modeling methods to the design of radiation dosimeter and a synthetic biology study. Furthermore, a novel study of protein ligand relationships powered by convolutional deep neural networks is described in this thesis.

In order to study the mechanism for the activation of artificial riboswitches, a parallel screening and quantitative structure-property relationship (QSPR) study was done to identify the key substituents of pertinent chemical structures. As an effective computational method for predicting bioactivities and properties of new molecules, QSPR was employed to identify the physiochemical properties and computed descriptor that affect the response value (fluorescence) to study the binding specificity with different riboswitches that could form apatasensors as a fluorescence quantity indicator.

In drug discovery, it is the ultimate task to find a binding ligand to an identified target that could lead to modifying the course of disease. Traditional methods such as simulation-based approaches, despite some success, need to go through huge amounts of calculations and screening to find potential targets. Machine learning approaches that could be several orders of magnitude faster have the potential problem of overfitting and limitations in the choice of input features. Here we present a deep learning approach that could not only learn from the spatial distribution of the protein surface but also generate a possible respective ligand surface that could bind into the target with an appropriate property and shape distribution. For this task, we designed the new representation that could be feed into other deep neural network studies related to protein ligand relationship. A comprehensive study on the performance of different architectures was also done with metrics of accuracy and visualization of generated ligand 3D surface information.