Matching Items (2)
Filtering by
- All Subjects: Machine Learning
Description
Transient protein-protein and protein-molecule interactions fluctuate between associated and dissociated states. They are widespread in nature and mediate most biological processes. These interactions are complex and are strongly influenced by factors such as concentration, structure, and environment. Understanding and utilizing these types of interactions is useful from both a fundamental and design perspective. In this dissertation, transient protein interactions are used as the sensing element of a biosensor for small molecule detection. This is done by using a transcription factor-small molecule pair that mediates the activation of a CRISPR/Cas12a complex. Activation of the Cas12a enzyme results in an amplified readout mechanism that is either fluorescence or paper based. This biosensor can successfully detect 9 different small molecules including antibiotics with a tuneable detection limit ranging from low µM to low nM. By combining protein and nucleic acid-based systems, this biosensor has the potential to report on almost any protein-molecule interaction, linking this to the intrinsic amplification that is possible when working with nucleic acid-based technologies. The second part of this dissertation focuses on understanding protein-molecule interactions at a more fundamental level, and, in so doing, exploring design rules required to generalize sensors like the ones described above. This is done by training a neural network algorithm with binding data from high density peptide micro arrays incubated with specific protein targets. Because the peptide sequences were chosen simply to evenly, though sparsely, represent all sequence space, the resulting network provides a comprehensive sequence/binding relationship for a given target protein. While past work had shown that this works well on the arrays, here I have explored how well the neural networks thus trained, predict sequence-dependent binding in the context of protein-protein and peptide-protein interactions. Amino acid sequences, either free in solution or embedded in protein structure, will display somewhat different binding properties than sequences affixed to the surface of a high-density array. However, the neural network trained on array sequences was able to both identify binding regions in between proteins and predict surface plasmon resonance-based binding propensities for peptides with statistically significant levels of accuracy.
ContributorsSwingle, Kirstie Lynn (Author) / Woodbury, Neal W (Thesis advisor) / Green, Alexander A (Thesis advisor) / Stephanopoulos, Nicholas (Committee member) / Borges, Chad (Committee member) / Arizona State University (Publisher)
Created2022
Description
Biopolymers perform the majority of essential functions necessary for life. From a small amount of components emerges considerable complexity in both structure and function. The separated timescales of dynamic processes and intricate intra- and inter-molecular interactions of these molecules necessitate the development and utilization of computational approaches for biopolymer study and nanotechnology applications. Biopolymer nanotechnology exploits the natural chemistry of biopolymers to perform novel functions at the nanoscale. Molecular dynamics is the numerical simulation of chemical entities according to the physical laws of motion and statistical mechanics. The number of atoms in biopolymers require coarse-grained methods to fully sample the dynamics of the system with reasonable resources. Accordingly, a coarse-grained molecular dynamics model for the characterization of hybrid nucleic acid-protein nanotechnology was developed. Proteins are represented as an anisotropic network model (ANM) which show good agreement with experimentally derived protein dynamics for a small computational cost. The model was subsequently applied to hybrid DNA-protein cages systems and exhibited excellent agreement with experimental results. Ongoing development efforts look to apply network models to oxDNA origami to create multiscale models for DNA origami. The network approximation will allow for detailed simulation of DNA origami association, of concern to DNA crystal and lattice formation. Identification and design of target-specific binders (aptamers) has received considerable attention on account of their diagnostic and therapeutic potential. Generated in selection cycles from extensive random libraries, biopolymer aptamers are of particular interest due to their potential non-immunogenic properties. Machine learning leverages the use of powerful statistical principles to train a model to transform an input into a desired output. Parameters of the model are iteratively adjusted according to the gradient of the cost function. An unsupervised and generative machine learning model was applied to Thrombin aptamer sequence data. From the model, sequence characteristics necessary for binding were identified and new aptamers capable of binding Thrombin were sampled and verified experimentally. Future work on the development and utilization of an unsupervised and interpretable machine learning model for unaligned sequence data is also discussed.
ContributorsProcyk, Jonah (Author) / Sulc, Petr (Thesis advisor) / Stephanopoulos, Nicholas (Thesis advisor) / Hariadi, Rizal (Committee member) / Heyden, Matthias (Committee member) / Arizona State University (Publisher)
Created2022